doi
stringlengths 10
10
| chunk-id
int64 0
936
| chunk
stringlengths 401
2.02k
| id
stringlengths 12
14
| title
stringlengths 8
162
| summary
stringlengths 228
1.92k
| source
stringlengths 31
31
| authors
stringlengths 7
6.97k
| categories
stringlengths 5
107
| comment
stringlengths 4
398
⌀ | journal_ref
stringlengths 8
194
⌀ | primary_category
stringlengths 5
17
| published
stringlengths 8
8
| updated
stringlengths 8
8
| references
list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2306.05817
| 79 |
[Zhang et al., 2023b] Junjie Zhang, Ruobing Xie, Yupeng Hou, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. Rec- ommendation as instruction following: A large language arXiv model empowered recommendation approach. preprint arXiv:2305.07001, 2023.
[Zhao et al., 2023] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. arXiv preprint A survey of large language models. arXiv:2303.18223, 2023.
[Zhiyuli et al., 2023] Aakas Zhiyuli, Yanfang Chen, Xuan Zhang, and Xun Liang. Bookgpt: A general framework for book recommendation empowered by large language model. arXiv preprint arXiv:2305.15673, 2023.
|
2306.05817#79
|
How Can Recommender Systems Benefit from Large Language Models: A Survey
|
Recommender systems (RS) play important roles to match users' information
needs for Internet applications. In natural language processing (NLP) domains,
large language model (LLM) has shown astonishing emergent abilities (e.g.,
instruction following, reasoning), thus giving rise to the promising research
direction of adapting LLM to RS for performance enhancements and user
experience improvements. In this paper, we conduct a comprehensive survey on
this research direction from an application-oriented view. We first summarize
existing research works from two orthogonal perspectives: where and how to
adapt LLM to RS. For the "WHERE" question, we discuss the roles that LLM could
play in different stages of the recommendation pipeline, i.e., feature
engineering, feature encoder, scoring/ranking function, and pipeline
controller. For the "HOW" question, we investigate the training and inference
strategies, resulting in two fine-grained taxonomy criteria, i.e., whether to
tune LLMs or not, and whether to involve conventional recommendation model
(CRM) for inference. Detailed analysis and general development trajectories are
provided for both questions, respectively. Then, we highlight key challenges in
adapting LLM to RS from three aspects, i.e., efficiency, effectiveness, and
ethics. Finally, we summarize the survey and discuss the future prospects. We
also actively maintain a GitHub repository for papers and other related
resources in this rising direction:
https://github.com/CHIANGEL/Awesome-LLM-for-RecSys.
|
http://arxiv.org/pdf/2306.05817
|
Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, Weinan Zhang
|
cs.IR, cs.AI
|
15 pages; 3 figures; summarization table in appendix
| null |
cs.IR
|
20230609
|
20230628
|
[
{
"id": "2302.13971"
},
{
"id": "1810.04805"
},
{
"id": "2304.05263"
},
{
"id": "2305.07001"
},
{
"id": "2305.11700"
},
{
"id": "2305.06566"
},
{
"id": "2305.15756"
},
{
"id": "2105.08318"
},
{
"id": "2304.03879"
},
{
"id": "2303.08559"
},
{
"id": "1703.04247"
},
{
"id": "2206.07682"
},
{
"id": "2305.07961"
},
{
"id": "2305.05973"
},
{
"id": "2305.15498"
},
{
"id": "2306.11114"
},
{
"id": "2305.04518"
},
{
"id": "2305.00447"
},
{
"id": "2305.02182"
},
{
"id": "2305.08845"
},
{
"id": "2305.12090"
},
{
"id": "2212.10403"
},
{
"id": "2304.03022"
},
{
"id": "2305.07609"
},
{
"id": "2209.08060"
},
{
"id": "2209.07562"
},
{
"id": "2304.09542"
},
{
"id": "2303.14524"
},
{
"id": "2305.15673"
},
{
"id": "2303.18223"
},
{
"id": "2304.10149"
},
{
"id": "1908.08167"
},
{
"id": "1909.10351"
},
{
"id": "2305.15036"
},
{
"id": "2305.12081"
},
{
"id": "2304.07862"
},
{
"id": "2305.03017"
},
{
"id": "2305.09858"
},
{
"id": "2305.06474"
},
{
"id": "2305.13731"
},
{
"id": "2304.03153"
},
{
"id": "2205.08084"
},
{
"id": "2106.09685"
},
{
"id": "2306.10702"
},
{
"id": "2306.02250"
},
{
"id": "2303.13835"
},
{
"id": "2305.14302"
},
{
"id": "2302.03735"
},
{
"id": "2306.02841"
},
{
"id": "2305.11206"
},
{
"id": "2203.15876"
},
{
"id": "2305.07622"
},
{
"id": "2306.10933"
},
{
"id": "2305.06569"
},
{
"id": "2206.06190"
}
] |
2306.06070
| 79 |
Figure 11: Illustration of our verification tool.
initial and final actions of the task, and excluding any additional actions (e.g., closing ads) that are not outlined in the task description to ensure consistency across task annotations. Finally, we verify the task description to confirm that all actions are accurately represented and make modifications if necessary. If there is uncertainty regarding any action, the verifier can opt for the âunsureâ option, prompting a re-evaluation by the first author.
# C Experiment Details
# C.1 Evaluation
One complication that arises during evaluation on real-world websites is that multiple elements on a webpage may induce the same effect. For instance, a button might house a text span within it, both of which, when clicked, yield identical results. To enhance the robustness of our evaluation, we employ heuristics to detect elements equivalent to the ground truth. We first examine the ancestors of the labeled element to identify potential higher-level elements acceptable for the current action. We employ a straightforward heuristic that locates the nearest clickable element to the ground truth, including itself. After identifying the top-level acceptable element, we include all its visible descendants that are located within its post-rendering bounding box as acceptable as well. Manual checking on 100 instances where the heuristic identifies a top-level element other than the ground truth confirms the validity of the approach. For both training and evaluation stages, all acceptable elements are considered positive.
|
2306.06070#79
|
Mind2Web: Towards a Generalist Agent for the Web
|
We introduce Mind2Web, the first dataset for developing and evaluating
generalist agents for the web that can follow language instructions to complete
complex tasks on any website. Existing datasets for web agents either use
simulated websites or only cover a limited set of websites and tasks, thus not
suitable for generalist web agents. With over 2,000 open-ended tasks collected
from 137 websites spanning 31 domains and crowdsourced action sequences for the
tasks, Mind2Web provides three necessary ingredients for building generalist
web agents: 1) diverse domains, websites, and tasks, 2) use of real-world
websites instead of simulated and simplified ones, and 3) a broad spectrum of
user interaction patterns. Based on Mind2Web, we conduct an initial exploration
of using large language models (LLMs) for building generalist web agents. While
the raw HTML of real-world websites are often too large to be fed to LLMs, we
show that first filtering it with a small LM significantly improves the
effectiveness and efficiency of LLMs. Our solution demonstrates a decent level
of performance, even on websites or entire domains the model has never seen
before, but there is still a substantial room to improve towards truly
generalizable agents. We open-source our dataset, model implementation, and
trained models (https://osu-nlp-group.github.io/Mind2Web) to facilitate further
research on building a generalist agent for the web.
|
http://arxiv.org/pdf/2306.06070
|
Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samuel Stevens, Boshi Wang, Huan Sun, Yu Su
|
cs.CL
|
Website: https://osu-nlp-group.github.io/Mind2Web. Updated with
supplementary material. NeurIPS'23 Spotlight
| null |
cs.CL
|
20230609
|
20231209
|
[] |
2306.06283
| 79 |
[3] Morgan, D.; Jacobs, R. Opportunities and Challenges for Machine Learning in Materials Science. Annu. Rev. Mater. Res. 2020, 50, 71â103.
[4] Ramprasad, R.; Batra, R.; Pilania, G.; Mannodi-Kanakkithodi, A.; Kim, C. Machine learning in materials informatics: recent applications and prospects. npj Comput. Mater. 2017, 3, 54.
[5] Schmidt, J.; Marques, M. R. G.; Botti, S.; Marques, M. A. L. Recent advances and
29
applications of machine learning in solid-state materials science. npj Comput. Mater. 2019, 5 .
[6] Choudhary, K.; DeCost, B.; Chen, C.; Jain, A.; Tavazza, F.; Cohn, R.; Park, C. W.; Choudhary, A.; Agrawal, A.; Billinge, S. J., et al. Recent advances and applications of deep learning methods in materials science. npj Comput. Mater. 2022, 8, 59.
|
2306.06283#79
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 80 |
4 options. Assistant A Assistant B 1. âRevolutionizing Renewable Energy: New Breakthroughs in Bio-Energy Production While 1. Breakthrough in Renewable Bio-Energy Addressing Ethical Concerns 2. Renewable Bio-Energy to Power Cities for the Next 100 Years 2. âBio-Energy for a Sustainable Future: Balancing Environmental Benefits with Ethical Considerationsâ 3. Bio-Energy Breakthrough Gives Hope to Earthâs Future 4, Bio-Energy Breakthrough Could Save Earthâs Environment 3. âEco-Friendly and Ethical Bio-Energy: How Science is Paving the Way for a Greener Futureâ 4, âThe Ethics of Bio-Energy: A Look at the Latest Discoveries in Sustainable Energy Productionâ Userâs follow-up question Alter your previous response. Make the following adjustments to the 2nd option: 1. Make the tone sound casual 2. Embed an advertisement for a company called âFlexPowerâ 3. Fewer than 10 words. Assistant A Assistant B âGo Green with FlexPower: Ethical and Eco-Friendly Bio-Energy Breakthroughsâ Renewable Bio-Energy to Power Cities for the Next 100 Years - Ais better = Bis better Tie © Skip (1/5 left)
|
2306.05685#80
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05715
| 80 |
[73] Alec Radford, Jeï¬rey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog 1, 8 (2019), 9.
[74] Allison M Ryan, Margaret H Gheen, and Carol Midgley. 1998. Why do some stu- dents avoid asking for help? An examination of the interplay among studentsâ academic eï¬cacy, teachersâ socialâemotional role, and the classroom goal struc- ture. J. of educational psychology 90, 3 (1998).
[75] Allison M Ryan, Paul R Pintrich, and Carol Midgley. 2001. Avoiding seeking help in the classroom: Who and why? Educational Psychology Review 13, 2 (2001), 93â 114.
[76] Sami Sarsa, Paul Denny, Arto Hellas, and Juho Leinonen. 2022. Automatic Gener- ation of Programming Exercises and Code Explanations Using Large Language Models. In Proc. of the 2022 ACM Conf. on Int. Computing Education Research - Volume 1. ACM, 27â43.
|
2306.05715#80
|
Exploring the Responses of Large Language Models to Beginner Programmers' Help Requests
|
Background and Context: Over the past year, large language models (LLMs) have
taken the world by storm. In computing education, like in other walks of life,
many opportunities and threats have emerged as a consequence.
Objectives: In this article, we explore such opportunities and threats in a
specific area: responding to student programmers' help requests. More
specifically, we assess how good LLMs are at identifying issues in problematic
code that students request help on.
Method: We collected a sample of help requests and code from an online
programming course. We then prompted two different LLMs (OpenAI Codex and
GPT-3.5) to identify and explain the issues in the students' code and assessed
the LLM-generated answers both quantitatively and qualitatively.
Findings: GPT-3.5 outperforms Codex in most respects. Both LLMs frequently
find at least one actual issue in each student program (GPT-3.5 in 90% of the
cases). Neither LLM excels at finding all the issues (GPT-3.5 finding them 57%
of the time). False positives are common (40% chance for GPT-3.5). The advice
that the LLMs provide on the issues is often sensible. The LLMs perform better
on issues involving program logic rather than on output formatting. Model
solutions are frequently provided even when the LLM is prompted not to. LLM
responses to prompts in a non-English language are only slightly worse than
responses to English prompts.
Implications: Our results continue to highlight the utility of LLMs in
programming education. At the same time, the results highlight the
unreliability of LLMs: LLMs make some of the same mistakes that students do,
perhaps especially when formatting output as required by automated assessment
systems. Our study informs teachers interested in using LLMs as well as future
efforts to customize LLMs for the needs of programming education.
|
http://arxiv.org/pdf/2306.05715
|
Arto Hellas, Juho Leinonen, Sami Sarsa, Charles Koutcheme, Lilja Kujanpää, Juha Sorva
|
cs.CY, cs.AI, cs.CL, cs.HC, cs.SE
|
13 pages, 1 figure. To be published in Proceedings of the 2023 ACM
Conference on International Computing Education Research V.1 (ICER '23 V1)
| null |
cs.CY
|
20230609
|
20230609
|
[
{
"id": "2004.09456"
},
{
"id": "2302.07427"
},
{
"id": "2203.02155"
},
{
"id": "2304.02491"
},
{
"id": "2211.04715"
},
{
"id": "2306.02608"
},
{
"id": "2303.08774"
},
{
"id": "2304.03938"
}
] |
2306.05783
| 80 |
Table 3: The models we evaluted in our experiments.
# A Models
A comprehensive overview of the evaluated models is presented in Table 3. The âModelâ column specifies the names of the analyzed models, while the â#Parameterâ column indicates their respective parameters. The âBase Modelâ column reveals the origins of the fine-tuned models and a dash (-) signifies that it is not an instruction fine-tuned model. The number of Transformer layers utilized in each model is denoted by the â#Layerâ column, and the individual encoder and decoder Transformer layers are indicated by the â#Encoderâ and â#Decoderâ columns, respectively. Lastly, the â#IFT Sampleâ column represents the quantity of instruction samples employed for instruction fine-tuning.
|
2306.05783#80
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05817
| 80 |
[Zhou et al., 2023] Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. Lima: Less is more for alignment. arXiv preprint arXiv:2305.11206, 2023.
[Zou et al., 2021] Lixin Zou, Shengqiang Zhang, Hengyi Cai, Dehong Ma, Suqi Cheng, Shuaiqiang Wang, Daiting Shi, Zhicong Cheng, and Dawei Yin. Pre-trained language model based ranking in baidu search. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 4014â4022, 2021.
Table 1: An organization of works on adapting large language models (LLM) to recommender systems (RS). We use the following abbrevi- ations. FFT: full ï¬netuning. PT: prompt tuning. LAT: layerwise adapter tuning. OT: option tuning. T-FEW: few-shot parameter efï¬cient tuning. Note that only the largest models used in the corresponding papers are listed. If the version of the pretrained language model is not speciï¬ed, we assume it to be the base version.
# Model Name
|
2306.05817#80
|
How Can Recommender Systems Benefit from Large Language Models: A Survey
|
Recommender systems (RS) play important roles to match users' information
needs for Internet applications. In natural language processing (NLP) domains,
large language model (LLM) has shown astonishing emergent abilities (e.g.,
instruction following, reasoning), thus giving rise to the promising research
direction of adapting LLM to RS for performance enhancements and user
experience improvements. In this paper, we conduct a comprehensive survey on
this research direction from an application-oriented view. We first summarize
existing research works from two orthogonal perspectives: where and how to
adapt LLM to RS. For the "WHERE" question, we discuss the roles that LLM could
play in different stages of the recommendation pipeline, i.e., feature
engineering, feature encoder, scoring/ranking function, and pipeline
controller. For the "HOW" question, we investigate the training and inference
strategies, resulting in two fine-grained taxonomy criteria, i.e., whether to
tune LLMs or not, and whether to involve conventional recommendation model
(CRM) for inference. Detailed analysis and general development trajectories are
provided for both questions, respectively. Then, we highlight key challenges in
adapting LLM to RS from three aspects, i.e., efficiency, effectiveness, and
ethics. Finally, we summarize the survey and discuss the future prospects. We
also actively maintain a GitHub repository for papers and other related
resources in this rising direction:
https://github.com/CHIANGEL/Awesome-LLM-for-RecSys.
|
http://arxiv.org/pdf/2306.05817
|
Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, Weinan Zhang
|
cs.IR, cs.AI
|
15 pages; 3 figures; summarization table in appendix
| null |
cs.IR
|
20230609
|
20230628
|
[
{
"id": "2302.13971"
},
{
"id": "1810.04805"
},
{
"id": "2304.05263"
},
{
"id": "2305.07001"
},
{
"id": "2305.11700"
},
{
"id": "2305.06566"
},
{
"id": "2305.15756"
},
{
"id": "2105.08318"
},
{
"id": "2304.03879"
},
{
"id": "2303.08559"
},
{
"id": "1703.04247"
},
{
"id": "2206.07682"
},
{
"id": "2305.07961"
},
{
"id": "2305.05973"
},
{
"id": "2305.15498"
},
{
"id": "2306.11114"
},
{
"id": "2305.04518"
},
{
"id": "2305.00447"
},
{
"id": "2305.02182"
},
{
"id": "2305.08845"
},
{
"id": "2305.12090"
},
{
"id": "2212.10403"
},
{
"id": "2304.03022"
},
{
"id": "2305.07609"
},
{
"id": "2209.08060"
},
{
"id": "2209.07562"
},
{
"id": "2304.09542"
},
{
"id": "2303.14524"
},
{
"id": "2305.15673"
},
{
"id": "2303.18223"
},
{
"id": "2304.10149"
},
{
"id": "1908.08167"
},
{
"id": "1909.10351"
},
{
"id": "2305.15036"
},
{
"id": "2305.12081"
},
{
"id": "2304.07862"
},
{
"id": "2305.03017"
},
{
"id": "2305.09858"
},
{
"id": "2305.06474"
},
{
"id": "2305.13731"
},
{
"id": "2304.03153"
},
{
"id": "2205.08084"
},
{
"id": "2106.09685"
},
{
"id": "2306.10702"
},
{
"id": "2306.02250"
},
{
"id": "2303.13835"
},
{
"id": "2305.14302"
},
{
"id": "2302.03735"
},
{
"id": "2306.02841"
},
{
"id": "2305.11206"
},
{
"id": "2203.15876"
},
{
"id": "2305.07622"
},
{
"id": "2306.10933"
},
{
"id": "2305.06569"
},
{
"id": "2206.06190"
}
] |
2306.05949
| 80 |
[27] N. Beniach and I. Hogarth. State of AI Report 2022. URL https://www.stateof.ai/.
[28] C. L. Bennett, C. Gleason, M. K. Scheuerman, J. P. Bigham, A. Guo, and A. To. âItâs Complicatedâ: Negotiating Accessibility and (Mis)Representation in Image Descriptions of Race, Gender, and Disability. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, pages 1â19. ACM, 2021-05-06. ISBN 978-1-4503-8096-6. doi: 10. 1145/3411764.3445498. URL https://dl.acm.org/doi/10.1145/3411764.3445498.
[29] J. Berg, M. Furrer, E. Harmon, U. Rani, and M. S. Silberman. Digital Labour Platforms and the Future of Work: Towards Decent Work in the Online World. International Labour Organization, 2018. ISBN 978-92-2-031024-3.
|
2306.05949#80
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06070
| 80 |
# C.2 Model Implementation Details
Candidate Generation. We use the Cross-Encoder implementation from Sentence-Transformers 4 and use DeBERTa as the backbone model. More specifically, we use DeBERTa-v3-base 5 for our experiments.
Action Prediction. We use the Seq2Seq model implementation from Transformers [38]. We experiment with the base 6, large 7 and xl 8 versions of Flan-T5 [10].
4https://www.sbert.net/examples/applications/cross-encoder/README.html 5https://huggingface.co/microsoft/deberta-v3-base 6https://huggingface.co/google/flan-t5-base 7https://huggingface.co/google/flan-t5-large 8https://huggingface.co/google/flan-t5-xl
20
Table 4: Hyperparameters used in experiments.
|
2306.06070#80
|
Mind2Web: Towards a Generalist Agent for the Web
|
We introduce Mind2Web, the first dataset for developing and evaluating
generalist agents for the web that can follow language instructions to complete
complex tasks on any website. Existing datasets for web agents either use
simulated websites or only cover a limited set of websites and tasks, thus not
suitable for generalist web agents. With over 2,000 open-ended tasks collected
from 137 websites spanning 31 domains and crowdsourced action sequences for the
tasks, Mind2Web provides three necessary ingredients for building generalist
web agents: 1) diverse domains, websites, and tasks, 2) use of real-world
websites instead of simulated and simplified ones, and 3) a broad spectrum of
user interaction patterns. Based on Mind2Web, we conduct an initial exploration
of using large language models (LLMs) for building generalist web agents. While
the raw HTML of real-world websites are often too large to be fed to LLMs, we
show that first filtering it with a small LM significantly improves the
effectiveness and efficiency of LLMs. Our solution demonstrates a decent level
of performance, even on websites or entire domains the model has never seen
before, but there is still a substantial room to improve towards truly
generalizable agents. We open-source our dataset, model implementation, and
trained models (https://osu-nlp-group.github.io/Mind2Web) to facilitate further
research on building a generalist agent for the web.
|
http://arxiv.org/pdf/2306.06070
|
Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samuel Stevens, Boshi Wang, Huan Sun, Yu Su
|
cs.CL
|
Website: https://osu-nlp-group.github.io/Mind2Web. Updated with
supplementary material. NeurIPS'23 Spotlight
| null |
cs.CL
|
20230609
|
20231209
|
[] |
2306.06283
| 80 |
[7] Jablonka, K. M.; Ongari, D.; Moosavi, S. M.; Smit, B. Big-Data Science in Porous Mate- rials: Materials Genomics and Machine Learning. Chem. Rev. 2020, 120, 8066â8129.
[8] Shi, J.; Quevillon, M. J.; Amorim Valen¸ca, P. H.; Whitmer, J. K. Predicting Adhesive Free Energies of PolymerâSurface Interactions with Machine Learning. ACS Appl. Mater. Interfaces 2022, 14, 37161â37169.
[9] Shi, J.; Albreiki, F.; Col´on, Y. J.; Srivastava, S.; Whitmer, J. K. Transfer Learning Facil- itates the Prediction of PolymerâSurface Adhesion Strength. J. Chem. Theory Comput. 2023,
[10] No´e, F.; Tkatchenko, A.; M¨uller, K.-R.; Clementi, C. Machine Learning for Molecular Simulation. Annu. Rev. Phys. Chem. 2020, 71, 361â390.
|
2306.06283#80
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 81 |
Figure 17: The screenshot of MT-bench data collection. We show an instruction similar to the prompt we give to GPT-4. We present questions from MT-bench and answers from two random anonymous assistants and ask which one is better. We present the first-turn conversation and ask humans to vote, then repeat the same procedure for the second-turn. A user can skip up to 5 questions if they are not confident. For some questions (e.g., math, reasoning), they can also see a reference solution.
|
2306.05685#81
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05715
| 81 |
[77] Sami Sarsa, Jesper Pettersson, and Arto Hellas. 2022. How to Help to Ask for Help? Help Request Prompt Structure Inï¬uence on Help Request Quantity and Course Retention. In 2022 IEEE Frontiers in Education Conference (FIE). IEEE, 1â9. [78] Daniel Seamark and Lynne Gabriel. 2018. Barriers to support: a qualitative ex- ploration into the help-seeking and avoidance factors of young adults. British J. of Guidance & Counselling 46, 1 (2018).
ICER â23 V1, August 7â11, 2023, Chicago, IL, USA
[79] Otto Seppälä, Petri Ihantola, Essi Isohanni, Juha Sorva, and Arto Vihavainen. 2015. Do we know how diï¬cult the rainfall problem is?. In Proc. of the 15th Koli Calling Conf. on Computing Education Research. 87â96.
[80] Rebecca Smith and Scott Rixner. 2019. The error landscape: Characterizing the mistakes of novice programmers. In Proceedings of the 50th ACM technical sym- posium on computer science education. 538â544.
|
2306.05715#81
|
Exploring the Responses of Large Language Models to Beginner Programmers' Help Requests
|
Background and Context: Over the past year, large language models (LLMs) have
taken the world by storm. In computing education, like in other walks of life,
many opportunities and threats have emerged as a consequence.
Objectives: In this article, we explore such opportunities and threats in a
specific area: responding to student programmers' help requests. More
specifically, we assess how good LLMs are at identifying issues in problematic
code that students request help on.
Method: We collected a sample of help requests and code from an online
programming course. We then prompted two different LLMs (OpenAI Codex and
GPT-3.5) to identify and explain the issues in the students' code and assessed
the LLM-generated answers both quantitatively and qualitatively.
Findings: GPT-3.5 outperforms Codex in most respects. Both LLMs frequently
find at least one actual issue in each student program (GPT-3.5 in 90% of the
cases). Neither LLM excels at finding all the issues (GPT-3.5 finding them 57%
of the time). False positives are common (40% chance for GPT-3.5). The advice
that the LLMs provide on the issues is often sensible. The LLMs perform better
on issues involving program logic rather than on output formatting. Model
solutions are frequently provided even when the LLM is prompted not to. LLM
responses to prompts in a non-English language are only slightly worse than
responses to English prompts.
Implications: Our results continue to highlight the utility of LLMs in
programming education. At the same time, the results highlight the
unreliability of LLMs: LLMs make some of the same mistakes that students do,
perhaps especially when formatting output as required by automated assessment
systems. Our study informs teachers interested in using LLMs as well as future
efforts to customize LLMs for the needs of programming education.
|
http://arxiv.org/pdf/2306.05715
|
Arto Hellas, Juho Leinonen, Sami Sarsa, Charles Koutcheme, Lilja Kujanpää, Juha Sorva
|
cs.CY, cs.AI, cs.CL, cs.HC, cs.SE
|
13 pages, 1 figure. To be published in Proceedings of the 2023 ACM
Conference on International Computing Education Research V.1 (ICER '23 V1)
| null |
cs.CY
|
20230609
|
20230609
|
[
{
"id": "2004.09456"
},
{
"id": "2302.07427"
},
{
"id": "2203.02155"
},
{
"id": "2304.02491"
},
{
"id": "2211.04715"
},
{
"id": "2306.02608"
},
{
"id": "2303.08774"
},
{
"id": "2304.03938"
}
] |
2306.05783
| 81 |
3https://huggingface.co/bigscience/bloom-560m 4https://huggingface.co/bigscience/bloomz-560m 5https://huggingface.co/EleutherAI/pythia-1b 6https://huggingface.co/bigscience/bloom-1b7 7https://huggingface.co/bigscience/bloomz-1b7 8https://huggingface.co/databricks/dolly-v2-3b 9https://huggingface.co/EleutherAI/pythia-2.8b 10https://huggingface.co/bigscience/bloom-3b 11https://huggingface.co/bigscience/bloomz-3b 12https://huggingface.co/stabilityai/stablelm-tuned-alpha-3b 13https://huggingface.co/THUDM/chatglm-6b 14https://github.com/xionghonglin/DoctorGLM 15https://huggingface.co/databricks/dolly-v2-7b
|
2306.05783#81
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05949
| 81 |
[30] BigScience Workshop:, T. L. Scao, A. Fan, C. Akiki, E. Pavlick, S. Ili´c, D. Hesslow, R. Castagné, A. S. Luccioni, F. Yvon, M. Gallé, J. Tow, A. M. Rush, S. Biderman, A. Webson, P. S. Ammanamanchi, T. Wang, B. Sagot, N. Muennighoff, A. V. del Moral, O. Ruwase, R. Bawden, S. Bekman, A. McMillan-Major, I. Beltagy, H. Nguyen, L. Saulnier, S. Tan, P. O. Suarez, V. Sanh, H. Laurençon, Y. Jernite, J. Launay, M. Mitchell, C. Raffel, A. Gokaslan, A. Simhi, A. Soroa, A. F. Aji, A. Alfassy, A. Rogers, A. K. Nitzav, C. Xu, C. Mou, C. Emezue, C. Klamm, C. Leong, D. van Strien, D. I.
|
2306.05949#81
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06070
| 81 |
20
Table 4: Hyperparameters used in experiments.
Method Model Hyperparamerters Candidate Generation deberta-v3-base batch_size:32, epoch:5, learning_rate:3eâ5 Action Prediction Generation flan-t5-base batch_size:32, epoch:5, learning_rate:5eâ5 MINDACT flan-t5-base, flan-t5-large, flan-t5-xl gpt-3.5-turbo, gpt-4 batch_size:32, epoch:5, learning_rate:5eâ5
Table 5: Step Success Rate for Flan-T5 models with different groups of options. Here we shown mean and standard deviation of 5 runs with different random seeds.
Flan-T5B Flan-T5L Flan-T5XL 41.5±0.7 49.9±0.2 51.9±0.8 30.0±0.8 35.7±0.5 39.5±0.2 31.3±0.5 36.7±0.3 39.6±0.2
|
2306.06070#81
|
Mind2Web: Towards a Generalist Agent for the Web
|
We introduce Mind2Web, the first dataset for developing and evaluating
generalist agents for the web that can follow language instructions to complete
complex tasks on any website. Existing datasets for web agents either use
simulated websites or only cover a limited set of websites and tasks, thus not
suitable for generalist web agents. With over 2,000 open-ended tasks collected
from 137 websites spanning 31 domains and crowdsourced action sequences for the
tasks, Mind2Web provides three necessary ingredients for building generalist
web agents: 1) diverse domains, websites, and tasks, 2) use of real-world
websites instead of simulated and simplified ones, and 3) a broad spectrum of
user interaction patterns. Based on Mind2Web, we conduct an initial exploration
of using large language models (LLMs) for building generalist web agents. While
the raw HTML of real-world websites are often too large to be fed to LLMs, we
show that first filtering it with a small LM significantly improves the
effectiveness and efficiency of LLMs. Our solution demonstrates a decent level
of performance, even on websites or entire domains the model has never seen
before, but there is still a substantial room to improve towards truly
generalizable agents. We open-source our dataset, model implementation, and
trained models (https://osu-nlp-group.github.io/Mind2Web) to facilitate further
research on building a generalist agent for the web.
|
http://arxiv.org/pdf/2306.06070
|
Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samuel Stevens, Boshi Wang, Huan Sun, Yu Su
|
cs.CL
|
Website: https://osu-nlp-group.github.io/Mind2Web. Updated with
supplementary material. NeurIPS'23 Spotlight
| null |
cs.CL
|
20230609
|
20231209
|
[] |
2306.06283
| 81 |
[11] Batzner, S.; Musaelian, A.; Sun, L.; Geiger, M.; Mailoa, J. P.; Kornbluth, M.; Molinari, N.; Smidt, T. E.; Kozinsky, B. E(3)equivariant graph neural networks for data eï¬cient and accurate interatomic potentials. Nat. Commun. 2022, 13, 2453.
[12] Sanchez-Lengeling, B.; Aspuru-Guzik, A. Inverse molecular design using machine learning: Generative models for matter engineering. Science 2018, 361, 360â365.
[13] Gonthier, J. F.; Steinmann, S. N.; Wodrich, M. D.; Corminboeuf, C. Quantiï¬cation of âfuzzyâ chemical concepts: a computational perspective. Chem. Soc. Rev. 2012, 41, 4671.
[14] Weininger, D. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. J. Chem. Inf. Comput. Sci. 1988, 28, 31â36.
[15] Jablonka, K. M.; Patiny, L.; Smit, B. Making the collective knowledge of chemistry open and machine actionable. Nat. Chem. 2022, 14, 365â376.
|
2306.06283#81
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 82 |
Instruction Please act as an impartial judge and evaluate the quality of the responses provided by two Al assistants to the user question displayed below. You should choose the assistant that follows the userâs instructions and answers the userâs question better. Your evaluation should consider factors such as the helpfulness, relevance, accuracy, depth, creativity, and level of detail of their responses. Begin your evaluation by comparing the two responses and provide a short explanation. Avoid any positional biases and ensure that the order in which the responses were presented does not influence your decision. Do not allow the length of the responses to influence your evaluation. Do not favor certain names of the assistants. Be as objective as possible. Finally, indicate your verdict by clicking one button. Please use the tie button sparsely. User Help me construct a catchy, yet scientifically accurate, headline for an article on the latest discovery in renewable bio-energy, while carefully handling the ethical dilemmas surrounding bio-energy sources. Propose 4 options. Assistant A Assistant B 1, âRevolutionizing Renewable Energy: New Breakthroughs in Bio-Energy Production While 1, Breakthrough in Renewable Bio-Energy
|
2306.05685#82
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05715
| 82 |
[81] Elliot Soloway, Jeï¬rey G. Bonar, and Kate Ehrlich. 1983. Cognitive strategies and looping constructs: An empirical study. Commun. ACM 26, 11 (1983), 853â860. https://doi.org/10.1145/182.358436
[82] Elliot Soloway, Kate Ehrlich, Jeï¬rey G. Bonar, and Judith Greenspan. 1982. What do novices know about programming? In Directions in HumanâComputer In- teractions, Albert Badre and Ben Shneiderman (Eds.). Vol. 6. Ablex Publishing, 27â54.
[83] James C Spohrer and Elliot Soloway. 1986. Novice mistakes: Are the folk wis- doms correct? Commun. ACM 29, 7 (1986), 624â632.
[84] Priyan Vaithilingam, Tianyi Zhang, and Elena L Glassman. 2022. Expectation vs. Experience: Evaluating the Usability of Code Generation Tools Powered by Large Language Models. In CHI Conf. on Human Factors in Computing Systems Extended Abstracts. 1â7.
|
2306.05715#82
|
Exploring the Responses of Large Language Models to Beginner Programmers' Help Requests
|
Background and Context: Over the past year, large language models (LLMs) have
taken the world by storm. In computing education, like in other walks of life,
many opportunities and threats have emerged as a consequence.
Objectives: In this article, we explore such opportunities and threats in a
specific area: responding to student programmers' help requests. More
specifically, we assess how good LLMs are at identifying issues in problematic
code that students request help on.
Method: We collected a sample of help requests and code from an online
programming course. We then prompted two different LLMs (OpenAI Codex and
GPT-3.5) to identify and explain the issues in the students' code and assessed
the LLM-generated answers both quantitatively and qualitatively.
Findings: GPT-3.5 outperforms Codex in most respects. Both LLMs frequently
find at least one actual issue in each student program (GPT-3.5 in 90% of the
cases). Neither LLM excels at finding all the issues (GPT-3.5 finding them 57%
of the time). False positives are common (40% chance for GPT-3.5). The advice
that the LLMs provide on the issues is often sensible. The LLMs perform better
on issues involving program logic rather than on output formatting. Model
solutions are frequently provided even when the LLM is prompted not to. LLM
responses to prompts in a non-English language are only slightly worse than
responses to English prompts.
Implications: Our results continue to highlight the utility of LLMs in
programming education. At the same time, the results highlight the
unreliability of LLMs: LLMs make some of the same mistakes that students do,
perhaps especially when formatting output as required by automated assessment
systems. Our study informs teachers interested in using LLMs as well as future
efforts to customize LLMs for the needs of programming education.
|
http://arxiv.org/pdf/2306.05715
|
Arto Hellas, Juho Leinonen, Sami Sarsa, Charles Koutcheme, Lilja Kujanpää, Juha Sorva
|
cs.CY, cs.AI, cs.CL, cs.HC, cs.SE
|
13 pages, 1 figure. To be published in Proceedings of the 2023 ACM
Conference on International Computing Education Research V.1 (ICER '23 V1)
| null |
cs.CY
|
20230609
|
20230609
|
[
{
"id": "2004.09456"
},
{
"id": "2302.07427"
},
{
"id": "2203.02155"
},
{
"id": "2304.02491"
},
{
"id": "2211.04715"
},
{
"id": "2306.02608"
},
{
"id": "2303.08774"
},
{
"id": "2304.03938"
}
] |
2306.05783
| 82 |
14https://github.com/xionghonglin/DoctorGLM 15https://huggingface.co/databricks/dolly-v2-7b 16https://huggingface.co/h2oai/h2ogpt-oig-oasst1-512-6.9b 17https://huggingface.co/EleutherAI/pythia-6.9b 18https://huggingface.co/tatsu-lab/alpaca-7b-wdiff 19https://huggingface.co/tloen/alpaca-lora-7b 20https://huggingface.co/project-baize/baize-lora-7B 21https://huggingface.co/project-baize/baize-healthcare-lora-7B
|
2306.05783#82
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05817
| 82 |
Feature Engineering GReaT [Borisov et al., 2023] GPT2-medium (355M) FFT N/A Tabular GENRE [Liu et al., 2023c] ChatGPT Frozen Retrieval Sequential RS News AnyPredict [Wang et al., 2023] ChatGPT Frozen N/A Tabular LLM4KGC [Chen et al., 2023] PaLM (540B) ChatGPT Frozen N/A E-commerce TagGPT [Li et al., 2023a] ChatGPT Frozen Item Tagging Food Video ICPC [Christakopoulou et al., 2023] LaMDA (137B) FFT/PT User Proï¬ling N/A DPLLM [Carranza et al., 2023] T5-XL (3B) FFT Retrieval Privacy Web Search KAR [Xi et al., 2023b] ChatGPT Frozen CTR Prediction Movie MINT [Petrov and Macdonald, 2023] GPT-3 (175B) Frozen Narrative RS POI Feature Encoder U-BERT [Qiu et al., 2021] BERT-base (110M) FFT Rating Prediction Business E-commerce UNBERT [Zhang et al., 2021a] BERT-base (110M) FFT
|
2306.05817#82
|
How Can Recommender Systems Benefit from Large Language Models: A Survey
|
Recommender systems (RS) play important roles to match users' information
needs for Internet applications. In natural language processing (NLP) domains,
large language model (LLM) has shown astonishing emergent abilities (e.g.,
instruction following, reasoning), thus giving rise to the promising research
direction of adapting LLM to RS for performance enhancements and user
experience improvements. In this paper, we conduct a comprehensive survey on
this research direction from an application-oriented view. We first summarize
existing research works from two orthogonal perspectives: where and how to
adapt LLM to RS. For the "WHERE" question, we discuss the roles that LLM could
play in different stages of the recommendation pipeline, i.e., feature
engineering, feature encoder, scoring/ranking function, and pipeline
controller. For the "HOW" question, we investigate the training and inference
strategies, resulting in two fine-grained taxonomy criteria, i.e., whether to
tune LLMs or not, and whether to involve conventional recommendation model
(CRM) for inference. Detailed analysis and general development trajectories are
provided for both questions, respectively. Then, we highlight key challenges in
adapting LLM to RS from three aspects, i.e., efficiency, effectiveness, and
ethics. Finally, we summarize the survey and discuss the future prospects. We
also actively maintain a GitHub repository for papers and other related
resources in this rising direction:
https://github.com/CHIANGEL/Awesome-LLM-for-RecSys.
|
http://arxiv.org/pdf/2306.05817
|
Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, Weinan Zhang
|
cs.IR, cs.AI
|
15 pages; 3 figures; summarization table in appendix
| null |
cs.IR
|
20230609
|
20230628
|
[
{
"id": "2302.13971"
},
{
"id": "1810.04805"
},
{
"id": "2304.05263"
},
{
"id": "2305.07001"
},
{
"id": "2305.11700"
},
{
"id": "2305.06566"
},
{
"id": "2305.15756"
},
{
"id": "2105.08318"
},
{
"id": "2304.03879"
},
{
"id": "2303.08559"
},
{
"id": "1703.04247"
},
{
"id": "2206.07682"
},
{
"id": "2305.07961"
},
{
"id": "2305.05973"
},
{
"id": "2305.15498"
},
{
"id": "2306.11114"
},
{
"id": "2305.04518"
},
{
"id": "2305.00447"
},
{
"id": "2305.02182"
},
{
"id": "2305.08845"
},
{
"id": "2305.12090"
},
{
"id": "2212.10403"
},
{
"id": "2304.03022"
},
{
"id": "2305.07609"
},
{
"id": "2209.08060"
},
{
"id": "2209.07562"
},
{
"id": "2304.09542"
},
{
"id": "2303.14524"
},
{
"id": "2305.15673"
},
{
"id": "2303.18223"
},
{
"id": "2304.10149"
},
{
"id": "1908.08167"
},
{
"id": "1909.10351"
},
{
"id": "2305.15036"
},
{
"id": "2305.12081"
},
{
"id": "2304.07862"
},
{
"id": "2305.03017"
},
{
"id": "2305.09858"
},
{
"id": "2305.06474"
},
{
"id": "2305.13731"
},
{
"id": "2304.03153"
},
{
"id": "2205.08084"
},
{
"id": "2106.09685"
},
{
"id": "2306.10702"
},
{
"id": "2306.02250"
},
{
"id": "2303.13835"
},
{
"id": "2305.14302"
},
{
"id": "2302.03735"
},
{
"id": "2306.02841"
},
{
"id": "2305.11206"
},
{
"id": "2203.15876"
},
{
"id": "2305.07622"
},
{
"id": "2306.10933"
},
{
"id": "2305.06569"
},
{
"id": "2206.06190"
}
] |
2306.05949
| 82 |
C. Xu, C. Mou, C. Emezue, C. Klamm, C. Leong, D. van Strien, D. I. Adelani, D. Radev, E. G. Ponferrada, E. Lev- kovizh, E. Kim, E. B. Natan, F. D. Toni, G. Dupont, G. Kruszewski, G. Pistilli, H. Elsahar, H. Benyamina, H. Tran, I. Yu, I. Abdulmumin, I. Johnson, I. Gonzalez-Dios, J. de la Rosa, J. Chim, J. Dodge, J. Zhu, J. Chang, J. Frohberg, J. Tobing, J. Bhattacharjee, K. Almubarak, K. Chen, K. Lo, L. V. Werra, L. Weber, L. Phan, L. B. allal, L. Tanguy, M. Dey, M. R. Muñoz, M. Masoud, M. Grandury, M. Šaško, M. Huang, M. Coavoux, M. Singh, M. T.-J. Jiang, M. C.
|
2306.05949#82
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06070
| 82 |
LLM In-context Learning. We use the OpenAI API for in-context learning with LLMs. We experiment with two versions of GPT models: gpt-3.5-turbo and gpt-4. We include three demonstration examples for in-context learning. The complete prompt is shown in Table 8.
Training Details. The flan-t5-xl and flan-t5-large models are trained on servers with 4*A100 80GB cards provided by Ohio Supercomputer Center [6]. All other models are trained with single A6000 48GB cards.
Please see Table 4 for all hyperparameters used in our experiments.
# D Additional Results
# D.1 Effect of Random Grouping Elements for Action Prediction
For both training and inference, we shuffle the elements in the webpage and randomly group them into multi-choice questions. The model might give different predictions when presented with different sets of choices, and leads to slightly different final evaluation scores. Here we show the average and standard deviation of 5 runs with different random seeds to show the effect of random grouping. As we can see from Table 5, the selection of choices only lead to small changes in overall performance with standard deviation less than 1 for all runs.
# D.2 Zero-shot Results for Flan-T5XL
|
2306.06070#82
|
Mind2Web: Towards a Generalist Agent for the Web
|
We introduce Mind2Web, the first dataset for developing and evaluating
generalist agents for the web that can follow language instructions to complete
complex tasks on any website. Existing datasets for web agents either use
simulated websites or only cover a limited set of websites and tasks, thus not
suitable for generalist web agents. With over 2,000 open-ended tasks collected
from 137 websites spanning 31 domains and crowdsourced action sequences for the
tasks, Mind2Web provides three necessary ingredients for building generalist
web agents: 1) diverse domains, websites, and tasks, 2) use of real-world
websites instead of simulated and simplified ones, and 3) a broad spectrum of
user interaction patterns. Based on Mind2Web, we conduct an initial exploration
of using large language models (LLMs) for building generalist web agents. While
the raw HTML of real-world websites are often too large to be fed to LLMs, we
show that first filtering it with a small LM significantly improves the
effectiveness and efficiency of LLMs. Our solution demonstrates a decent level
of performance, even on websites or entire domains the model has never seen
before, but there is still a substantial room to improve towards truly
generalizable agents. We open-source our dataset, model implementation, and
trained models (https://osu-nlp-group.github.io/Mind2Web) to facilitate further
research on building a generalist agent for the web.
|
http://arxiv.org/pdf/2306.06070
|
Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samuel Stevens, Boshi Wang, Huan Sun, Yu Su
|
cs.CL
|
Website: https://osu-nlp-group.github.io/Mind2Web. Updated with
supplementary material. NeurIPS'23 Spotlight
| null |
cs.CL
|
20230609
|
20231209
|
[] |
2306.06283
| 82 |
[16] Bommasani, R. et al. On the Opportunities and Risks of Foundation Models. CoRR 2021, abs/2108.07258 .
17 Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; Polosukhin, I. Attention is all you need. Advances in neural information processing systems 2017, 30.
[18] Chowdhery, A. et al. PaLM: Scaling Language Modeling with Pathways. arXiv preprint:
30
Arxiv-2204.02311. 2022.
[19] Hoï¬mann, J. et al. Training Compute-Optimal Large Language Models. arXiv preprint: Arxiv-2203.15556. 2022.
[20] Brown, T.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J. D.; Dhariwal, P.; Neelakan- tan, A.; Shyam, P.; Sastry, G.; Askell, A., et al. Language models are few-shot learners. Advances in neural information processing systems 2020, 33, 1877â1901.
|
2306.06283#82
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 83 |
4 options. Assistant A Assistant B 1, âRevolutionizing Renewable Energy: New Breakthroughs in Bio-Energy Production While 1, Breakthrough in Renewable Bio-Energy Addressing Ethical Concerns 2. Renewable Bio-Energy to Power Cities for the Next 100 Years 2. âBio-Energy for a Sustainable Future: Balancing Environmental Benefits with Ethical Considerationsâ 3. Bio-Energy Breakthrough Gives Hope to Earthâs Future 4. Bio-Energy Breakthrough Could Save Earthâs Environment 3, âEco-Friendly and Ethical Bio-Energy: How Science is Paving the Way for a Greener Futureâ 4. âThe Ethics of Bio-Energy: A Look at the Latest Discoveries in Sustainable Energy Productionâ NOTICE Here is a judgment that disagrees with your choice. Is it reasonable? Judgment Assistant Aâs response provides four headlines that are catchy, scientifically accurate, and address the ethical dilemmas surrounding bio-energy sources. On the other hand, Assistant Bâs response does not address the ethical dilemmas in any of the proposed headlines. Therefore, Assistant Aâs response is better. My final verdict is: ([A]] Reasonable; | want to change my choice. Reasonable; | still keep
|
2306.05685#83
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05715
| 83 |
[85] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Åukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems 30 (2017).
[86] Arto Vihavainen, Juha Helminen, and Petri Ihantola. 2014. How novices tackle their ï¬rst lines of code in an ide: Analysis of programming session traces. In Pro- ceedings of the 14th koli calling international conference on computing education research. 109â116.
[87] Arto Vihavainen, Craig S Miller, and Amber Settle. 2015. Beneï¬ts of self- explanation in introductory programming. In Proceedings of the 46th ACM Tech- nical Symposium on Computer Science Education. 284â289.
[88] Christopher Watson, Frederick WB Li, and Jamie L Godwin. 2013. Predicting performance in an introductory programming course by logging and analyzing student programming behavior. In 2013 IEEE 13th international conference on ad- vanced learning technologies. IEEE, 319â323.
|
2306.05715#83
|
Exploring the Responses of Large Language Models to Beginner Programmers' Help Requests
|
Background and Context: Over the past year, large language models (LLMs) have
taken the world by storm. In computing education, like in other walks of life,
many opportunities and threats have emerged as a consequence.
Objectives: In this article, we explore such opportunities and threats in a
specific area: responding to student programmers' help requests. More
specifically, we assess how good LLMs are at identifying issues in problematic
code that students request help on.
Method: We collected a sample of help requests and code from an online
programming course. We then prompted two different LLMs (OpenAI Codex and
GPT-3.5) to identify and explain the issues in the students' code and assessed
the LLM-generated answers both quantitatively and qualitatively.
Findings: GPT-3.5 outperforms Codex in most respects. Both LLMs frequently
find at least one actual issue in each student program (GPT-3.5 in 90% of the
cases). Neither LLM excels at finding all the issues (GPT-3.5 finding them 57%
of the time). False positives are common (40% chance for GPT-3.5). The advice
that the LLMs provide on the issues is often sensible. The LLMs perform better
on issues involving program logic rather than on output formatting. Model
solutions are frequently provided even when the LLM is prompted not to. LLM
responses to prompts in a non-English language are only slightly worse than
responses to English prompts.
Implications: Our results continue to highlight the utility of LLMs in
programming education. At the same time, the results highlight the
unreliability of LLMs: LLMs make some of the same mistakes that students do,
perhaps especially when formatting output as required by automated assessment
systems. Our study informs teachers interested in using LLMs as well as future
efforts to customize LLMs for the needs of programming education.
|
http://arxiv.org/pdf/2306.05715
|
Arto Hellas, Juho Leinonen, Sami Sarsa, Charles Koutcheme, Lilja Kujanpää, Juha Sorva
|
cs.CY, cs.AI, cs.CL, cs.HC, cs.SE
|
13 pages, 1 figure. To be published in Proceedings of the 2023 ACM
Conference on International Computing Education Research V.1 (ICER '23 V1)
| null |
cs.CY
|
20230609
|
20230609
|
[
{
"id": "2004.09456"
},
{
"id": "2302.07427"
},
{
"id": "2203.02155"
},
{
"id": "2304.02491"
},
{
"id": "2211.04715"
},
{
"id": "2306.02608"
},
{
"id": "2303.08774"
},
{
"id": "2304.03938"
}
] |
2306.05783
| 83 |
14
# B Data Statistic
Table 4 enumerates the comprehensive statistical information for the given dataset. The âXiezhiâ column represents the total number of data points in the Xiezhi Benchmark. Furthermore, the âXiezhi-Specâ and âXiezhi-Interâ columns correspond to the data points in the Xiezhi-Specialty and Xiezhi-Interdiscipline benchmarks, respectively. Additionally, the âXiezhi-Trainâ column signifies the number of data points within the Xiezhi-Train dataset that pertain to few-shot learning.
Philosophy (Level 1 subject) Marxist Philosophy Chinese Philosophy Foreign Philosophy Logic Ethics Aesthetics Religion Philosophy of Science and Technology Applied Economics National Economics Regional Economics Finance (including: Taxation) Finance (including: Insurance) Industrial Economics International Trade Labour economics Statistics Quantitative Economics Defence Economics Theoretical Economics Political Economy History of Economic Thought Economic history Western economics World Economics 102 48 0 0 0 0 0 0 0 0 354 158 2 0 34 25 0 1 48 0 0 0 146 50 0 0 2 0 45 40 5 5 5 5 5 5 5 0 95 55 5 5 5 5 5 5 5 5 5 5 35 5 5 5 5 5
|
2306.05783#83
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05817
| 83 |
BERT-base (110M) FFT Rating Prediction Business E-commerce UNBERT [Zhang et al., 2021a] BERT-base (110M) FFT Sequential RS News PLM-NR [Wu et al., 2021] RoBERTa-base (125M) FFT Sequential RS News Pyramid-ERNIE [Zou et al., 2021] ERNIE (110M) FFT Ranking Web Search ERNIE-RS [Liu et al., 2021] ERNIE (110M) FFT Retrieval Web Search CTR-BERT [Muhamed et al., 2021] Customized BERT (1.5B) FFT CTR Prediction E-commerce ZESRec [Ding et al., 2021] BERT-base (110M) Frozen Sequential RS E-commerce UniSRec [Hou et al., 2022] BERT-base (110M) Frozen Sequential RS E-commerce PREC [Liu et al., 2022b] BERT-base (110M) FFT CTR Prediction News MM-Rec [Wu et al., 2022] BERT-base (110M) FFT Sequential RS News Tiny-NewsRec [Yu et al., 2022b] UniLMv2-base (110M) FFT Sequential RS News
|
2306.05817#83
|
How Can Recommender Systems Benefit from Large Language Models: A Survey
|
Recommender systems (RS) play important roles to match users' information
needs for Internet applications. In natural language processing (NLP) domains,
large language model (LLM) has shown astonishing emergent abilities (e.g.,
instruction following, reasoning), thus giving rise to the promising research
direction of adapting LLM to RS for performance enhancements and user
experience improvements. In this paper, we conduct a comprehensive survey on
this research direction from an application-oriented view. We first summarize
existing research works from two orthogonal perspectives: where and how to
adapt LLM to RS. For the "WHERE" question, we discuss the roles that LLM could
play in different stages of the recommendation pipeline, i.e., feature
engineering, feature encoder, scoring/ranking function, and pipeline
controller. For the "HOW" question, we investigate the training and inference
strategies, resulting in two fine-grained taxonomy criteria, i.e., whether to
tune LLMs or not, and whether to involve conventional recommendation model
(CRM) for inference. Detailed analysis and general development trajectories are
provided for both questions, respectively. Then, we highlight key challenges in
adapting LLM to RS from three aspects, i.e., efficiency, effectiveness, and
ethics. Finally, we summarize the survey and discuss the future prospects. We
also actively maintain a GitHub repository for papers and other related
resources in this rising direction:
https://github.com/CHIANGEL/Awesome-LLM-for-RecSys.
|
http://arxiv.org/pdf/2306.05817
|
Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, Weinan Zhang
|
cs.IR, cs.AI
|
15 pages; 3 figures; summarization table in appendix
| null |
cs.IR
|
20230609
|
20230628
|
[
{
"id": "2302.13971"
},
{
"id": "1810.04805"
},
{
"id": "2304.05263"
},
{
"id": "2305.07001"
},
{
"id": "2305.11700"
},
{
"id": "2305.06566"
},
{
"id": "2305.15756"
},
{
"id": "2105.08318"
},
{
"id": "2304.03879"
},
{
"id": "2303.08559"
},
{
"id": "1703.04247"
},
{
"id": "2206.07682"
},
{
"id": "2305.07961"
},
{
"id": "2305.05973"
},
{
"id": "2305.15498"
},
{
"id": "2306.11114"
},
{
"id": "2305.04518"
},
{
"id": "2305.00447"
},
{
"id": "2305.02182"
},
{
"id": "2305.08845"
},
{
"id": "2305.12090"
},
{
"id": "2212.10403"
},
{
"id": "2304.03022"
},
{
"id": "2305.07609"
},
{
"id": "2209.08060"
},
{
"id": "2209.07562"
},
{
"id": "2304.09542"
},
{
"id": "2303.14524"
},
{
"id": "2305.15673"
},
{
"id": "2303.18223"
},
{
"id": "2304.10149"
},
{
"id": "1908.08167"
},
{
"id": "1909.10351"
},
{
"id": "2305.15036"
},
{
"id": "2305.12081"
},
{
"id": "2304.07862"
},
{
"id": "2305.03017"
},
{
"id": "2305.09858"
},
{
"id": "2305.06474"
},
{
"id": "2305.13731"
},
{
"id": "2304.03153"
},
{
"id": "2205.08084"
},
{
"id": "2106.09685"
},
{
"id": "2306.10702"
},
{
"id": "2306.02250"
},
{
"id": "2303.13835"
},
{
"id": "2305.14302"
},
{
"id": "2302.03735"
},
{
"id": "2306.02841"
},
{
"id": "2305.11206"
},
{
"id": "2203.15876"
},
{
"id": "2305.07622"
},
{
"id": "2306.10933"
},
{
"id": "2305.06569"
},
{
"id": "2206.06190"
}
] |
2306.05949
| 83 |
Grandury, M. Šaško, M. Huang, M. Coavoux, M. Singh, M. T.-J. Jiang, M. C. Vu, M. A. Jauhar, M. Ghaleb, N. Subramani, N. Kassner, N. Khamis, O. Nguyen, O. Espejel, O. de Gibert, P. Villegas, P. Henderson, P. Colombo, P. Amuok, Q. Lhoest, R. Har- liman, R. Bommasani, R. L. López, R. Ribeiro, S. Osei, S. Pyysalo, S. Nagel, S. Bose, S. H. Muhammad, S. Sharma, S. Longpre, S. Nikpoor, S. Silberberg, S. Pai, S. Zink, T. T. Tor- rent, T. Schick, T. Thrush, V. Danchev, V. Nikoulina, V. Laippala, V. Lepercq, V. Prabhu, Z. Alyafeai, Z. Talat, A. Raja, B. Heinzerling, C. Si, D. E. Ta¸sar,
|
2306.05949#83
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06070
| 83 |
# D.2 Zero-shot Results for Flan-T5XL
Since Flan-T5 is tuned with multi-choice format, it can also do element selection in zero-shot. However, as we can see from Table 6, while the model still gets some elements correct, it is much lower compared to the fine-tuned model, and 3-shot GPT 3.5/4. This is expected, since Flan-T5 is not tuned for HTML and coding related tasks.
Table 6: Zero-shot element selection results for Flan-T5XL compared with the fine-tuned counterpart.
Cross-Task Cross-Website Cross-Domain Flan-T5XL Zero-Shot Flan-T5XL Fine-Tuned 10.8 52.0 7.8 38.9 11.7 39.6
21
Table 7: Step Success Rate for all methods on the 50 tasks subsets we used to evaluate GPT-4. Numbers in parentheses are the results on the full test set (same as Table 2)
|
2306.06070#83
|
Mind2Web: Towards a Generalist Agent for the Web
|
We introduce Mind2Web, the first dataset for developing and evaluating
generalist agents for the web that can follow language instructions to complete
complex tasks on any website. Existing datasets for web agents either use
simulated websites or only cover a limited set of websites and tasks, thus not
suitable for generalist web agents. With over 2,000 open-ended tasks collected
from 137 websites spanning 31 domains and crowdsourced action sequences for the
tasks, Mind2Web provides three necessary ingredients for building generalist
web agents: 1) diverse domains, websites, and tasks, 2) use of real-world
websites instead of simulated and simplified ones, and 3) a broad spectrum of
user interaction patterns. Based on Mind2Web, we conduct an initial exploration
of using large language models (LLMs) for building generalist web agents. While
the raw HTML of real-world websites are often too large to be fed to LLMs, we
show that first filtering it with a small LM significantly improves the
effectiveness and efficiency of LLMs. Our solution demonstrates a decent level
of performance, even on websites or entire domains the model has never seen
before, but there is still a substantial room to improve towards truly
generalizable agents. We open-source our dataset, model implementation, and
trained models (https://osu-nlp-group.github.io/Mind2Web) to facilitate further
research on building a generalist agent for the web.
|
http://arxiv.org/pdf/2306.06070
|
Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samuel Stevens, Boshi Wang, Huan Sun, Yu Su
|
cs.CL
|
Website: https://osu-nlp-group.github.io/Mind2Web. Updated with
supplementary material. NeurIPS'23 Spotlight
| null |
cs.CL
|
20230609
|
20231209
|
[] |
2306.06283
| 83 |
[21] Edwards, C. N.; Lai, T.; Ros, K.; Honke, G.; Ji, H. Translation between Molecules and Natural Language. Conference On Empirical Methods In Natural Language Processing 2022,
[22] Eloundou, T.; Manning, S.; Mishkin, P.; Rock, D. GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. arXiv preprint: Arxiv- 2303.10130 2023,
[23] Srivastava, A. et al. Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models. arXiv preprint: Arxiv-2206.04615. 2022.
[24] Bubeck, S.; Chandrasekaran, V.; Eldan, R.; Gehrke, J.; Horvitz, E.; Kamar, E.; Lee, P.; Lee, Y. T.; Li, Y.; Lundberg, S.; Nori, H.; Palangi, H.; Ribeiro, M. T.; Zhang, Y. Sparks of Artiï¬cial General Intelligence: Early experiments with GPT-4. arXiv preprint: Arxiv- 2303.12712 2023,
|
2306.06283#83
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05817
| 84 |
FFT Sequential RS News Tiny-NewsRec [Yu et al., 2022b] UniLMv2-base (110M) FFT Sequential RS News PTM4Tag [He et al., 2022] CodeBERT (125M) FFT Top-N RS posts TwHIN-BERT [Zhang et al., 2022] BERT-base (110M) FFT Social RS posts TransRec [Wang et al., 2022] BERT-base (110M) FFT Cross-domain RS Sequential RS News Video VQ-Rec [Hou et al., 2023a] BERT-base (110M) Frozen Sequential RS E-commerce IDRec vs MoRec [Yuan et al., 2023] BERT-base (110M) FFT Sequential RS News Video E-commerce TransRec [Fu et al., 2023a] RoBERTa-base (125M) LAT Cross-domain RS Sequential RS News Video E-commerce LSH [Rahmani et al., 2023] BERT-base (110M) FFT Top-N RS Code TCF [Li et al., 2023d] OPT-175B (175B) Frozen/FFT Sequential RS Top-N RS
|
2306.05817#84
|
How Can Recommender Systems Benefit from Large Language Models: A Survey
|
Recommender systems (RS) play important roles to match users' information
needs for Internet applications. In natural language processing (NLP) domains,
large language model (LLM) has shown astonishing emergent abilities (e.g.,
instruction following, reasoning), thus giving rise to the promising research
direction of adapting LLM to RS for performance enhancements and user
experience improvements. In this paper, we conduct a comprehensive survey on
this research direction from an application-oriented view. We first summarize
existing research works from two orthogonal perspectives: where and how to
adapt LLM to RS. For the "WHERE" question, we discuss the roles that LLM could
play in different stages of the recommendation pipeline, i.e., feature
engineering, feature encoder, scoring/ranking function, and pipeline
controller. For the "HOW" question, we investigate the training and inference
strategies, resulting in two fine-grained taxonomy criteria, i.e., whether to
tune LLMs or not, and whether to involve conventional recommendation model
(CRM) for inference. Detailed analysis and general development trajectories are
provided for both questions, respectively. Then, we highlight key challenges in
adapting LLM to RS from three aspects, i.e., efficiency, effectiveness, and
ethics. Finally, we summarize the survey and discuss the future prospects. We
also actively maintain a GitHub repository for papers and other related
resources in this rising direction:
https://github.com/CHIANGEL/Awesome-LLM-for-RecSys.
|
http://arxiv.org/pdf/2306.05817
|
Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, Weinan Zhang
|
cs.IR, cs.AI
|
15 pages; 3 figures; summarization table in appendix
| null |
cs.IR
|
20230609
|
20230628
|
[
{
"id": "2302.13971"
},
{
"id": "1810.04805"
},
{
"id": "2304.05263"
},
{
"id": "2305.07001"
},
{
"id": "2305.11700"
},
{
"id": "2305.06566"
},
{
"id": "2305.15756"
},
{
"id": "2105.08318"
},
{
"id": "2304.03879"
},
{
"id": "2303.08559"
},
{
"id": "1703.04247"
},
{
"id": "2206.07682"
},
{
"id": "2305.07961"
},
{
"id": "2305.05973"
},
{
"id": "2305.15498"
},
{
"id": "2306.11114"
},
{
"id": "2305.04518"
},
{
"id": "2305.00447"
},
{
"id": "2305.02182"
},
{
"id": "2305.08845"
},
{
"id": "2305.12090"
},
{
"id": "2212.10403"
},
{
"id": "2304.03022"
},
{
"id": "2305.07609"
},
{
"id": "2209.08060"
},
{
"id": "2209.07562"
},
{
"id": "2304.09542"
},
{
"id": "2303.14524"
},
{
"id": "2305.15673"
},
{
"id": "2303.18223"
},
{
"id": "2304.10149"
},
{
"id": "1908.08167"
},
{
"id": "1909.10351"
},
{
"id": "2305.15036"
},
{
"id": "2305.12081"
},
{
"id": "2304.07862"
},
{
"id": "2305.03017"
},
{
"id": "2305.09858"
},
{
"id": "2305.06474"
},
{
"id": "2305.13731"
},
{
"id": "2304.03153"
},
{
"id": "2205.08084"
},
{
"id": "2106.09685"
},
{
"id": "2306.10702"
},
{
"id": "2306.02250"
},
{
"id": "2303.13835"
},
{
"id": "2305.14302"
},
{
"id": "2302.03735"
},
{
"id": "2306.02841"
},
{
"id": "2305.11206"
},
{
"id": "2203.15876"
},
{
"id": "2305.07622"
},
{
"id": "2306.10933"
},
{
"id": "2305.06569"
},
{
"id": "2206.06190"
}
] |
2306.05949
| 84 |
Z. Alyafeai, Z. Talat, A. Raja, B. Heinzerling, C. Si, D. E. Ta¸sar, E. Salesky, S. J. Mielke, W. Y. Lee, A. Sharma, A. Santilli, A. Chaffin, A. Stiegler, D. Datta, E. Szczechla, G. Chhablani, H. Wang, H. Pandey, H. Strobelt, J. A. Fries, J. Rozen, L. Gao, L. Sutawika, M. S. Bari, M. S. Al-shaibani, M. Manica, N. Nayak, R. Teehan, S. Albanie, S. Shen, S. Ben-David, S. H. Bach, T. Kim, T. Bers, T. Fevry, T. Neeraj, U. Thakker, V. Raunak, X. Tang, Z.-X. Yong, Z. Sun, S. Brody, Y. Uri, H. Tojarieh, A. Roberts, H. W. Chung, J. Tae, J. Phang, O. Press, C. Li, D. Narayanan,
|
2306.05949#84
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06070
| 84 |
Flan-T5B Flan-T5L Flan-T5XL GPT-3.5 GPT-4 43.3 (41.0) 48.1 (50.3) 47.9 (52.0) 15.2 (17.4) 36.2 25.3 (29.5) 30.8 (35.3) 33.3 (38.9) 15.1 (16.2) 30.1 28.1 (31.6) 27.6 (37.3) 34.6 (39.6) 16.7 (18.6) 26.4
# Cross-Task Cross-Website Cross-Domain
# D.3 Results on the 50 task subsets
Due to budget constraint, we only run GPT-4 on 50 tasks for each setting. Here we show the step success rate results for other methods on the same 50 examples that GPT-4 is tested on, As we can see from Table 7, the results on the 50 tasks subsets are consistent with the results on the respective full test set, and the relative performance across methods and splits remains the same.
22
Table 8: Prompt for action prediction in MINDACT with GPT models. Only part of the HTML snippet is shown here to save space.
# Role
# Content
# system
# user
|
2306.06070#84
|
Mind2Web: Towards a Generalist Agent for the Web
|
We introduce Mind2Web, the first dataset for developing and evaluating
generalist agents for the web that can follow language instructions to complete
complex tasks on any website. Existing datasets for web agents either use
simulated websites or only cover a limited set of websites and tasks, thus not
suitable for generalist web agents. With over 2,000 open-ended tasks collected
from 137 websites spanning 31 domains and crowdsourced action sequences for the
tasks, Mind2Web provides three necessary ingredients for building generalist
web agents: 1) diverse domains, websites, and tasks, 2) use of real-world
websites instead of simulated and simplified ones, and 3) a broad spectrum of
user interaction patterns. Based on Mind2Web, we conduct an initial exploration
of using large language models (LLMs) for building generalist web agents. While
the raw HTML of real-world websites are often too large to be fed to LLMs, we
show that first filtering it with a small LM significantly improves the
effectiveness and efficiency of LLMs. Our solution demonstrates a decent level
of performance, even on websites or entire domains the model has never seen
before, but there is still a substantial room to improve towards truly
generalizable agents. We open-source our dataset, model implementation, and
trained models (https://osu-nlp-group.github.io/Mind2Web) to facilitate further
research on building a generalist agent for the web.
|
http://arxiv.org/pdf/2306.06070
|
Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samuel Stevens, Boshi Wang, Huan Sun, Yu Su
|
cs.CL
|
Website: https://osu-nlp-group.github.io/Mind2Web. Updated with
supplementary material. NeurIPS'23 Spotlight
| null |
cs.CL
|
20230609
|
20231209
|
[] |
2306.06283
| 84 |
[25] Schick, T.; Dwivedi-Yu, J.; Dess`ı, R.; Raileanu, R.; Lomeli, M.; Zettlemoyer, L.; Can- cedda, N.; Scialom, T. Toolformer: Language Models Can Teach Themselves to Use Tools. arXiv preprint: Arxiv-2302.04761 2023,
[26] Karpas, E. et al. MRKL Systems: A modular, neuro-symbolic architecture that combines large language models, external knowledge sources and discrete reasoning. arXiv preprint: Arxiv-2205.00445 2022,
[27] Shen, Y.; Song, K.; Tan, X.; Li, D.; Lu, W.; Zhuang, Y. HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in HuggingFace. arXiv preprint: Arxiv-2303.17580. 2023.
[28] White, A. paper-qa. https://github.com/whitead/paper-qa, 2022.
[29] Liu, J. LlamaIndex. 2022; https://github.com/jerryjliu/llama_index, last accessed 2023-05-30.
|
2306.06283#84
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 85 |
Figure 18: The screenshot of MT-bench data collection. When humanâs vote differs from GPT-4, we additionally show GPT-4âs judgment (red region in the screenshot) and ask the user to click one of the three buttons to decide whether GPT-4âs judgment is reasonable.
23
To invite participants, we obtained their consent by letting them sign an application form. We pay them $20 for judging 20 questions, which corresponds to an hourly rate of around $35. The participants are mostly graduate students from more than ten universities.
# C.2 Chatbot Arena
Figure 19 shows a screenshot of Chatbot Arena. Users are required to accept the terms of use, which obtain their consent and give us the right to release the conversation data. The instructions are shown at the top of the interface. This is a free website. We do not pay users and any user can use this platform without registration. More introductions and analyses can be found at https: //lmsys.org/blog/2023-05-03-arena/.
|
2306.05685#85
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 85 |
22https://huggingface.co/decapoda-research/llama-7b-hf 23https://huggingface.co/stabilityai/stablelm-tuned-alpha-7b 24https://github.com/lm-sys/FastChat#vicuna-weights 25https://huggingface.co/BelleGroup/BELLE-7B-0.2M 26https://huggingface.co/BelleGroup/BELLE-7B-0.6M 27https://huggingface.co/BelleGroup/BELLE-7B-1M 28https://huggingface.co/BelleGroup/BELLE-7B-2M 29https://huggingface.co/bigscience/bloom-7b1 30https://huggingface.co/bigscience/bloomz-7b1 31https://huggingface.co/bigscience/bloomz-7b1-mt 32https://huggingface.co/bigscience/bloomz-7b1-p3 33https://huggingface.co/databricks/dolly-v2-12b
|
2306.05783#85
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05817
| 85 |
Table 1 continued from previous page Model Name LLM Backbone LLM Tuning Strategy Scoring/Ranking Function (Item Scoring Task) LMRecSys [Zhang et al., 2021b] GPT2-XL (1.5B) FFT PTab [Liu et al., 2022a] BERT-base (110M) FFT UniTRec [Mao et al., 2023] BART (406M) FFT Prompt4NR [Zhang and Wang, 2023] BERT-base (110M) FFT RecFormer [Li et al., 2023b] LongFormer (149M) FFT TabLLM [Hegselmann et al., 2023] T0 (11B) T-FEW Zero-shot GPT [Sileo et al., 2022] GPT2-medium (355M) Frozen FLAN-T5 [Kang et al., 2023] FLAN-T5-XXL (11B) FFT BookGPT [Zhiyuli et al., 2023] ChatGPT Frozen TALLRec [Bao et al., 2023] LLaMA (7B) LoRA PBNR [Li et al., 2023f] T5-small (60M) FFT Scoring/Ranking Function (Item Generation Task)
|
2306.05817#85
|
How Can Recommender Systems Benefit from Large Language Models: A Survey
|
Recommender systems (RS) play important roles to match users' information
needs for Internet applications. In natural language processing (NLP) domains,
large language model (LLM) has shown astonishing emergent abilities (e.g.,
instruction following, reasoning), thus giving rise to the promising research
direction of adapting LLM to RS for performance enhancements and user
experience improvements. In this paper, we conduct a comprehensive survey on
this research direction from an application-oriented view. We first summarize
existing research works from two orthogonal perspectives: where and how to
adapt LLM to RS. For the "WHERE" question, we discuss the roles that LLM could
play in different stages of the recommendation pipeline, i.e., feature
engineering, feature encoder, scoring/ranking function, and pipeline
controller. For the "HOW" question, we investigate the training and inference
strategies, resulting in two fine-grained taxonomy criteria, i.e., whether to
tune LLMs or not, and whether to involve conventional recommendation model
(CRM) for inference. Detailed analysis and general development trajectories are
provided for both questions, respectively. Then, we highlight key challenges in
adapting LLM to RS from three aspects, i.e., efficiency, effectiveness, and
ethics. Finally, we summarize the survey and discuss the future prospects. We
also actively maintain a GitHub repository for papers and other related
resources in this rising direction:
https://github.com/CHIANGEL/Awesome-LLM-for-RecSys.
|
http://arxiv.org/pdf/2306.05817
|
Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, Weinan Zhang
|
cs.IR, cs.AI
|
15 pages; 3 figures; summarization table in appendix
| null |
cs.IR
|
20230609
|
20230628
|
[
{
"id": "2302.13971"
},
{
"id": "1810.04805"
},
{
"id": "2304.05263"
},
{
"id": "2305.07001"
},
{
"id": "2305.11700"
},
{
"id": "2305.06566"
},
{
"id": "2305.15756"
},
{
"id": "2105.08318"
},
{
"id": "2304.03879"
},
{
"id": "2303.08559"
},
{
"id": "1703.04247"
},
{
"id": "2206.07682"
},
{
"id": "2305.07961"
},
{
"id": "2305.05973"
},
{
"id": "2305.15498"
},
{
"id": "2306.11114"
},
{
"id": "2305.04518"
},
{
"id": "2305.00447"
},
{
"id": "2305.02182"
},
{
"id": "2305.08845"
},
{
"id": "2305.12090"
},
{
"id": "2212.10403"
},
{
"id": "2304.03022"
},
{
"id": "2305.07609"
},
{
"id": "2209.08060"
},
{
"id": "2209.07562"
},
{
"id": "2304.09542"
},
{
"id": "2303.14524"
},
{
"id": "2305.15673"
},
{
"id": "2303.18223"
},
{
"id": "2304.10149"
},
{
"id": "1908.08167"
},
{
"id": "1909.10351"
},
{
"id": "2305.15036"
},
{
"id": "2305.12081"
},
{
"id": "2304.07862"
},
{
"id": "2305.03017"
},
{
"id": "2305.09858"
},
{
"id": "2305.06474"
},
{
"id": "2305.13731"
},
{
"id": "2304.03153"
},
{
"id": "2205.08084"
},
{
"id": "2106.09685"
},
{
"id": "2306.10702"
},
{
"id": "2306.02250"
},
{
"id": "2303.13835"
},
{
"id": "2305.14302"
},
{
"id": "2302.03735"
},
{
"id": "2306.02841"
},
{
"id": "2305.11206"
},
{
"id": "2203.15876"
},
{
"id": "2305.07622"
},
{
"id": "2306.10933"
},
{
"id": "2305.06569"
},
{
"id": "2206.06190"
}
] |
2306.05949
| 85 |
Tojarieh, A. Roberts, H. W. Chung, J. Tae, J. Phang, O. Press, C. Li, D. Narayanan, H. Bourfoune, J. Casper, J. Rasley, M. Ryabinin, M. Mishra, M. Zhang, M. Shoeybi, M. Peyrounette, N. Patry, N. Tazi, O. Sanseviero, P. von Platen, P. Cornette, P. F. Lavallée, R. Lacroix, S. Rajbhandari, S. Gandhi, S. Smith, S. Requena, S. Patil, T. Dettmers, A. Baruwa, A. Singh, A. Cheveleva, A.-L. Ligozat, A. Subramonian, A. Névéol, C. Lovering, D. Garrette, D. Tunuguntla, E. Reiter, E. Taktasheva, E. Voloshina, E. Bogdanov, G. I. Winata, H. Schoelkopf, J.-C. Kalo, J. Novikova, J. Z. Forde, J. Clive, J.
|
2306.05949#85
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06070
| 85 |
22
Table 8: Prompt for action prediction in MINDACT with GPT models. Only part of the HTML snippet is shown here to save space.
# Role
# Content
# system
# user
You are a helpful assistant that is great at website design, navigation, and executing tasks for the user ``` <html> <div> <div> <a tock home page /> <button id=0 book a reservation. toggle open> <span> Book a reservation </span> </button> <button book a reservation. toggle open> </button> </div> <div> <select id=1 type> <option reservations true> Dine in </option> ... </html> ``` Based on the HTML webpage above, try to complete the following task: Task: Check for pickup restaurant available in Boston, NY on March 18, 5pm with just one guest Previous actions: None What should be the next action? Please select from the following choices (If the correct action is not in the page above, please select A. âNone of the aboveâ):
A. None of the above B. <button id=0 book a reservation. toggle open> <span> Book a C. <select id=1 type> <option reservations true> Dine in </option> <option D. <div id=2> <p> Celebrating and supporting leading women shaking up
assistant Answer: C.
user
|
2306.06070#85
|
Mind2Web: Towards a Generalist Agent for the Web
|
We introduce Mind2Web, the first dataset for developing and evaluating
generalist agents for the web that can follow language instructions to complete
complex tasks on any website. Existing datasets for web agents either use
simulated websites or only cover a limited set of websites and tasks, thus not
suitable for generalist web agents. With over 2,000 open-ended tasks collected
from 137 websites spanning 31 domains and crowdsourced action sequences for the
tasks, Mind2Web provides three necessary ingredients for building generalist
web agents: 1) diverse domains, websites, and tasks, 2) use of real-world
websites instead of simulated and simplified ones, and 3) a broad spectrum of
user interaction patterns. Based on Mind2Web, we conduct an initial exploration
of using large language models (LLMs) for building generalist web agents. While
the raw HTML of real-world websites are often too large to be fed to LLMs, we
show that first filtering it with a small LM significantly improves the
effectiveness and efficiency of LLMs. Our solution demonstrates a decent level
of performance, even on websites or entire domains the model has never seen
before, but there is still a substantial room to improve towards truly
generalizable agents. We open-source our dataset, model implementation, and
trained models (https://osu-nlp-group.github.io/Mind2Web) to facilitate further
research on building a generalist agent for the web.
|
http://arxiv.org/pdf/2306.06070
|
Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samuel Stevens, Boshi Wang, Huan Sun, Yu Su
|
cs.CL
|
Website: https://osu-nlp-group.github.io/Mind2Web. Updated with
supplementary material. NeurIPS'23 Spotlight
| null |
cs.CL
|
20230609
|
20231209
|
[] |
2306.06283
| 85 |
[29] Liu, J. LlamaIndex. 2022; https://github.com/jerryjliu/llama_index, last accessed 2023-05-30.
[30] Andrej Karpathy [@karpathy], The Hottest New Programming Language Is English. 2023; https://twitter.com/karpathy/status/1617979122625712128, last accessed 2023-05- 11.
31
[31] Hocky, G. M.; White, A. D. Natural language processing models that automate program- ming will transform chemistry research and teaching. Digital Discovery 2022, 1, 79â83.
[32] Jablonka, K. M.; Schwaller, P.; Ortega-Guerrero, A.; Smit, B. Is GPT-3 all you need for low-data discovery in chemistry? ChemRxiv preprint 10.26434/chemrxiv-2023-fw8n4 2023,
[33] White, A. D.; Hocky, G. M.; Gandhi, H. A.; Ansari, M.; Cox, S.; Wellawatte, G. P.; Sasmal, S.; Yang, Z.; Liu, K.; Singh, Y., et al. Assessment of chemistry knowledge in large language models that generate code. Digital Discovery 2023,
|
2306.06283#85
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 86 |
>< Chatbot Arena >< Rules © Chat with two anonymous models side-by-side and vote for which one is better! © Youcan do multiple rounds of conversations before voting, The names of the models will be revealed after your vote. Conversations with identity keywords (e.g., ChatGPT, Bard, Vicuna) or any votes after the names are revealed will not count towards the leaderboard. © Click âClear historyâ to start anew round. © [Blog] (GitHub) [Twitter] [Discord] Terms of use By using this service, users are required to agree to the following terms: The service is a research preview intended for non-commercial use only. It only provides limited safety measures and may generate offensive content. It must not be used for any illegal, harmful, violent, racist, or sexual purposes. The service collects user dialogue data and reserves the right to distribute it under a Creative Commons Attribution (CC-BY) license. The demo works better on desktop devices with a wide screen. Battle Please scroll down and start chatting. You can view a leaderboard of participating models in the fourth tab above labeled âLeaderboardâ or by clicking here. The models include both closed-source models and open-source models. Model A Model
|
2306.05685#86
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 86 |
33https://huggingface.co/databricks/dolly-v2-12b 34https://huggingface.co/h2oai/h2ogpt-oasst1-512-12b 35https://huggingface.co/OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5 36https://huggingface.co/EleutherAI/pythia-12b 37https://huggingface.co/project-baize/baize-lora-13B 38https://huggingface.co/decapoda-research/llama-13b-hf 39https://github.com/lm-sys/FastChat#vicuna-weights 40https://huggingface.co/fnlp/moss-moon-003-sft 41https://huggingface.co/fnlp/moss-moon-003-sft-plugin 42https://huggingface.co/EleutherAI/gpt-neox-20b 43https://huggingface.co/h2oai/h2ogpt-oasst1-512-20b 44https://huggingface.co/project-baize/baize-lora-30B
|
2306.05783#86
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05817
| 86 |
(7B) LoRA PBNR [Li et al., 2023f] T5-small (60M) FFT Scoring/Ranking Function (Item Generation Task) GPT4Rec [Li et al., 2023c] GPT2 (110M) FFT UP5 [Hua et al., 2023a] T5-base (223M) FFT VIP5 [Geng et al., 2023] T5-base (223M) LAT P5-ID [Hua et al., 2023b] T5-small (61M) FFT FaiRLLM [Zhang et al., 2023a] ChatGPT Frozen PALR [Chen, 2023] LLaMA (7B) FFT ChatGPT-3 [Hou et al., 2023b] ChatGPT Frozen AGR [Lin and Zhang, 2023] ChatGPT Frozen NIR [Wang and Lim, 2023] GPT-3 (175B) Frozen GPTRec [Petrov and Macdonald, 2023] GPT2-medium (355M) FFT ChatNews [Li et al., 2023g] ChatGPT Frozen Scoring/Ranking Function (Hybrid Task) RS Task Top-N RS N/A Sequential RS Sequential RS Sequential RS N/A Rating
|
2306.05817#86
|
How Can Recommender Systems Benefit from Large Language Models: A Survey
|
Recommender systems (RS) play important roles to match users' information
needs for Internet applications. In natural language processing (NLP) domains,
large language model (LLM) has shown astonishing emergent abilities (e.g.,
instruction following, reasoning), thus giving rise to the promising research
direction of adapting LLM to RS for performance enhancements and user
experience improvements. In this paper, we conduct a comprehensive survey on
this research direction from an application-oriented view. We first summarize
existing research works from two orthogonal perspectives: where and how to
adapt LLM to RS. For the "WHERE" question, we discuss the roles that LLM could
play in different stages of the recommendation pipeline, i.e., feature
engineering, feature encoder, scoring/ranking function, and pipeline
controller. For the "HOW" question, we investigate the training and inference
strategies, resulting in two fine-grained taxonomy criteria, i.e., whether to
tune LLMs or not, and whether to involve conventional recommendation model
(CRM) for inference. Detailed analysis and general development trajectories are
provided for both questions, respectively. Then, we highlight key challenges in
adapting LLM to RS from three aspects, i.e., efficiency, effectiveness, and
ethics. Finally, we summarize the survey and discuss the future prospects. We
also actively maintain a GitHub repository for papers and other related
resources in this rising direction:
https://github.com/CHIANGEL/Awesome-LLM-for-RecSys.
|
http://arxiv.org/pdf/2306.05817
|
Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, Weinan Zhang
|
cs.IR, cs.AI
|
15 pages; 3 figures; summarization table in appendix
| null |
cs.IR
|
20230609
|
20230628
|
[
{
"id": "2302.13971"
},
{
"id": "1810.04805"
},
{
"id": "2304.05263"
},
{
"id": "2305.07001"
},
{
"id": "2305.11700"
},
{
"id": "2305.06566"
},
{
"id": "2305.15756"
},
{
"id": "2105.08318"
},
{
"id": "2304.03879"
},
{
"id": "2303.08559"
},
{
"id": "1703.04247"
},
{
"id": "2206.07682"
},
{
"id": "2305.07961"
},
{
"id": "2305.05973"
},
{
"id": "2305.15498"
},
{
"id": "2306.11114"
},
{
"id": "2305.04518"
},
{
"id": "2305.00447"
},
{
"id": "2305.02182"
},
{
"id": "2305.08845"
},
{
"id": "2305.12090"
},
{
"id": "2212.10403"
},
{
"id": "2304.03022"
},
{
"id": "2305.07609"
},
{
"id": "2209.08060"
},
{
"id": "2209.07562"
},
{
"id": "2304.09542"
},
{
"id": "2303.14524"
},
{
"id": "2305.15673"
},
{
"id": "2303.18223"
},
{
"id": "2304.10149"
},
{
"id": "1908.08167"
},
{
"id": "1909.10351"
},
{
"id": "2305.15036"
},
{
"id": "2305.12081"
},
{
"id": "2304.07862"
},
{
"id": "2305.03017"
},
{
"id": "2305.09858"
},
{
"id": "2305.06474"
},
{
"id": "2305.13731"
},
{
"id": "2304.03153"
},
{
"id": "2205.08084"
},
{
"id": "2106.09685"
},
{
"id": "2306.10702"
},
{
"id": "2306.02250"
},
{
"id": "2303.13835"
},
{
"id": "2305.14302"
},
{
"id": "2302.03735"
},
{
"id": "2306.02841"
},
{
"id": "2305.11206"
},
{
"id": "2203.15876"
},
{
"id": "2305.07622"
},
{
"id": "2306.10933"
},
{
"id": "2305.06569"
},
{
"id": "2206.06190"
}
] |
2306.05949
| 86 |
Winata, H. Schoelkopf, J.-C. Kalo, J. Novikova, J. Z. Forde, J. Clive, J. Kasai, K. Kawamura, L. Hazan, M. Carpuat, M. Clinciu, N. Kim, N. Cheng, O. Serikov, O. Antverg, O. van der Wal, R. Zhang, R. Zhang, S. Gehrmann, S. Mirkin, S. Pais, T. Shavrina, T. Scialom, T. Yun, T. Limisiewicz, V. Rieser, V. Protasov, V. Mikhailov, Y. Pruksachatkun, Y. Belinkov, Z. Bamberger, Z. Kasner, A. Rueda, A. Pestana, A. Feizpour, A. Khan, A. Faranak, A. Santos, A. Hevia, A. Unldreaj, A. Aghagol, A. Abdollahi, A. Tammour, A. HajiHosseini, B. Behroozi, B. Ajibade, B. Saxena, C. M. Ferrandis, D. Contractor,
|
2306.05949#86
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06070
| 86 |
assistant Answer: C.
user
Action: SELECT Value: Pickup ``` <html> <div> <main main> <section tabpanel> <div> <ul tablist> <li tab heading level 3 search and> </li> <li id=0 tab heading level 3 search and> <span> Hotel </span> </li> <li tab heading level 3 search and> </li> <li tab heading level 3 search and> </li> </ul> <div tabpanel> <div id=1> <div> <span> Dates* </span> <button button clear dates /> </div> ... </html> ``` Based on the HTML webpage above, try to complete the following task: Task: Compare the fare types to book a 1-adult ticket from Springfiels, IL to Austin, TX for April 29th 2023 Previous actions: [combobox] Enter your departing city, airport name, or airpor... SPRINGFIELD [button] Springfield, IL, US (SPI) -> CLICK [combobox] Enter your destination city, airport name, or airp... -> TYPE: AUSTIN [button] Austin, TX, US (AUS) -> CLICK What should be the next action? Please select from the following choices (If the correct action is not in the page above, please select A. âNone of the aboveâ):
|
2306.06070#86
|
Mind2Web: Towards a Generalist Agent for the Web
|
We introduce Mind2Web, the first dataset for developing and evaluating
generalist agents for the web that can follow language instructions to complete
complex tasks on any website. Existing datasets for web agents either use
simulated websites or only cover a limited set of websites and tasks, thus not
suitable for generalist web agents. With over 2,000 open-ended tasks collected
from 137 websites spanning 31 domains and crowdsourced action sequences for the
tasks, Mind2Web provides three necessary ingredients for building generalist
web agents: 1) diverse domains, websites, and tasks, 2) use of real-world
websites instead of simulated and simplified ones, and 3) a broad spectrum of
user interaction patterns. Based on Mind2Web, we conduct an initial exploration
of using large language models (LLMs) for building generalist web agents. While
the raw HTML of real-world websites are often too large to be fed to LLMs, we
show that first filtering it with a small LM significantly improves the
effectiveness and efficiency of LLMs. Our solution demonstrates a decent level
of performance, even on websites or entire domains the model has never seen
before, but there is still a substantial room to improve towards truly
generalizable agents. We open-source our dataset, model implementation, and
trained models (https://osu-nlp-group.github.io/Mind2Web) to facilitate further
research on building a generalist agent for the web.
|
http://arxiv.org/pdf/2306.06070
|
Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samuel Stevens, Boshi Wang, Huan Sun, Yu Su
|
cs.CL
|
Website: https://osu-nlp-group.github.io/Mind2Web. Updated with
supplementary material. NeurIPS'23 Spotlight
| null |
cs.CL
|
20230609
|
20231209
|
[] |
2306.06283
| 86 |
[34] Ramos, M. C.; Michtavy, S. S.; Porosoï¬, M. D.; White, A. D. Bayesian Optimization of Catalysts With In-context Learning. arXiv preprint: Arxiv-2304.05341 2023,
[35] White, A. D. The future of chemistry is language. Nat. Rev. Chem. 2023,
[36] Ramakrishnan, R.; Dral, P. O.; Rupp, M.; Von Lilienfeld, O. A. Big data meets quantum chemistry approximations: the â-machine learning approach. J. Chem. Theory Comput. 2015, 11, 2087â2096.
[37] Dinh, T.; Zeng, Y.; Zhang, R.; Lin, Z.; Gira, M.; Rajput, S.; Sohn, J.-Y.; Papail- iopoulos, D.; Lee, K. LIFT: Language-Interfaced Fine-Tuning for Non-Language Machine Learning Tasks. arXiv preprint: Arxiv-2206.06565. 2022.
|
2306.06283#86
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 87 |
of participating models in the fourth tab above labeled âLeaderboardâ or by clicking here. The models include both closed-source models and open-source models. Model A Model Draft a professional email seeking your supervisor's feedback on the âQuarterly Financial Reportâ Draft a professional email seeking your supervisor's feedback on the âQuarterly Financial Reportâ you prepared. Ask specifically about the data analysis, presentation style, and the clarity of you prepared. Ask specifically about the data analysis, presentation style, and the clarity of conclusions drawn. Keep the email short and to the point. conclusions drawn. Keep the email short and to the point. Subject: Request for Feedback on Quarterly Financial Report Subject: Request for Feedback onQuarterly Financial Report Dear (Supervisor's Name], Dear [Supervisor's Name], I hope this email finds you well. | am writing to seek your feedback on the Quarterly Financial | hope this email finds you well. | am writing to seek your feedback on the Quarterly Financial Report that | prepared for our company. The report provides an overview of our financial Report | prepared for the company. | am confident that the report provides valuable insights performance for the past quarter, and | would greatly appreciate any comments or suggestions into our
|
2306.05685#87
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 87 |
44https://huggingface.co/project-baize/baize-lora-30B 45https://huggingface.co/decapoda-research/llama-30b-hf 46https://huggingface.co/decapoda-research/llama-65b-hf 47https://huggingface.co/bigscience/bloom 48https://huggingface.co/bigscience/bloomz 49https://huggingface.co/bigscience/bloomz-mt 50https://huggingface.co/bigscience/bloomz-p3 51https://platform.openai.com/docs/models/gpt-3-5 52https://platform.openai.com/docs/models/gpt-4
|
2306.05783#87
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05817
| 87 |
ChatGPT Frozen Scoring/Ranking Function (Hybrid Task) RS Task Top-N RS N/A Sequential RS Sequential RS Sequential RS N/A Rating Prediction Rating Prediction Sequential RS Top-N RS Summary Recommendation Sequential RS Sequential RS Sequential RS Retrieval Sequential RS Sequential RS Top-N RS Explaination Generation Sequential RS Top-N RS Sequential RS Sequential RS Conversational RS Sequential RS Sequential RS Sequential RS RS Scenario Movie Tabular News Question Social Media News Product Tabular Movie Book Movie Book Book Movie News E-commerce Movie Insurance E-commerce Business E-commerce Music Movie Movie E-commerce Movie E-commerce N/A Movie Movie News
|
2306.05817#87
|
How Can Recommender Systems Benefit from Large Language Models: A Survey
|
Recommender systems (RS) play important roles to match users' information
needs for Internet applications. In natural language processing (NLP) domains,
large language model (LLM) has shown astonishing emergent abilities (e.g.,
instruction following, reasoning), thus giving rise to the promising research
direction of adapting LLM to RS for performance enhancements and user
experience improvements. In this paper, we conduct a comprehensive survey on
this research direction from an application-oriented view. We first summarize
existing research works from two orthogonal perspectives: where and how to
adapt LLM to RS. For the "WHERE" question, we discuss the roles that LLM could
play in different stages of the recommendation pipeline, i.e., feature
engineering, feature encoder, scoring/ranking function, and pipeline
controller. For the "HOW" question, we investigate the training and inference
strategies, resulting in two fine-grained taxonomy criteria, i.e., whether to
tune LLMs or not, and whether to involve conventional recommendation model
(CRM) for inference. Detailed analysis and general development trajectories are
provided for both questions, respectively. Then, we highlight key challenges in
adapting LLM to RS from three aspects, i.e., efficiency, effectiveness, and
ethics. Finally, we summarize the survey and discuss the future prospects. We
also actively maintain a GitHub repository for papers and other related
resources in this rising direction:
https://github.com/CHIANGEL/Awesome-LLM-for-RecSys.
|
http://arxiv.org/pdf/2306.05817
|
Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, Weinan Zhang
|
cs.IR, cs.AI
|
15 pages; 3 figures; summarization table in appendix
| null |
cs.IR
|
20230609
|
20230628
|
[
{
"id": "2302.13971"
},
{
"id": "1810.04805"
},
{
"id": "2304.05263"
},
{
"id": "2305.07001"
},
{
"id": "2305.11700"
},
{
"id": "2305.06566"
},
{
"id": "2305.15756"
},
{
"id": "2105.08318"
},
{
"id": "2304.03879"
},
{
"id": "2303.08559"
},
{
"id": "1703.04247"
},
{
"id": "2206.07682"
},
{
"id": "2305.07961"
},
{
"id": "2305.05973"
},
{
"id": "2305.15498"
},
{
"id": "2306.11114"
},
{
"id": "2305.04518"
},
{
"id": "2305.00447"
},
{
"id": "2305.02182"
},
{
"id": "2305.08845"
},
{
"id": "2305.12090"
},
{
"id": "2212.10403"
},
{
"id": "2304.03022"
},
{
"id": "2305.07609"
},
{
"id": "2209.08060"
},
{
"id": "2209.07562"
},
{
"id": "2304.09542"
},
{
"id": "2303.14524"
},
{
"id": "2305.15673"
},
{
"id": "2303.18223"
},
{
"id": "2304.10149"
},
{
"id": "1908.08167"
},
{
"id": "1909.10351"
},
{
"id": "2305.15036"
},
{
"id": "2305.12081"
},
{
"id": "2304.07862"
},
{
"id": "2305.03017"
},
{
"id": "2305.09858"
},
{
"id": "2305.06474"
},
{
"id": "2305.13731"
},
{
"id": "2304.03153"
},
{
"id": "2205.08084"
},
{
"id": "2106.09685"
},
{
"id": "2306.10702"
},
{
"id": "2306.02250"
},
{
"id": "2303.13835"
},
{
"id": "2305.14302"
},
{
"id": "2302.03735"
},
{
"id": "2306.02841"
},
{
"id": "2305.11206"
},
{
"id": "2203.15876"
},
{
"id": "2305.07622"
},
{
"id": "2306.10933"
},
{
"id": "2305.06569"
},
{
"id": "2206.06190"
}
] |
2306.06070
| 87 |
A. None of the above B. <li id=0 tab heading level 3 search and> <span> Hotel C. <div id=1> <div> <span> Dates* </span> <button button clear dates D. <ul id=2> <a mobile tools> </a> <a open unitedâs tiktok
# assistant Answer: A.
Continued on next page
23
Table 8 â continued from previous page
# Role
# Role
# Content
# user
``` <html> <div> <nav main menu> <ul> <li> <div button> Car Sales </div> <div id=0> <div> <div> <div> Buy A Car </div> <div> Plan Your Purchase </div> </div> <div> <h4> Its Tax Refund Time. Treat Yourself to an Upgrade. </h4> <p> With a variety of options, invest your refund in what you really want - a quality, used vehicle from Enterprise. </p> ... </html> ```
|
2306.06070#87
|
Mind2Web: Towards a Generalist Agent for the Web
|
We introduce Mind2Web, the first dataset for developing and evaluating
generalist agents for the web that can follow language instructions to complete
complex tasks on any website. Existing datasets for web agents either use
simulated websites or only cover a limited set of websites and tasks, thus not
suitable for generalist web agents. With over 2,000 open-ended tasks collected
from 137 websites spanning 31 domains and crowdsourced action sequences for the
tasks, Mind2Web provides three necessary ingredients for building generalist
web agents: 1) diverse domains, websites, and tasks, 2) use of real-world
websites instead of simulated and simplified ones, and 3) a broad spectrum of
user interaction patterns. Based on Mind2Web, we conduct an initial exploration
of using large language models (LLMs) for building generalist web agents. While
the raw HTML of real-world websites are often too large to be fed to LLMs, we
show that first filtering it with a small LM significantly improves the
effectiveness and efficiency of LLMs. Our solution demonstrates a decent level
of performance, even on websites or entire domains the model has never seen
before, but there is still a substantial room to improve towards truly
generalizable agents. We open-source our dataset, model implementation, and
trained models (https://osu-nlp-group.github.io/Mind2Web) to facilitate further
research on building a generalist agent for the web.
|
http://arxiv.org/pdf/2306.06070
|
Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samuel Stevens, Boshi Wang, Huan Sun, Yu Su
|
cs.CL
|
Website: https://osu-nlp-group.github.io/Mind2Web. Updated with
supplementary material. NeurIPS'23 Spotlight
| null |
cs.CL
|
20230609
|
20231209
|
[] |
2306.06283
| 87 |
[38] Krenn, M.; H¨ase, F.; Nigam, A.; Friederich, P.; Aspuru-Guzik, A. Self-referencing embed- ded strings (SELFIES): A 100% robust molecular string representation. Mach. Learn.: Sci. Technol. 2020, 1, 045024.
[39] Krenn, M.; Ai, Q.; Barthel, S.; Carson, N.; Frei, A.; Frey, N. C.; Friederich, P.; Gaudin, T.; Gayle, A. A.; Jablonka, K. M., et al. SELFIES and the future of molecular string repre- sentations. Patterns 2022, 3, 100588.
[40] Curtiss, L. A.; Redfern, P. C.; Raghavachari, K. Gaussian-4 theory using reduced order perturbation theory. J. Chem. Phys. 2007, 127, 124105.
[41] Ramakrishnan, R.; Dral, P. O.; Rupp, M.; Von Lilienfeld, O. A. Quantum chemistry structures and properties of 134 kilo molecules. Sci. Data 2014, 1, 1â7.
|
2306.06283#87
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 88 |
prepared for the company. | am confident that the report provides valuable insights performance for the past quarter, and | would greatly appreciate any comments or suggestions into our financial performance and would like to hear your thoughts on the data analysis, you may have about the data analysis, presentation style, and clarity of the conclusions presentation style, and clarity of conclusions drawn. drawn. Please let me know if you have any specific areas of concern or if there is any information you The report is a crucial part of our financial planning and decision-making, and | want to ensure would like me to update or expand upon. | am eager to improve and ensure that our financial that it meets the highest standards of accuracy and professionalism. Your expertise in the field reports are as accurate and informative as possible. of finance will be invaluable in guiding me towards improvement in these areas. Thank you for taking the time to read my email. | look forward to hearing from you soon, I have attached the report to this email for your reference. If you could provide your feedback Best regards, within the next few days, | would be grateful. [Your Name] Thank you in advance for your time and support. Best regards, {Your Name] ~® Ais better
|
2306.05685#88
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 88 |
15
Xiezhi Xiezhi-Spec Xiezhi-Inter Xiezhi-Train 48 60 1189 52188 48 125 431 36638 53 6554 48 383 40 45 48 59 48 163 50 1876 48 1413 32 262 271 905 39 44 41 70 47 53 48 95 48 67 166 874 Law (Level 1 subject) 0 222 Legal Theory 0 190 History of Law 2 51 Constitutional and Administrative Jurisprudence 0 54 Criminal Law 40 45 Civil and Commercial Law (including: Labour Law, Social Security Law) 0 44 Litigation Law 0 44 Economic Law 40 Environmental and Resource Protection Law 47 36 International Law (including: Public International Law, Private International Law, International Economic Law) 41 0 58 Military Law 177 4011 40 45 48 137 0 939 39 44 48 1063 450 6984 218 752 48 62 26 37 48 63 48 171 135 755 0 0 47 57 40 46 48 601 0 215 0 37 0 93 0 37 0 11 0 2 0 38 0 29 0 46 0 47 806 26971 283 1763 0 671 42 54 2 80 50 286 1 296 48 104 48 59 46 107 373 764 0 57 48 67 40 49 0 47 42 82 48 76 40 46 41 75 1 71 48 63 17 22 48 333 0 143 0 64 297 49 162353 248 1180 6732 11973 424 637 1177 410 61
16
|
2306.05783#88
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05817
| 88 |
# P5 [Geng et al., 2022]
T5-base (223M)
# FFT
# FFT
# Rating Prediction Top-N RS Sequential RS Explanation Generation Review Summarization
# Business E-commerce
# M6-Rec [Cui et al., 2022]
M6-base (300M)
# OT
# Retrieval Ranking Explanation Generation Conversational RS
# E-commerce
Table 1 continued from previous page
Model Name InstructRec [Zhang et al., 2023b] ChatGPT-1 [Liu et al., 2023a] ChatGPT-2 [Dai et al., 2023] ChatGPT-4 [Sun et al., 2023] Chat-REC [Gao et al., 2023] RecLLM [Friedman et al., 2023] LLM Backbone LLM Tuning Strategy Flan-T5-XL (3B) FFT ChatGPT Frozen ChatGPT Frozen ChatGPT Frozen Pipeline Controller ChatGPT Frozen LLaMA (7B) FFT RS Task Sequential RS Product Search Personalized Search Matching-then-reranking Rating Prediction Top-N RS Sequential RS Explanation Generation Review Summarization Pointwise Ranking Pairwise Ranking List-wise Ranking Passage Reranking Rating Prediction Top-N RS Conversational RS RS Scenario E-commerce E-commerce News Movie E-commerce Web Search Movie Video
|
2306.05817#88
|
How Can Recommender Systems Benefit from Large Language Models: A Survey
|
Recommender systems (RS) play important roles to match users' information
needs for Internet applications. In natural language processing (NLP) domains,
large language model (LLM) has shown astonishing emergent abilities (e.g.,
instruction following, reasoning), thus giving rise to the promising research
direction of adapting LLM to RS for performance enhancements and user
experience improvements. In this paper, we conduct a comprehensive survey on
this research direction from an application-oriented view. We first summarize
existing research works from two orthogonal perspectives: where and how to
adapt LLM to RS. For the "WHERE" question, we discuss the roles that LLM could
play in different stages of the recommendation pipeline, i.e., feature
engineering, feature encoder, scoring/ranking function, and pipeline
controller. For the "HOW" question, we investigate the training and inference
strategies, resulting in two fine-grained taxonomy criteria, i.e., whether to
tune LLMs or not, and whether to involve conventional recommendation model
(CRM) for inference. Detailed analysis and general development trajectories are
provided for both questions, respectively. Then, we highlight key challenges in
adapting LLM to RS from three aspects, i.e., efficiency, effectiveness, and
ethics. Finally, we summarize the survey and discuss the future prospects. We
also actively maintain a GitHub repository for papers and other related
resources in this rising direction:
https://github.com/CHIANGEL/Awesome-LLM-for-RecSys.
|
http://arxiv.org/pdf/2306.05817
|
Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, Weinan Zhang
|
cs.IR, cs.AI
|
15 pages; 3 figures; summarization table in appendix
| null |
cs.IR
|
20230609
|
20230628
|
[
{
"id": "2302.13971"
},
{
"id": "1810.04805"
},
{
"id": "2304.05263"
},
{
"id": "2305.07001"
},
{
"id": "2305.11700"
},
{
"id": "2305.06566"
},
{
"id": "2305.15756"
},
{
"id": "2105.08318"
},
{
"id": "2304.03879"
},
{
"id": "2303.08559"
},
{
"id": "1703.04247"
},
{
"id": "2206.07682"
},
{
"id": "2305.07961"
},
{
"id": "2305.05973"
},
{
"id": "2305.15498"
},
{
"id": "2306.11114"
},
{
"id": "2305.04518"
},
{
"id": "2305.00447"
},
{
"id": "2305.02182"
},
{
"id": "2305.08845"
},
{
"id": "2305.12090"
},
{
"id": "2212.10403"
},
{
"id": "2304.03022"
},
{
"id": "2305.07609"
},
{
"id": "2209.08060"
},
{
"id": "2209.07562"
},
{
"id": "2304.09542"
},
{
"id": "2303.14524"
},
{
"id": "2305.15673"
},
{
"id": "2303.18223"
},
{
"id": "2304.10149"
},
{
"id": "1908.08167"
},
{
"id": "1909.10351"
},
{
"id": "2305.15036"
},
{
"id": "2305.12081"
},
{
"id": "2304.07862"
},
{
"id": "2305.03017"
},
{
"id": "2305.09858"
},
{
"id": "2305.06474"
},
{
"id": "2305.13731"
},
{
"id": "2304.03153"
},
{
"id": "2205.08084"
},
{
"id": "2106.09685"
},
{
"id": "2306.10702"
},
{
"id": "2306.02250"
},
{
"id": "2303.13835"
},
{
"id": "2305.14302"
},
{
"id": "2302.03735"
},
{
"id": "2306.02841"
},
{
"id": "2305.11206"
},
{
"id": "2203.15876"
},
{
"id": "2305.07622"
},
{
"id": "2306.10933"
},
{
"id": "2305.06569"
},
{
"id": "2206.06190"
}
] |
2306.06070
| 88 |
Based on the HTML webpage above, try to complete the following task: Task: Find a mini van at Brooklyn City from April 5th to April 8th for a 22 year old renter. Previous actions: [searchbox] Pick-up & Return Location (ZIP, City or Airport) (... Brooklyn [option] Brooklyn, NY, US Select -> CLICK What should be the next action? Please select from the following choices (If the correct action is not in the page above, please select A. âNone of the aboveâ):
A. None of the above B. <div id=0> <div> <div> <div> Buy A Car </div> <div> C. <div id=1> Enterprise Fleet Management </div> D. <button id=2 selected pick-up date 03/19/2023> <span> <span> 19 </span>
# assistant Answer: D.
# Action: CLICK
24
|
2306.06070#88
|
Mind2Web: Towards a Generalist Agent for the Web
|
We introduce Mind2Web, the first dataset for developing and evaluating
generalist agents for the web that can follow language instructions to complete
complex tasks on any website. Existing datasets for web agents either use
simulated websites or only cover a limited set of websites and tasks, thus not
suitable for generalist web agents. With over 2,000 open-ended tasks collected
from 137 websites spanning 31 domains and crowdsourced action sequences for the
tasks, Mind2Web provides three necessary ingredients for building generalist
web agents: 1) diverse domains, websites, and tasks, 2) use of real-world
websites instead of simulated and simplified ones, and 3) a broad spectrum of
user interaction patterns. Based on Mind2Web, we conduct an initial exploration
of using large language models (LLMs) for building generalist web agents. While
the raw HTML of real-world websites are often too large to be fed to LLMs, we
show that first filtering it with a small LM significantly improves the
effectiveness and efficiency of LLMs. Our solution demonstrates a decent level
of performance, even on websites or entire domains the model has never seen
before, but there is still a substantial room to improve towards truly
generalizable agents. We open-source our dataset, model implementation, and
trained models (https://osu-nlp-group.github.io/Mind2Web) to facilitate further
research on building a generalist agent for the web.
|
http://arxiv.org/pdf/2306.06070
|
Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samuel Stevens, Boshi Wang, Huan Sun, Yu Su
|
cs.CL
|
Website: https://osu-nlp-group.github.io/Mind2Web. Updated with
supplementary material. NeurIPS'23 Spotlight
| null |
cs.CL
|
20230609
|
20231209
|
[] |
2306.06283
| 88 |
[42] Narayanan, B.; Redfern, P. C.; Assary, R. S.; Curtiss, L. A. Accurate quantum chemical energies for 133000 organic molecules. Chem. Sci. 2019, 10, 7449â7455.
[43] Gupta, A. K.; Raghavachari, K. Three-Dimensional Convolutional Neural Networks Uti- lizing Molecular Topological Features for Accurate Atomization Energy Predictions. J. Chem. Theory Comput. 2022, 18, 2132â2143.
32
[44] Ward, L.; Blaiszik, B.; Foster, I.; Assary, R. S.; Narayanan, B.; Curtiss, L. Machine learning prediction of accurate atomization energies of organic molecules from low-ï¬delity quantum chemical calculations. MRS Commun. 2019, 9, 891â899.
[45] Ramakrishnan, R.; Dral, P. O.; Rupp, M.; von Lilienfeld, O. A. Big Data Meets Quan- tum Chemistry Approximations: The â-Machine Learning Approach. J. Chem. Theory Comput. 2015, 11, 2087â2096.
[46] Becke, A. D. Density-functional thermochemistry. III. The role of exact exchange. J. Chem. Phys. 1993, 98, 5648â5652.
|
2306.06283#88
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05783
| 89 |
16
Subject Ancient Chinese History Modern Chinese History World History Archaeology Science History of Science and Technology History of Science and Technology (Level 2 subject) Chemistry Inorganic chemistry Analytical chemistry Organic chemistry Physical Chemistry Polymer chemistry and physics Geophysics Solid State Geophysics Space physics Geography Physical Geography Human Geography Cartography and GIS Geology Mineralogy Petrology Ore Deposits Geochemistry Palaeontology and Stratigraphy Palaeoanthropology Tectonic Geology Quaternary Geology Atmospheric Sciences Meteorology Atmospheric Physics and Atmospheric Environment Astronomy Astrophysics Astrometry and Celestial Mechanics Mathematics Basic Mathematics Computational Mathematics Probability Theory and Mathematical Statistics Applied Mathematics Operations Research and Cybernetics Marine Science Physical Oceanography Marine Chemistry Marine Biology Marine Geology Physics Theoretical Physics Particle Physics and Atomic Nucleus Physics Atomic and molecular physics Plasma physics Condensed Matter Physics Acoustics Optics Radio physics Ecology Biology Botany Zoology Physiology Aquatic Biology Microbiology Neurobiology Genetics Developmental Biology Cell Biology Biochemistry and Molecular Biology Biophysics Ecology Systems Science Systems Theory Systems Analysis and Integration Statistics Engineering Optical Engineering Optical Engineering (Level 2 discipline)
|
2306.05783#89
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05949
| 89 |
I. Sedenko, I. Nejadgholi, J. Passmore, J. Seltzer, J. B. Sanz, L. Dutra, M. Samagaio, M. El- badri, M. Mieskes, M. Gerchick, M. Akinlolu, M. McKenna, M. Qiu, M. Ghauri, M. Burynok, N. Abrar, N. Rajani, N. Elkott, N. Fahmy, O. Samuel, R. An, R. Kromann, R. Hao, S. Alizadeh, S. Shubber, S. Wang, S. Roy, S. Viguier, T. Le, T. Oyebade, T. Le, Y. Yang, Z. Nguyen, A. R. Kashyap, A. Palasciano, A. Callahan, A. Shukla, A. Miranda-Escalada, A. Singh, B. Beil- harz, B. Wang, C. Brito, C. Zhou, C. Jain, C. Xu, C. Fourrier, D. L. Periñán, D. Molano, D. Yu, E. Manjavacas, F. Barth, F.
|
2306.05949#89
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06283
| 89 |
[47] Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; Chen, W. LoRA: Low-Rank Adaptation of Large Language Models. arXiv preprint: Arxiv-2106.09685 2021,
[48] Radford, A.; Wu, J.; Child, R.; Luan, D.; Amodei, D.; Sutskever, I. Language Models are Unsupervised Multitask Learners. 2019, https://d4mucfpksywv.cloudfront. net/better-language-models/language_models_are_unsupervised_multitask_ learners.pdf.
[49] Scrivener, K. L.; John, V. M.; Gartner, E. M. Eco-eï¬cient cements: Potential economi- cally viable solutions for a low-CO2 cement-based materials industry. Cem. Concr. Res. 2018, 114, 2â26.
|
2306.06283#89
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 90 |
Figure 19: The screenshot of Chatbot Arena.
# C.3 Data Release
We will clean the Personal Identifiable Information (PII) and tag toxic conversations with OpenAI moderation APIs for our dataset release.
24
# D Additional Experimental Results
We present some additional experimental results.
# D.1 Position bias
We test two more prompts and present the full results in Table 9 âscoreâ changes the default prompt to let the model output two absolute scores instead of which one is better. âshortâ is a simplified version of our default prompt by removing instructions like âAvoid any position bias..â, âBegin your evaluation ... and provide a short explanationâ. We can find different prompts have different effects on different models. For example, the "score" prompt can increase the consistency of GPT-3.5 but decreases it for Claude-v1 and GPT-4.
Table 9: Position bias on different models and prompts. Consistency is the percentage of cases where a judge gives consistent results when swapping the order of two assistants. âBiased toward firstâ is the percentage of cases when a judge favors the first answer. âErrorâ indicates wrong output formats. The two largest numbers in each column are in bold.
|
2306.05685#90
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 90 |
Xiezhi Xiezhi-Spec Xiezhi-Inter Xiezhi-Train 3290 5003 2618 3158 423 1180 413 1509 1261 18563 77 306 0 45 7 399 0 46 0 59 1 46 1 46 0 36 16 208 1 46 1 45 649 5537 9 89 110 189 46 103 322 941 2 50 1 46 0 41 56 113 33 99 153 249 1 46 0 41 2 368 1 270 0 27 12 170 1 31 2 33 67 1080 1 31 2 29 1 46 0 45 4 31 24 299 1 42 1 46 1 46 3 49 35 1148 1 44 0 44 0 60 0 44 1 60 1 46 4 65 0 45 49 313 138 2831 8 117 47 414 2 51 0 45 0 27 0 45 1 37 2 61 1 46 1 45 0 44 49 313 0 163 0 45 0 33 22 147 562 35917 0 771 0 60 0 686 0 43 25 934 0 30 0 56 9 63 2 46 1 781 1 46 0 45 4 809 1 28 1 55 22 944 1 47
|
2306.05783#90
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05949
| 90 |
C. Fourrier, D. L. Periñán, D. Molano, D. Yu, E. Manjavacas, F. Barth, F. Fuhrimann, G. Altay, G. Bayrak, G. Burns, H. U. Vrabec, I. Bello, I. Dash, J. Kang, J. Giorgi, J. Golde, J. D. Posada, K. R. Sivaraman, L. Bulchandani, L. Liu, L. Shinzato, M. H. de Bykhovetz, M. Takeuchi, M. Pà mies, M. A. Castillo, M. Nezhu- rina, M. Sänger, M. Samwald, M. Cullan, M. Weinberg, M. D. Wolf, M. Mihaljcic, M. Liu, M. Freidank, M. Kang, N. Seelam, N. Dahlberg, N. M. Broad, N. Muellner, P. Fung, P. Haller, R. Chandrasekhar, R. Eisenberg, R. Martin, R. Canalli, R. Su, R. Su, S. Cahyawijaya, S. Garda, S.
|
2306.05949#90
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06283
| 90 |
[50] V¨olker, C.; Benjami Moreno Torres,; Tehseen Rug,; Firdous, R.; Ghezal Ahmad,; Zia, J.; L¨uders, S.; Scaï¬no, H. L.; H¨opler, M.; B¨ohmer, F.; Pfaï¬, M.; Stephan, D.; Kruschwitz, S. Green building materials: a new frontier in data-driven sustainable concrete design. Preprint 10.13140/RG.2.2.29079.85925. 2023.
[51] Rao, G. M.; Rao, T. D. G. A quantitative method of approach in designing the mix proportions of ï¬y ash and GGBS-based geopolymer concrete. Aust. J. Civ. Eng. 2018, 16, 53â63.
[52] Tshitoyan, V.; Dagdelen, J.; Weston, L.; Dunn, A.; Rong, Z.; Kononova, O.; Pers- son, K. A.; Ceder, G.; Jain, A. Unsupervised word embeddings capture latent knowledge from materials science literature. Nature 2019, 571, 95â98.
|
2306.06283#90
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 91 |
Judge Prompt Consistency Biased toward first Biased toward second Error claude-v1 default rename score short 23.8% 56.2% 20.0% 22.5% 75.0% 11.2% 80.0% 75.0% 0.0% 28.7% 0.0% 2.5% 1.2% 3.8% 0.0% 0.0% gpt-3.5-turbo default rename score short 46.2% 51.2% 55.0% 38.8% 50.0% 38.8% 33.8% 57.5% 1.2% 6.2% 11.2% 3.8% 2.5% 3.8% 0.0% 0.0% gpt-4 default rename score short 65.0% 66.2% 51.2% 62.5% 30.0% 28.7% 46.2% 35.0% 5.0% 5.0% 2.5% 2.5% 0.0% 0.0% 0.0% 0.0%
|
2306.05685#91
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 91 |
0 0 0 1 2347 88 40 200 41 48 40 40 31 128 40 39 204 48 48 48 386 41 40 36 47 48 48 40 36 70 0 22 102 25 26 171 25 22 40 40 22 164 36 40 40 0 301 38 39 48 39 48 0 0 40 8 420 48 48 0 40 22 40 0 48 40 39 39 8 76 0 28 0 4717 96 48 86 38 208 25 48 48 39 128 40 40 118 22 48 162 41
5 5 5 5 340 10 5 30 5 5 5 5 5 15 5 5 20 5 5 5 45 5 5 5 5 5 5 5 5 15 5 5 15 5 5 30 5 5 5 5 5 25 5 5 5 5 45 5 5 5 5 5 5 5 5 5 65 5 5 5 5 5 5 5 5 5 5 5 5 15 5 5 5 770 10 5 10 5 25 5 5 5 5 15 5 5 15 5 5 25 5
# Biomedical Engineering (Level 2 discipline)
# Traffic and Transport Engineering
Road and Railway Engineering Traffic Information Engineering and Control Transport Planning and Management Vehicle Operation Engineering Instrument Science and Technology
# Precision Instruments and Machinery Test and Measurement Technology and Instrumentation
# Information and Communication Engineering Communication and Information Systems Signal and Information Processing
# Weapons Science and Technology
# Weapon Systems and Operations Engineering
17
|
2306.05783#91
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05949
| 91 |
R. Eisenberg, R. Martin, R. Canalli, R. Su, R. Su, S. Cahyawijaya, S. Garda, S. S. Deshmukh, S. Mishra, S. Kiblawi, S. Ott, S. Sang-aroonsiri, S. Kumar, S. Schweter, S. Bharati, T. Laud, T. Gigant, T. Kainuma, W. Kusa, Y. Labrak, Y. S. Bajaj, Y. Venkatraman, Y. Xu, Y. Xu, Y. Xu, Z. Tan, Z. Xie, Z. Ye, M. Bras, Y. Belkada, and T. Wolf. Bloom: A 176b-parameter open-access multilingual language model, 2023.
|
2306.05949#91
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06283
| 91 |
[53] Mikolov, T.; Chen, K.; Corrado, G.; Dean, J. Eï¬cient Estimation of Word Representations in Vector Space. International Conference On Learning Representations. 2013.
[54] Olivetti, E. A.; Cole, J. M.; Kim, E.; Kononova, O.; Ceder, G.; Han, T. Y.-J.; Hiszpan- ski, A. M. Data-driven materials research enabled by natural language processing and
33
information extraction. Appl. Phys. Rev. 2020, 7, 041317.
[55] Selva Birunda, S.; Kanniga Devi, R. A review on word embedding techniques for text classiï¬cation. Innovative Data Communication Technologies and Application: Proceedings of ICIDCA 2020 2021, 267â281.
[56] Hong, Z.; Ajith, A.; Pauloski, G.; Duede, E.; Malamud, C.; Magoulas, R.; Chard, K.; Foster, I. ScholarBERT: Bigger is Not Always Better. arXiv preprint: Arxiv-2205.11342. 2022.
[57] Dai, H. et al. AugGPT: Leveraging ChatGPT for Text Data Augmentation. arXiv preprint: Arxiv-2302.13007. 2023.
|
2306.06283#91
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 92 |
As shown in Table 10, position bias is more noticeable on open questions like writing and stem/hu- manity knowledge questions. On math and coding questions, LLM judges are more confident even though their judgments can often be wrong, as we show in Section 3.3. Finally, we study how the model pairs influence position bias by using GPT-4 and the default prompt to judge three different model pairs. As shown in Table 11, the position bias is more noticeable for models with close performance and can almost disappear when the performance of the two models differs a lot.
Table 10: Position bias on different categories. The two largest numbers in each column are in bold.
Category writing roleplay reasoning math coding extraction stem humanities 42.0% 68.0% 76.0% 86.0% 86.0% 78.0% 44.0% 36.0% 46.0% 30.0% 20.0% 4.0% 14.0% 12.0% 54.0% 60.0% 12.0% 2.0% 4.0% 10.0% 0.0% 10.0% 2.0% 4.0%
Consistent Biased toward first Biased toward second
# Table 11: Position bias on different model pairs.
|
2306.05685#92
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 92 |
Subject Weapon Launch Theory and Technology Artillery, Automatic Weapons and Ammunition Engineering Military Chemistry and Pyrotechnics Agricultural Engineering Agricultural Mechanization Engineering Agricultural Soil and Water Engineering Agricultural Bioenvironmental and Energy Engineering Agricultural Electrification and Automation Metallurgical Engineering Metallurgical Physical Chemistry Mechanics General Mechanics Solid Mechanics Fluid Mechanics Engineering Mechanics Power Engineering and Engineering Thermophysics Engineering Thermophysics Thermal Engineering Power Mechanics and Engineering Fluid Mechanics and Engineering Refrigeration and Cryogenic Engineering Chemical Process Machinery Chemical Engineering and Technology Chemical Engineering Chemical Processes Biochemistry Applied Chemistry Industrial Catalysis Civil Engineering Geotechnical Engineering Structural Engineering Municipal Engineering Heating, Gas, Ventilation and Air Conditioning Engineering Disaster Prevention and Mitigation Engineering and Protection Engineering Bridge and Tunnel Engineering Geological Resources and Geological Engineering Mineral Census and Exploration Earth Exploration and Information Technology Geological Engineering Urban and Rural Planning Safety Science and Engineering Architecture History and Theory of Architecture Architectural Design and Theory Urban Planning and Design (including: Landscape Architecture Planning and Design)
|
2306.05783#92
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05949
| 92 |
[31] A. Birhane, W. Isaac, V. Prabhakaran, M. Diaz, M. C. Elish, I. Gabriel, and S. Mohamed. Power to the people? opportunities and challenges for participatory AI. In Equity and Access in Algorithms, Mechanisms, and Optimization. ACM, oct 2022. doi: 10.1145/3551624.3555290. URL https://doi.org/10.1145%2F3551624.3555290.
[32] A. Birhane, E. Ruane, T. Laurent, M. S. Brown, J. Flowers, A. Ventresque, and C. L. Dancy. The Forgotten Margins of AI Ethics. In 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 948â958. ACM, 2022-06-21. ISBN 978-1-4503-9352-2. doi: 10. 1145/3531146.3533157. URL https://dl.acm.org/doi/10.1145/3531146.3533157.
|
2306.05949#92
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06283
| 92 |
[58] Venkatasubramanian, V.; Chan, K.; Caruthers, J. M. Computer-aided molecular design using genetic algorithms. Comput. Chem. Eng. 1994, 18, 833â844.
[59] Flam-Shepherd, D.; Aspuru-Guzik, A. Language models can generate molecules, materi- als, and protein binding sites directly in three dimensions as XYZ, CIF, and PDB ï¬les. arXiv preprint: Arxiv-2305.05708. 2023.
[60] Taylor, R.; Kardas, M.; Cucurull, G.; Scialom, T.; Hartshorn, A.; Saravia, E.; Poulton, A.; Kerkez, V.; Stojnic, R. Galactica: A Large Language Model for Science. arXiv preprint: Arxiv-2211.09085 2022,
[61] Schwaller, P.; Gaudin, T.; L´anyi, D.; Bekas, C.; Laino, T. âFound in Translationâ: predict- ing outcomes of complex organic chemistry reactions using neural sequence-to-sequence models. Chem. Sci. 2018, 9, 6091â6098.
|
2306.06283#92
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 93 |
Consistent Biased toward first Biased toward second
# Table 11: Position bias on different model pairs.
Pair Consistent Biased toward first Biased toward second 67.5% GPT-3.5 vs Claude-V1 GPT-3.5 vs Vicuna-13B 73.8% GPT-3.5 vs LLaMA-13B 98.8% 23.8% 23.8% 1.2% 8.8% 2.5% 0.0%
25
# D.2 Few-shot judge
We examine how few-shot examples improve LLM judges. As shown in Table 12, they improve the consistency of all three LLM judges significantly. It almost alleviates the position bias of GPT-4, but moves the position bias of GPT-3.5 from the first position to the second position. We then measure the agreement between few-shot GPT-4 pairwise comparison and humans on MT-bench, but found it performs similarly to zero-shot GPT-4 pairwise comparison.
# Table 12: Improvements of the few-shot judge on consistency for position bias.
|
2306.05685#93
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 93 |
Science and Engineering Architecture History and Theory of Architecture Architectural Design and Theory Urban Planning and Design (including: Landscape Architecture Planning and Design) Building Technology Science Control Science and Engineering Control Theory and Control Engineering Testing Technology and Automation Systems Engineering Pattern Recognition and Intelligent Systems Navigation, Guidance and Control Mechanical Engineering Mechanical Manufacturing and Automation Mechatronics Engineering Mechanical Design and Theory Vehicle Engineering Materials Science and Engineering Materials Physics and Chemistry Materials Science Materials Processing Engineering Forestry Engineering Forest Engineering Wood Science and Technology Forestry Chemical Processing Engineering Nuclear Science and Technology Nuclear Energy Science and Engineering Nuclear Fuel Cycle and Materials Nuclear Technology and Applications Radiation Protection and Environmental Protection Water Resources Engineering Hydrology and Water Resources Hydraulics and River Dynamics Hydraulic Structural Engineering Hydraulic and Hydroelectric Engineering Port, Coastal and Offshore Engineering
|
2306.05783#93
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05949
| 93 |
[33] S. L. Blodgett, S. Barocas, H. Daumé III, and H. Wallach. Language (Technology) is Power: A Critical Survey of âBiasâ in NLP. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454â5476. Association for Computational Linguistics, 2020. doi: 10.18653/v1/2020.acl-main.485. URL https://www.aclweb.org/anthology/ 2020.acl-main.485.
[34] M. J. Bockarie. We need to end âparachuteâ research which sidelines the URL https://qz.com/africa/1536355/ work of African scientists, Jan. 2019. african-scientists-are-sidelined-by-parachute-research-teams.
|
2306.05949#93
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06283
| 93 |
[62] Yao, S.; Zhao, J.; Yu, D.; Du, N.; Shafran, I.; Narasimhan, K.; Cao, Y. ReAct: Synergizing Reasoning and Acting in Language Models. arXiv preprint: Arxiv-2210.03629 2023,
[63] Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Chi, E.; Xia, F.; Le, Q.; Zhou, D. Chain- of-Thought Prompting Elicits Reasoning in Large Language Models. Neural Information Processing Systems 2022,
[64] OpenAI, GPT-4 Technical Report. arXiv preprint: Arxiv-2303.08774v3. 2023.
[65] Bran, A. M.; Cox, S.; White, A. D.; Schwaller, P. ChemCrow: Augmenting large-language models with chemistry tools. arXiv preprint: Arxiv-2304.05376 2023,
[66] Boiko, D. A.; MacKnight, R.; Gomes, G. Emergent autonomous scientiï¬c research capa- bilities of large language models. arXiv preprint: Arxiv-2304.05332 2023,
|
2306.06283#93
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 94 |
# Table 12: Improvements of the few-shot judge on consistency for position bias.
Model Prompt Consistency Biased toward first Biased toward second Error Claude-v1 zero-shot few-shot 23.8% 63.7% 75.0% 21.2% 0.0% 11.2% 1.2% 3.8% GPT-3.5 zero-shot few-shot 46.2% 55.0% 50.0% 16.2% 1.2% 28.7% 2.5% 0.0% GPT-4 zero-shot few-shot 65.0% 77.5% 30.0% 10.0% 5.0% 12.5% 0.0% 0.0%
# D.3 Agreement Evaluation
Agreement calculation. We define the agreement between two types of judges as the probability of randomly selected individuals (but not identical) of each type agreeing on a randomly selected question. For example, if we are comparing GPT-4 and Claude, the agreement is the probability of GPT-4 and Claude agreeing on the vote for a randomly selected question. If we are comparing GPT-4 and humans, the agreement is the probability of GPT-4 and a randomly selected human agreeing on the vote for a randomly selected question. The agreement among humans themselves is the probability of two randomly selected but not identical humans agreeing on the vote for a randomly selected question.
|
2306.05685#94
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 94 |
Xiezhi Xiezhi-Spec Xiezhi-Inter Xiezhi-Train 61 30 60 890 66 60 59 55 772 64 902 45 46 44 47 994 51 48 60 46 57 59 911 60 65 45 46 50 1089 45 45 45 54 172 45 759 44 58 50 104 77 984 67 135 45 56 883 44 45 46 43 45 852 50 45 46 45 806 45 72 45 698 45 45 45 841 45 45 45 42 1057 52 45 46 54 45 772 29 45 65 1086 394 46 84 819 30 39 45 45 844 31
48 25 0 238 46 48 48 48 96 48 206 40 40 39 39 48 0 0 0 0 0 0 48 0 0 0 0 0 184 0 0 0 48 48 40 93 0 45 0 49 48 218 41 49 40 40 205 39 40 0 38 40 161 34 40 39 0 48 0 0 0 128 0 40 40 205 40 40 40 37 256 40 40 39 41 40 118 24 0 46 48 0 0 48 128 0 0 40 40 231 26
|
2306.05783#94
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05949
| 94 |
[35] R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill, E. Brynjolfsson, S. Buch, D. Card, R. Castellon, N. Chatterji, A. Chen, K. Creel, J. Q. Davis, D. Demszky, C. Donahue, M. Doumbouya, E. Durmus, S. Ermon, J. Etchemendy, K. Ethayarajh, L. Fei-Fei, C. Finn, T. Gale, L. Gillespie, K. Goel, N. Goodman, S. Grossman, N. Guha, T. Hashimoto, P. Henderson, J. Hewitt, D. E. Ho, J. Hong, K. Hsu, J. Huang, T. Icard, S. Jain, D. Jurafsky, P. Kalluri, S. Karamcheti, G. Keeling, F. Khani, O. Khattab, P. W. Koh, M. Krass,
|
2306.05949#94
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06283
| 94 |
[67] Jain, A.; Ong, S. P.; Hautier, G.; Chen, W.; Richards, W. D.; Dacek, S.; Cholia, S.; Gunter, D.; Skinner, D.; Ceder, G.; Persson, K. A. Commentary: The Materials Project:
34
A materials genome approach to accelerating materials innovation. APL Materials 2013,
1, 011002.
[68] Rego, N.; Koes, D. 3Dmol.js: molecular visualization with WebGL. Bioinformatics 2014, 31, 1322â1324.
[69] White, A.; Hocky, G. marvis - VMD Audio/Text control with natural language. https: //github.com/whitead/marvis, 2022.
[70] Humphrey, W.; Dalke, A.; Schulten, K. VMD: Visual molecular dynamics. J. Mol. Graph. 1996, 14, 33â38.
[71] Radford, A.; Kim, J. W.; Xu, T.; Brockman, G.; McLeavey, C.; Sutskever, I. Robust speech recognition via large-scale weak supervision. arXiv preprint: ArXiv-2212.04356. 2022.
|
2306.06283#94
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 95 |
Note that the agreement among humans could be a lower estimation compared to the agreement of GPT4 and humans. Consider three humans who voted âAâ, âAâ, and âBâ for a question, respectively. The agreement among them is only 1 3 , as there are three pairs â(A, A)â, â(A, B)â, and â(A, B)â. But the agreement between GPT4 and those three is 2
Therefore, to have a more comprehensive understanding of what happened, we introduce a new judge type called human-majority, which considers the majority of human votes for each question. The agreement between GPT4 and human-majority is then calculated as the probability of GPT4 agreeing with the majority of human votes on a randomly selected question. The upper bound of the agreement between GPT-4 and humans is the agreement between human-majority and human. When there is no majority vote for a question, the agreement is counted by an even split. For example, if there are an equal number of âAâ and âBâ human votes for a question, and GPT4 votes âAâ, the agreement is counted as 1
|
2306.05685#95
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 95 |
1 0 0 51 15 5 2 1 14 3 6 0 0 0 3 7 4 2 0 1 0 0 12 0 4 0 0 9 121 1 0 0 0 115 0 12 0 8 4 14 0 107 21 70 0 11 1 0 0 1 0 0 22 11 0 2 0 37 0 23 0 0 0 0 0 2 0 0 0 0 75 7 0 2 8 0 18 0 1 14 60 57 2 0 1 0 0 0 0 3 0
5 5 5 25 5 5 5 5 10 5 25 5 5 5 5 35 5 5 5 5 5 5 30 5 5 5 5 5 35 5 5 5 5 5 5 20 5 5 5 5 5 25 5 5 5 5 30 5 5 5 5 5 25 5 5 5 5 20 5 5 5 20 5 5 5 25 5 5 5 5 30 5 5 5 5 5 20 5 5 5 15 5 5 5 25 5 5 5 5 30 5
Geodesy and Surveying Engineering Photogrammetry and Remote Sensing Cartography and Geographic Information Engineering
# Environmental Science and Engineering
# Environmental Science Environmental Engineering
# Bioengineering Electronic Science and Technology
Physical Electronics Circuits and Systems Microelectronics and Solid State Electronics Electromagnetic Field and Microwave Technology
# Electrical Engineering
# Electrical Machines and Appliances
18
|
2306.05783#95
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05949
| 95 |
Kalluri, S. Karamcheti, G. Keeling, F. Khani, O. Khattab, P. W. Koh, M. Krass, R. Krishna, R. Kuditipudi, A. Kumar, F. Ladhak, M. Lee, T. Lee, J. Leskovec, I. Levent, X. L. Li, X. Li, T. Ma, A. Malik, C. D. Manning, S. Mirchandani, E. Mitchell, Z. Munyikwa, S. Nair, A. Narayan, D. Narayanan, B. Newman, A. Nie, J. C. Niebles, H. Nilforoshan, J. Nyarko, G. Ogut, L. Orr, I. Papadimitriou, J. S. Park, C. Piech, E. Portelance, C. Potts, A. Raghunathan, R. Reich, H. Ren, F. Rong, Y. Roohani, C. Ruiz, J. Ryan, C. Ré, D. Sadigh, S. Sagawa, K. Santhanam, A. Shih, K. Srinivasan, A. Tamkin, R.
|
2306.05949#95
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06283
| 95 |
[72] Baek, M. et al. Accurate prediction of protein structures and interactions using a three- track neural network. Science 2021, 373, 871â876.
[73] Watson, J. L. et al. Broadly applicable and accurate protein design by integrat- ing structure prediction networks and diï¬usion generative models. bioRxiv preprint: 10.1101/2022.12.09.519842. 2022.
[74] Andersen, C. W. et al. OPTIMADE, an API for exchanging materials data. Sci. Data 2021, 8, 217.
[75] Volk, A. A.; Epps, R. W.; Yonemoto, D. T.; Masters, B. S.; Castellano, F. N.; Reyes, K. G.; Abolhasani, M. AlphaFlow: autonomous discovery and optimization of multi-step chem- istry using a self-driven ï¬uidic lab guided by reinforcement learning. Nat. Commun. 2023, 14, 1403.
|
2306.06283#95
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 96 |
More results. Table 13 shows more agreement results on MT-bench. In addition to expert labelers (denoted as âHumanâ), we also include author votes (denoted as âAuthorâ).
# D.4 Category-wise scores with single-answer grading
We use single-answer grading to evaluate 6 models on MT-bench and plot the category-wise scores in Figure 20.
26
Table 13: Agreement between two types of judges on MT-bench. âG4-Pâ and âG4-Sâ denote GPT-4 with pairwise comparison and single-answer grading, respectively. âCâ denotes Claude. âHumanâ denotes expert labelers (excluding authors). âHuman-Mâ denotes the majority vote of humans. The single-answer grading can be converted into pairwise comparison results for calculating the agreement. We report two setups: âS1â includes non-tie, tie, and inconsistent (due to position bias) votes and counts inconsistent as a tie; âS2â only includes non-tie votes. The agreement between two random judges under each setup is denoted as âR=â. The top value in each cell is the agreement, and the bottom gray value is #votes.
|
2306.05685#96
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 96 |
Subject Power Systems and Automation High Voltage and Insulation Technology Power Electronics and Power Transmission Electrical Engineering Theory and New Technology Petroleum and Natural Gas Engineering Oil and Gas Well Engineering Oil and Gas Field Development Engineering Oil and Gas Storage and Transportation Engineering Mining Engineering Mining Engineering Mineral Processing Engineering Safety Technology and Engineering Textile Science and Engineering Textile Engineering Textile Materials and Textile Design Textile Chemistry and Dyeing and Finishing Engineering Clothing Cyberspace Security Aerospace Science and Technology Aircraft Design Aerospace Propulsion Theory and Engineering Aerospace Manufacturing Engineering Human-Machine and Environmental Engineering Marine and Offshore Engineering Ship and Marine Structures Design and Manufacture Turbine Engineering Hydroacoustic Engineering Computer Science and Technology Computer System Architecture Computer Software and Theory Computer Application Technology Software Engineering Light Industry Technology and Engineering Pulp and Paper Engineering Sugar Engineering Fermentation Engineering Leather Chemistry and Engineering Iron and Steel Metallurgy Non-Ferrous Metallurgy Landscape Architecture Food Science and Engineering Food Science Grain, Oil and Vegetable Protein Engineering Agricultural Products Processing and Storage Engineering Fish Processing and Storage Engineering Agronomy Crop Science Crop Cultivation and Farming Crop Genetics and Breeding
|
2306.05783#96
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.06283
| 96 |
[76] Shields, B. J.; Stevens, J.; Li, J.; Parasram, M.; Damani, F.; Alvarado, J. I. M.; Janey, J. M.; Adams, R. P.; Doyle, A. G. Bayesian reaction optimization as a tool for chemical synthesis. Nature 2021, 590, 89â96.
[77] Prieto Kullmer, C. N.; Kautzky, J. A.; Krska, S. W.; Nowak, T.; Dreher, S. D.; MacMil- lan, D. W. Accelerating reaction generality and mechanistic insight through additive map- ping. Science 2022, 376, 532â539.
[78] Rankovi´c, B.; Griï¬ths, R.-R.; Moss, H. B.; Schwaller, P. Bayesian optimisation for addi- tive screening and yield improvements in chemical reactions â beyond one-hot encodings. ChemRxiv preprint 10.26434/chemrxiv-2022-nll2j. 2022.
[79] Dunn, A.; Dagdelen, J.; Walker, N.; Lee, S.; Rosen, A. S.; Ceder, G.; Persson, K. A.;
35
|
2306.06283#96
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 97 |
Setup S1 (R = 33%) S2 (R = 50%) Judge G4-S C Author Human Human-M G4-S C Author Human Human-M G4-P 70% 1138 63% 1198 69% 345 66% 1343 67% 821 97% 662 94% 582 92% 201 85% 859 85% 546 G4-S - 66% 1136 67% 324 60% 1280 60% 781 - 90% 563 94% 175 85% 739 85% 473 C - - 58% 343 54% 1341 55% 820 - - 89% 141 85% 648 86% 414 Author - - 69% 49 65% 428 55% 93 - - 87% 31 83% 262 76% 46 Human - - - 63% 721 81% 892 - - - 81% 479 90% 631
(a) First Turn
|
2306.05685#97
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 97 |
Engineering Agricultural Products Processing and Storage Engineering Fish Processing and Storage Engineering Agronomy Crop Science Crop Cultivation and Farming Crop Genetics and Breeding Veterinary Medicine Basic Veterinary Medicine Preventive Veterinary Medicine Clinical Veterinary Medicine Agricultural Resource Utilisation Soil Science Plant Nutrition Horticulture Fruit Tree Science Vegetable Science Tea Forestry Forest Genetic Breeding Forest Breeding Forest Conservation Forest Management Wildlife Conservation and Utilisation Landscape Plants and Ornamental Horticulture Soil and Water Conservation and Desertification Control Plant Protection Plant Pathology Agricultural Insects and Pest Control Pesticides Aquaculture Aquaculture Capture Science Fishery Resources
|
2306.05783#97
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05949
| 97 |
Sorting Things Out Classification and Its Consequences. ISBN 978-0-262-52295-3. URL https://mitpress.mit.edu/ [36] G. Bowker and S. L. Star. The MIT Press, 2000. 9780262522953/sorting-things-out/.
[37] M. Brereton, P. Roe, R. Schroeter, and A. Lee Hong. Beyond ethnography: engagement and reciprocity as foundations for design research out here. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, CHI â14, pages 1183â1186, New York,
23
NY, USA, Apr. 2014. Association for Computing Machinery. ISBN 978-1-4503-2473-1. doi: 10.1145/2556288.2557374. URL https://doi.org/10.1145/2556288.2557374.
[38] S. Briscoe. U.S. Laws Address Deepfakes, Dec. 2021. URL http://www.asisonline. org/security-management-magazine/latest-news/today-in-security/2021/ january/U-S-Laws-Address-Deepfakes/.
|
2306.05949#97
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06283
| 97 |
[79] Dunn, A.; Dagdelen, J.; Walker, N.; Lee, S.; Rosen, A. S.; Ceder, G.; Persson, K. A.;
35
Jain, A. Structured information extraction from complex scientiï¬c text with ï¬ne-tuned large language models. arXiv preprint: Arxiv-2212.05238 2022,
[80] Walker, N.; Dagdelen, J.; Cruse, K.; Lee, S.; Gleason, S.; Dunn, A.; Ceder, G.; Alivisatos, A. P.; Persson, K. A.; Jain, A. Extracting Structured Seed-Mediated Gold Nanorod Growth Procedures from Literature with GPT-3. arXiv preprint: Arxiv- 2304.13846 2023,
[81] Neo4j, Neo4j - The Worldâs Leading Graph Database. 2012; http://neo4j.org/.
[82] Kearnes, S. M.; Maser, M. R.; Wleklinski, M.; Kast, A.; Doyle, A. G.; Dreher, S. D.; Hawkins, J. M.; Jensen, K. F.; Coley, C. W. The Open Reaction Database. J. Am. Chem. Soc. 143, 18820â18826.
|
2306.06283#97
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 98 |
(a) First Turn
Setup S1 (R = 33%) S2 (R = 50%) Judge G4-S Author Human Human-M G4-S Author Human Human-M G4-P 70% 1161 66% 341 66% 1325 68% 812 95% 727 88% 205 85% 864 85% 557 G4-S - 65% 331 59% 1285 61% 783 - 89% 193 84% 776 85% 506 Author - 67% 49 68% 413 63% 87 - 87% 31 86% 273 84% 54 Human - - 67% 707 83% 877 - - 82% 474 91% 629
(b) Second Turn
Writing model Roleplay ââ GPT-4 ~~ Claude-v1 ââ GPT-3.5-turbo Vicuna-13B â- Alpaca-13B 4qreasoning LLaMA-13B Humanities STEM Extraction Math Coding
Figure 20: Category-wise scores of 6 models on MT-bench.
27
# E Training Details of Vicuna Models
|
2306.05685#98
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 98 |
Xiezhi Xiezhi-Spec Xiezhi-Inter Xiezhi-Train 46 44 43 45 763 44 44 45 729 30 44 29 836 49 31 45 48 78 853 36 46 45 56 745 43 44 45 787 45 45 34 91 869 67 63 58 43 117 45 84 872 69 45 60 45 11148 926 61 61 852 45 45 52 872 46 44 803 34 47 45 959 45 34 43 45 30 77 46 793 45 60 55 796 45 49 45 835 58 59 40 44 91 11238 890 60 1726 148 72 123 44 73
40 39 38 40 48 0 0 0 136 25 39 24 114 0 26 40 0 48 88 0 40 0 0 86 38 0 0 88 0 40 0 48 86 0 0 0 38 88 40 48 176 0 40 48 40 1245 147 48 48 175 40 40 47 48 0 0 48 0 0 0 306 40 29 37 40 25 48 39 136 40 48 0 119 40 39 40 218 48 48 35 39 0 2130 96 48 597 48 48 48 0 48
|
2306.05783#98
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05949
| 98 |
[39] H. Brown, K. Lee, F. Mireshghallah, R. Shokri, and F. Tramèr. What Does it Mean for a Language Model to Preserve Privacy?, Feb. 2022. URL http://arxiv.org/abs/2202. 05520. arXiv:2202.05520 [cs, stat].
[40] T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei. Language Models are Few-Shot Learners, 2020-07-22. URL http://arxiv.org/abs/2005. 14165.
|
2306.05949#98
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06283
| 98 |
[83] Taori, R.; Gulrajani, I.; Zhang, T.; Dubois, Y.; Li, X.; Guestrin, C.; Liang, P.; Hashimoto, T. B. Stanford Alpaca: An Instruction-following LLaMA model. https: //github.com/tatsu-lab/stanford_alpaca, 2023.
[84] Alpaca-LoRA. https://github.com/tloen/alpaca-lora.
[85] Touvron, H.; Lavril, T.; Izacard, G.; Martinet, X.; Lachaux, M.-A.; Lacroix, T.; Rozi`ere, B.; Goyal, N.; Hambro, E.; Azhar, F., et al. Llama: Open and eï¬cient foun- dation language models. arXiv preprint:2302.13971 2023,
[86] Mamaghani, Z. G.; Hawboldt, K. A.; MacQuarrie, S. Adsorption of CO2 using biochar - Review of the impact of gas mixtures and water on adsorption. J. Environ. Chem. Eng. 2023, 11, 109643.
|
2306.06283#98
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 99 |
Figure 20: Category-wise scores of 6 models on MT-bench.
27
# E Training Details of Vicuna Models
Vicuna is created by fine-tuning a LLaMA base model using user-shared conversations gathered from ShareGPT.com with its public APIs. ShareGPT is a website where users can share their ChatGPT conversations. To ensure data quality, we convert the HTML back to markdown and filter out some inappropriate or low-quality samples, which results in 125K conversations after data cleaning.4 We then divide lengthy conversations into smaller segments that fit the modelâs maximum context length.
We construct three training datasets with different scales from this cleaned ShareGPT dataset. Their statistics are in Table 8, where we also compare it with Alpaca [38] dataset. âAllâ is the full dataset. âSingleâ only includes the first turn of each conversation. âSelectedâ is a small high-quality dataset of 3K sequences. To construct the âSelectedâ dataset, we pick sequences that include at least 3 turns of conversations generated by GPT-4 and run a clustering algorithm to divide them into 3K clusters and pick the centroid of each cluster.
|
2306.05685#99
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 99 |
1 0 0 0 0 0 0 0 3 0 0 0 9 4 0 0 2 0 6 0 1 0 0 5 0 0 0 8 0 0 5 0 7 3 2 0 0 4 0 0 5 3 0 0 0 142 44 2 0 1 0 0 0 24 2 0 11 0 1 0 29 0 0 1 0 0 21 2 1 0 1 0 2 0 2 0 2 0 0 0 0 4 271 52 0 127 46 9 29 3 5
5 5 5 5 20 5 5 5 20 5 5 5 25 5 5 5 5 5 25 5 5 5 5 20 5 5 5 20 5 5 5 5 25 5 5 5 5 10 5 5 25 5 5 5 5 185 15 5 5 20 5 5 5 15 5 5 20 5 5 5 40 5 5 5 5 5 5 5 20 5 5 5 20 5 5 5 25 5 5 5 5 5 325 10 5 70 5 5 5 5 5
# Animal Husbandry
Animal Genetic Breeding and Reproduction Animal Nutrition and Feed Science Grassland Science Special economic animal husbandry (including: silkworms, bees, etc.)
# Herbology
# Medicine
# Traditional Chinese Medicine
# Chinese herbal medicine (Level 2 discipline)
# Chinese Medicine
Basic Theory of Chinese Medicine Clinical Foundations of Chinese Medicine Chinese Medical History and Literature Formulary Diagnostics of Chinese Medicine
19
|
2306.05783#99
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
2306.05949
| 99 |
[41] B. Buchanan, A. Lohn, M. Musser, and K. Sedova. Truth, Lies, and Au- URL https://cset.georgetown.edu/publication/ tomation, May truth-lies-and-automation/. 2021.
[42] D. Bui, B. Tang, and K. G. Shin. Do Opt-Outs Really Opt Me Out? In Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security, pages 425â439, Los Angeles CA USA, Nov. 2022. ACM. ISBN 978-1-4503-9450-5. doi: 10.1145/3548606. 3560574. URL https://dl.acm.org/doi/10.1145/3548606.3560574.
[43] J. Buolamwini and T. Gebru. Gender Shades: Intersectional Accuracy Disparities in Com- mercial Gender Classification. In S. A. Friedler and C. Wilson, editors, Proceedings of the 1st Conference on Fairness, Accountability and Transparency, volume 81 of Proceedings of Machine Learning Research, pages 77â91, New York, NY, USA, Feb. 2018. PMLR. URL http://proceedings.mlr.press/v81/buolamwini18a.html.
|
2306.05949#99
|
Evaluating the Social Impact of Generative AI Systems in Systems and Society
|
Generative AI systems across modalities, ranging from text, image, audio, and
video, have broad social impacts, but there exists no official standard for
means of evaluating those impacts and which impacts should be evaluated. We
move toward a standard approach in evaluating a generative AI system for any
modality, in two overarching categories: what is able to be evaluated in a base
system that has no predetermined application and what is able to be evaluated
in society. We describe specific social impact categories and how to approach
and conduct evaluations in the base technical system, then in people and
society. Our framework for a base system defines seven categories of social
impact: bias, stereotypes, and representational harms; cultural values and
sensitive content; disparate performance; privacy and data protection;
financial costs; environmental costs; and data and content moderation labor
costs. Suggested methods for evaluation apply to all modalities and analyses of
the limitations of existing evaluations serve as a starting point for necessary
investment in future evaluations. We offer five overarching categories for what
is able to be evaluated in society, each with their own subcategories:
trustworthiness and autonomy; inequality, marginalization, and violence;
concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm. We
are concurrently crafting an evaluation repository for the AI research
community to contribute existing evaluations along the given categories. This
version will be updated following a CRAFT session at ACM FAccT 2023.
|
http://arxiv.org/pdf/2306.05949
|
Irene Solaiman, Zeerak Talat, William Agnew, Lama Ahmad, Dylan Baker, Su Lin Blodgett, Hal Daumé III, Jesse Dodge, Ellie Evans, Sara Hooker, Yacine Jernite, Alexandra Sasha Luccioni, Alberto Lusoli, Margaret Mitchell, Jessica Newman, Marie-Therese Png, Andrew Strait, Apostol Vassilev
|
cs.CY, cs.AI
| null | null |
cs.CY
|
20230609
|
20230612
|
[
{
"id": "2007.04068"
},
{
"id": "2305.09800"
},
{
"id": "1908.09203"
},
{
"id": "2202.05520"
},
{
"id": "2302.10329"
},
{
"id": "2107.03374"
},
{
"id": "2210.06245"
},
{
"id": "2211.02001"
},
{
"id": "2212.08073"
},
{
"id": "2303.08774"
},
{
"id": "2301.10226"
},
{
"id": "2202.02647"
},
{
"id": "2112.10752"
},
{
"id": "2206.04615"
},
{
"id": "2202.00885"
},
{
"id": "2010.15581"
},
{
"id": "2305.09941"
},
{
"id": "2301.04246"
},
{
"id": "2304.12298"
},
{
"id": "2203.09509"
},
{
"id": "2207.14157"
},
{
"id": "2102.09692"
},
{
"id": "1804.10999"
},
{
"id": "2303.11156"
},
{
"id": "2104.06390"
},
{
"id": "2002.05651"
}
] |
2306.06283
| 99 |
[87] Peng, Y.; Krungleviciute, V.; Eryazici, I.; Hupp, J. T.; Farha, O. K.; Yildirim, T. Methane Storage in MetalâOrganic Frameworks: Current Records, Surprise Findings, and Chal- lenges. J. Am. Chem. Soc. 2013, 135, 11887â11894.
[88] Sahoo, B.; Pandey, V.; Dogonchi, A.; Mohapatra, P.; Thatoi, D.; Nayak, N.; Nayak, M. A state-of-art review on 2D material-boosted metal oxide nanoparticle electrodes: Super- capacitor applications. J. Energy Storage 2023, 65, 107335.
[89] Suppiah, D. D.; Daud, W. M. A. W.; Johan, M. R. Supported Metal Oxide Catalysts for CO2 FischerâTropsch Conversion to Liquid Fuels-A Review. Energy Fuels. 2021, 35, 17261â17278.
|
2306.06283#99
|
14 Examples of How LLMs Can Transform Materials Science and Chemistry: A Reflection on a Large Language Model Hackathon
|
Large-language models (LLMs) such as GPT-4 caught the interest of many
scientists. Recent studies suggested that these models could be useful in
chemistry and materials science. To explore these possibilities, we organized a
hackathon.
This article chronicles the projects built as part of this hackathon.
Participants employed LLMs for various applications, including predicting
properties of molecules and materials, designing novel interfaces for tools,
extracting knowledge from unstructured data, and developing new educational
applications.
The diverse topics and the fact that working prototypes could be generated in
less than two days highlight that LLMs will profoundly impact the future of our
fields. The rich collection of ideas and projects also indicates that the
applications of LLMs are not limited to materials science and chemistry but
offer potential benefits to a wide range of scientific disciplines.
|
http://arxiv.org/pdf/2306.06283
|
Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M Bran, Stefan Bringuier, L. Catherine Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nicolas Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Imran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majumdar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel G. Rodriques, Jacob N. Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben E. Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean Warren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scourtas, KJ Schmidt, Ian Foster, Andrew D. White, Ben Blaiszik
|
cond-mat.mtrl-sci, cs.LG, physics.chem-ph
| null | null |
cond-mat.mtrl-sci
|
20230609
|
20230714
|
[
{
"id": "2209.08203"
},
{
"id": "2212.04450"
}
] |
2306.05685
| 100 |
All models (Vicuna-7B/13B) are trained with the same hyperparameters: global batch size=128, learning=2e-5, epochs=3, seq length=2048. Except for âSelectedâ, which we train for 5 epochs. The training code is built on top of the Alpaca code but additionally handles multi-turn conversations. The training is done with 8x A100 GPUs. The longest single training run takes around 2 days. We utilize SkyPilot [49] managed spot instances for saving training costs and FlashAttention [11] for memory optimizations. The training code is available at https://github.com/lm-sys/FastChat.
# Table 14: Dataset statistics Selected Alpaca
Dataset Name Single All #Token #Sequence Avg. turns of conversation Avg. response length (token) 4.4M 52K 1.0 65 4.8M 3K 4.0 343 184M 370M 257K 257K 2.9 1.0 373 473
4In this study, we use more data (125K) than the version in our earlier blog post (70K).
28
# F Exploring Vicuna as a judge
|
2306.05685#100
|
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
|
Evaluating large language model (LLM) based chat assistants is challenging
due to their broad capabilities and the inadequacy of existing benchmarks in
measuring human preferences. To address this, we explore using strong LLMs as
judges to evaluate these models on more open-ended questions. We examine the
usage and limitations of LLM-as-a-judge, including position, verbosity, and
self-enhancement biases, as well as limited reasoning ability, and propose
solutions to mitigate some of them. We then verify the agreement between LLM
judges and human preferences by introducing two benchmarks: MT-bench, a
multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our
results reveal that strong LLM judges like GPT-4 can match both controlled and
crowdsourced human preferences well, achieving over 80% agreement, the same
level of agreement between humans. Hence, LLM-as-a-judge is a scalable and
explainable way to approximate human preferences, which are otherwise very
expensive to obtain. Additionally, we show our benchmark and traditional
benchmarks complement each other by evaluating several variants of LLaMA and
Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with
human preferences are publicly available at
https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
|
http://arxiv.org/pdf/2306.05685
|
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica
|
cs.CL, cs.AI
|
NeurIPS 2023 Datasets and Benchmarks Track
| null |
cs.CL
|
20230609
|
20231224
|
[
{
"id": "2302.13971"
},
{
"id": "1905.07830"
},
{
"id": "1803.05457"
},
{
"id": "2305.10403"
},
{
"id": "2304.07327"
},
{
"id": "2201.11903"
},
{
"id": "2009.03300"
},
{
"id": "2304.12244"
},
{
"id": "2306.12420"
},
{
"id": "2304.06364"
},
{
"id": "2107.03374"
},
{
"id": "2306.04751"
},
{
"id": "2211.09110"
},
{
"id": "2301.13688"
},
{
"id": "2004.14602"
},
{
"id": "2110.14168"
},
{
"id": "2305.15717"
},
{
"id": "2211.05719"
},
{
"id": "2206.04615"
},
{
"id": "2204.05862"
},
{
"id": "2305.01937"
},
{
"id": "2305.14387"
},
{
"id": "2305.17926"
},
{
"id": "2304.03277"
},
{
"id": "2303.12712"
},
{
"id": "2305.14314"
},
{
"id": "2303.15056"
},
{
"id": "2109.01652"
},
{
"id": "2305.11206"
},
{
"id": "2109.07958"
},
{
"id": "2302.07736"
}
] |
2306.05783
| 100 |
19
Internal Medicine External Medicine Orthopaedics and Traumatology of Chinese Medicine TCM Gynecology TCM Paediatrics Traditional Chinese Medicine Acupuncture, Moxibustion and Tuina Ethnic Medicine Combination of Chinese and Western Medicine Fundamentals of Integrated Chinese and Western Medicine Combined Chinese and Western Medicine Clinical Clinical Medicine Internal Medicine Pediatrics Geriatric Medicine Neurology Psychiatry and Mental Health Dermatology and Venereology Imaging Medicine and Nuclear Medicine Diagnostic Clinical Laboratory Medicine Nursing Surgery Obstetrics and Gynaecology Ophthalmology Otorhinolaryngology Oncology Rehabilitation Medicine and Physiotherapy Sports Medicine Anaesthesiology Emergency Medicine Public Health and Preventive Medicine Epidemiology and Health Statistics Labour Hygiene and Environmental Health Nutrition and Food Hygiene Paediatric and Child Health and Maternal and Child Health Health Toxicology Military Preventive Medicine Dentistry Basic Dental Medicine Dental Clinical Medicine Basic Medicine Human Anatomy and Histoplasty Immunology Pathogenic Biology Pathology and Pathophysiology Forensic Medicine Radiological Medicine Aviation, Aerospace and Maritime Medicine Nursing Speciality Medicine Pharmacy Medicinal Chemistry Pharmacy Biopharmaceutics Pharmaceutical Analysis Microbiology and Biochemical Pharmacy Pharmacology Military Political Work Military Political Work Studies (Level 2 discipline) Military Logistics Military logistics and military equipment science Military Logistics Rear professional logistics Military Equipment Studies Military Thought and Military History Military Thought Military History Military Equipment Studies Military Training Military Systems Military Organization Military Management Military Command
|
2306.05783#100
|
Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation
|
New Natural Langauge Process~(NLP) benchmarks are urgently needed to align
with the rapid development of large language models (LLMs). We present Xiezhi,
the most comprehensive evaluation suite designed to assess holistic domain
knowledge. Xiezhi comprises multiple-choice questions across 516 diverse
disciplines ranging from 13 different subjects with 249,587 questions and
accompanied by Xiezhi-Specialty and Xiezhi-Interdiscipline, both with 15k
questions. We conduct evaluation of the 47 cutting-edge LLMs on Xiezhi. Results
indicate that LLMs exceed average performance of humans in science,
engineering, agronomy, medicine, and art, but fall short in economics,
jurisprudence, pedagogy, literature, history, and management. We anticipate
Xiezhi will help analyze important strengths and shortcomings of LLMs, and the
benchmark is released in~\url{https://github.com/MikeGu721/XiezhiBenchmark}.
|
http://arxiv.org/pdf/2306.05783
|
Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu He, Rui Xu, Wenhao Huang, Zili Wang, Shusen Wang, Weiguo Zheng, Hongwei Feng, Yanghua Xiao
|
cs.CL
|
Under review of NeurIPS 2023
| null |
cs.CL
|
20230609
|
20230615
|
[
{
"id": "2301.13126"
},
{
"id": "2302.13971"
},
{
"id": "2303.04048"
},
{
"id": "1905.07830"
},
{
"id": "2304.12986"
},
{
"id": "2304.07854"
},
{
"id": "2211.05100"
},
{
"id": "1909.00277"
},
{
"id": "2305.10263"
},
{
"id": "1909.06058"
},
{
"id": "2206.07682"
},
{
"id": "2304.06364"
},
{
"id": "2211.09110"
},
{
"id": "2305.08322"
},
{
"id": "2210.11416"
},
{
"id": "2212.08073"
},
{
"id": "2210.09261"
},
{
"id": "2206.04615"
},
{
"id": "2303.18223"
},
{
"id": "2302.09419"
},
{
"id": "2303.14742"
},
{
"id": "2111.10952"
},
{
"id": "2301.12726"
},
{
"id": "2304.01933"
},
{
"id": "2106.09685"
},
{
"id": "2303.12712"
},
{
"id": "2212.09251"
},
{
"id": "2304.01196"
},
{
"id": "2105.09938"
}
] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.