arxiv_id
stringlengths 9
12
| abstract
stringlengths 431
13.8k
|
|---|---|
2404.04633 | To answer a question, language models often need to integrate prior knowledge learned during pretraining and new information presented in context.
We hypothesize that models perform this integration in a predictable way across different questions and contexts: models will rely more on prior knowledge for questions about entities (e.g., persons, places, etc.) that they are more familiar with due to higher exposure in the training corpus, and be more easily persuaded by some contexts than others.
To formalize this problem, we propose two mutual information-based metrics to measure a model's dependency on a context and on its prior about an entity: first, thepersuasion scoreof a given context represents how much a model depends on the context in its decision, and second, thesusceptibility scoreof a given entity represents how much the model can be swayed away from its original answer distribution about an entity.
Following well-established measurement modeling methods, we empirically test for the validity and reliability of these metrics.
Finally, we explore and find a relationship between the scores and the model's expected familiarity with an entity, and provide two use cases to illustrate their benefits. https://github.com/kdu4108/measureLM |
2404.19705 | In this paper, we demonstrate how Large Language Models (LLMs) can effectively learn to use an off-the-shelf information retrieval (IR) system specifically when additional context is required to answer a given question.
Given the performance of IR systems, the optimal strategy for question answering does not always entail external information retrieval; rather, it often involves leveraging the parametric memory of the LLM itself. Prior research has identified this phenomenon in the PopQA dataset, wherein the most popular questions are effectively addressed using the LLM’s parametric memory, while less popular ones require IR system usage. Following this, we propose a tailored training approach for LLMs, leveraging existing open-domain question answering datasets. Here, LLMs are trained to generate a special token,RET, when they do not know the answer to a question. Our evaluation of the Adaptive Retrieval LLM (Adapt-LLM) on the PopQA dataset showcases improvements over the same LLM under three configurations: (i) retrieving information for all the questions, (ii) using always the parametric memory of the LLM, and (iii) using a popularity threshold to decide when to use a retriever. Through our analysis, we demonstrate thatAdapt-LLMis able to generate theRETtoken when it determines that it does not know how to answer a question, indicating the need for IR, while it achieves notably high accuracy levels when it chooses to rely only on its parametric memory. |
2405.01535 | Proprietary LMs such as GPT-4 are often employed to assess the quality of responses from various LMs. However, concerns including transparency, controllability, and affordability strongly motivate the development of open-source LMs specialized in evaluations. On the other hand, existing open evaluator LMs exhibit critical shortcomings: 1) they issue scores that significantly diverge from those assigned by humans, and 2) they lack the flexibility to perform both direct assessment and pairwise ranking, the two most prevalent forms of assessment. Additionally, they do not possess the ability to evaluate based oncustom evaluation criteria, focusing instead on general attributes like helpfulness and harmlessness. To address these issues, we introduce Prometheus 2, a more powerful evaluator LM than it’s predecessor that closely mirrors human and GPT-4 judgements. Moreover, it is capable of processing both direct assessment and pair-wise ranking formats grouped with a user-defined evaluation criteria. On four direct assessment benchmarks and four pairwise ranking benchmarks,Prometheus 2scores the highest correlation and agreement with humans and proprietary LM judges among all tested open evaluator LMs. Our models, code, and data are all publicly available111https://github.com/prometheus-eval/prometheus-eval. |
2306.12747 | Recent works have shown that line search methods can speed up Stochastic Gradient Descent (SGD) and Adam in modern over-parameterized settings. However, existing line searches may take steps that are smaller than necessary since they require a monotone decrease of the (mini-) batch objective function. We explore nonmonotone line search methods to relax this condition and possibly accept larger step sizes. Despite the lack of a monotonic decrease, we prove the same fast rates of convergence as in the monotone case. Our experiments show that nonmonotone methods improve the speed of convergence and generalization properties of SGD/Adam even beyond the previous monotone line searches. We propose a POlyak NOnmonotone Stochastic (PoNoS) method, obtained by combining a nonmonotone line search with a Polyak initial step size. Furthermore, we develop a new resetting technique that in the majority of the iterations reduces the amount of backtracks to zero while still maintaining a large initial step size. To the best of our knowledge, a first runtime comparison shows that the epoch-wise advantage of line-search-based methods gets reflected in the overall computational time. |
2405.00332v1 | Large language models (LLMs) have achieved impressive success on many benchmarks for mathematical reasoning.
However, there is growing concern that some of this performance actually reflects dataset contamination, where data closely resembling benchmark questions leaks into the training data, instead of true reasoning ability.
To investigate this claim rigorously, we commissionGrade School Math 1000(GSM1k) . GSM1k is designed to mirror the style and complexity of the established GSM8k benchmark,
the gold standard for measuring elementary mathematical reasoning. We ensure that the two benchmarks are comparable across important metrics such as human solve rates, number of steps in solution, answer magnitude, and more.
When evaluating leading open- and closed-source LLMs on GSM1k, we observe accuracy drops of up to 13%, with several families of models (e.g. Phi and Mistral) showing evidence of systematic overfitting across almost all model sizes.
At the same time, many models, especially those on the frontier, (e.g. Gemini/GPT/Claude) show minimal signs of overfitting.
Further analysis suggests a positive relationship (Spearman’s) between a model’s probability of generating an example from GSM8k and its performance gap between GSM8k and GSM1k, suggesting that many models may have partially memorized GSM8k. |
2404.19733 | Iterative preference optimization methods have recently been shown to
perform well for
general instruction tuning tasks, but typically make little improvement on
reasoning tasks.
In this work we develop an iterative approach that optimizes the preference between competing generated Chain-of-Thought (CoT) candidates by optimizing for winning vs. losing reasoning steps that lead to the correct answer. We train
using a modified DPO losswith an additional
negative log-likelihood term, which we find to be crucial. We show reasoning improves across repeated iterations of this scheme. While only relying on examples in the training set, our approach results in increasing accuracy for Llama-2-70B-Chat from 55.6% to 81.6% on GSM8K (and 88.7% with majority voting out of 32 samples) , from 12.5% to 20.8% on MATH, and from 77.8% to 86.7% on ARC-Challenge, which outperforms other Llama-2-based models not relying on additionally sourced datasets. |
2405.00675 | Traditional reinforcement learning from human feedback (RLHF) approaches relying on parametric models like the Bradley-Terry model fall short in capturing the intransitivity and irrationality in human preferences. Recent advancements suggest that directly working with preference probabilities can yield a more accurate reflection of human preferences, enabling more flexible and accurate language model alignment. In this paper, we propose a self-play-based method for language model alignment, which treats the problem as a constant-sum two-player game aimed at identifying the Nash equilibrium policy. Our approach, dubbedSelf-Play Preference Optimization(SPPO) ,
approximates the Nash equilibrium through iterative policy updates and enjoys a theoretical convergence guarantee. Our method can effectively increase the log-likelihood of the chosen response and decrease that of the rejected response, which cannot be trivially achieved by symmetric pairwise loss such as Direct Preference Optimization (DPO) and Identity Preference Optimization (IPO) . In our experiments, using only 60k prompts (without responses) from the UltraFeedback dataset and without any prompt augmentation, by leveraging a pre-trained preference model PairRM with only 0.4B parameters, SPPO can obtain a model from fine-tuning Mistral-7B-Instruct-v0.2 that achieves the state-of-the-art length-controlled win-rate of 28.53% against GPT-4-Turbo on AlpacaEval 2.0. It also outperforms the (iterative) DPO and IPO on MT-Bench and the Open LLM Leaderboard. Notably, the strong performance of SPPO is achieved without additional external supervision (e.g., responses, preferences, etc.) from GPT-4 or other stronger language models. |
2304.01982 | Multi-vector retrieval models such as ColBERTallow token-level interactions between queries and documents, and hence achieve state of the art on many information retrieval benchmarks.
However, their non-linear scoring function cannot be scaled to millions of documents, necessitating a three-stage process for inference: retrieving initial candidates via token retrieval, accessing all token vectors, and scoring the initial candidate documents.
The non-linear scoring function is applied over all token vectors of each candidate document, making the inference process complicated and slow.
In this paper, we aim to simplify the multi-vector retrieval by rethinking the role of token retrieval. We present XTR, ConteXtualizedTokenRetriever, which introduces a simple, yet novel, objective function that encourages the model to retrieve the most important document tokens first.
The improvement to token retrieval allows XTR to rank candidates only using the retrieved tokens rather than all tokens in the document, and enables a newly designed scoring stage that is two-to-three orders of magnitude cheaper than that of ColBERT.
On the popular BEIR benchmark,XTRadvances the state-of-the-art by 2.8 nDCG@10 without any distillation.
Detailed analysis confirms our decision to revisit the token retrieval stage, asXTRdemonstrates much better recall of the token retrieval stage compared to ColBERT. |
2303.08896 | Generative Large Language Models (LLMs) such as GPT-3 are capable of generating highly fluent responses to a wide variety of user prompts. However, LLMs are known to hallucinate facts and make non-factual statements which can undermine trust in their output. Existing fact-checking approaches either require access to the output probability distribution (which may not be available for systems such as ChatGPT) or external databases that are interfaced via separate, often complex, modules. In this work, we propose "SelfCheckGPT", a simple sampling-based approach that can be used to fact-check the responses of black-box models in a zero-resource fashion, i.e. without an external database. SelfCheckGPT leverages the simple idea that if an LLM has knowledge of a given concept, sampled responses are likely to be similar and contain consistent facts. However, for hallucinated facts, stochastically sampled responses are likely to diverge and contradict one another. We investigate this approach by using GPT-3 to generate passages about individuals from the WikiBio dataset, and manually annotate the factuality of the generated passages. We demonstrate that SelfCheckGPT can: i) detect non-factual and factual sentences; and ii) rank passages in terms of factuality. We compare our approach to several baselines and show that our approach has considerably higher AUC-PR scores in sentence-level hallucination detection and higher correlation scores in passage-level factuality assessment compared to grey-box methods.111Code and dataset can be found on the project page athttps://github.com/potsawee/selfcheckgpt. |
2402.09371 | Length generalization, defined as the ability to extrapolate from shorter training sequences to longer test ones, is a significant challenge for language models. This issue persists even with large-scale Transformers handling relatively straightforward tasks. In this paper, we test the Transformer’s ability of length generalization using the task of addition of two integers. We show that the success of length generalization is intricately linked to the data format and the type of position encoding. Using the right combination of data format and position encodings, we show for the first time that standard Transformers can extrapolate to a sequence length that isthe input length. Nevertheless, unlike in-distribution generalization, length generalization remains fragile, significantly influenced by factors like random weight initialization and training data order, leading to large variances across different random seeds. |
2404.07503 | The success of AI models relies on the availability of large, diverse, and high-quality datasets, which can be challenging to obtain due to data scarcity, privacy concerns, and high costs. Synthetic data has emerged as a promising solution by generating artificial data that mimics real-world patterns. This paper provides an overview of synthetic data research, discussing its applications, challenges, and future directions. We present empirical evidence from prior art to demonstrate its effectiveness and highlight the importance of ensuring its factuality, fidelity, and unbiasedness. We emphasize the need for responsible use of synthetic data to build more powerful, inclusive, and trustworthy language models. |
2404.14723 | Large Language Models (LLMs) have demonstrated remarkable performance across a spectrum of tasks. Recently, Direct Preference Optimization (DPO) has emerged as an RL-free approach to optimize the policy model on human preferences. However, several limitations hinder the widespread adoption of this method. To address these shortcomings, various versions of DPO have been introduced. Yet, a comprehensive evaluation of these variants across diverse tasks is still lacking. In this study, we aim to bridge this gap by investigating the performance of alignment methods across three distinct scenarios:(1) keeping the Supervised Fine-Tuning (SFT) part,(2) skipping the SFT part, and(3) skipping the SFT part and utilizing an instruction-tuned model. Furthermore, we explore the impact of different training sizes on their performance.
Our evaluation spans a range of tasks including dialogue systems, reasoning, mathematical problem-solving, question answering, truthfulness, and multi-task understanding, encompassing 13 benchmarks such as MT-Bench, Big Bench, and Open LLM Leaderboard.
Key observations reveal that alignment methods achieve optimal performance with smaller training data subsets, exhibit limited effectiveness in reasoning tasks yet significantly impact mathematical problem-solving, and employing an instruction-tuned model notably influences truthfulness.
We anticipate that our findings will catalyze further research aimed at developing more robust models to address alignment challenges. |
2403.14608 | Large models represent a groundbreaking advancement in multiple application fields, enabling remarkable achievements across various tasks. However, their unprecedented scale comes with significant computational costs. These models, often consisting of billions of parameters, require vast amounts of computational resources for execution. Especially, the expansive scale and computational demands pose considerable challenges when customizing them for particular downstream tasks, particularly over the hardware platforms constrained by computational capabilities. Parameter Efficient Fine-Tuning (PEFT) provides a practical solution by efficiently adapt the large models over the various downstream tasks. In particular, PEFT refers to the process of adjusting the parameters of a pre-trained large models to adapt it to a specific task or domain while minimizing the number of additional parameters introduced or computational resources required. This approach is particularly important when dealing with large-scale language models with high parameter counts, as fine-tuning these models from scratch can be computationally expensive and resource-intensive, posing considerable challenges in the supporting system platform design. In this survey, we present comprehensive studies of various PEFT algorithms, examining their performance and computational overhead. Moreover, we provide an overview of applications developed using different PEFT algorithms and discuss common techniques employed to mitigate computation costs for PEFT.
In addition to the algorithmic perspective, we overview various real-world system designs to investigate the implementation costs associated with different PEFT algorithms. This survey serves as an indispensable resource for researchers aiming to understand both the PEFT algorithm and its system implementation, offering detailed insights into recent advancements and practical applications. |
2307.02477 | The impressive performance of recent language models across a wide range of tasks suggests that they possess a degree of abstract reasoning skills. Are these skills general and transferable, or specialized to specific tasks seen during pretraining? To disentangle these effects, we propose an evaluation framework based on “counterfactual” task variants that deviate from
the default assumptions underlying standard tasks. Across a suite of 11 tasks, we observe nontrivial performance on the counterfactual variants, but nevertheless find that performance substantially and consistently degrades compared to the default conditions.
This suggests that while current LMs may possess abstract task-solving skills to a degree, they often also rely on narrow, non-transferable procedures for task-solving.
These results motivate a more careful interpretation of language model performance that teases apart these aspects of behavior. |
2311.12997 | Transformers trained on huge text corpora exhibit a remarkable set of capabilities, e.g., performing basic arithmetic. Given the inherent compositional nature of language, one can expect the model to learn to compose these capabilities, potentially yielding acombinatorial explosionof what operations it can perform on an input. Motivated by the above, we train autoregressive Transformer models on a synthetic data-generating process that involves compositions of a set of well-defined monolithic capabilities. Through a series of extensive and systematic experiments on this data-generating process, we show that:(1) autoregressive Transformers can learn compositional structures from small amounts of training data and generalize to exponentially or even combinatorially many functions;(2) generating intermediate outputs when composing functions is more effective for generalizing to new, unseen compositions than not generating any intermediate outputs(3) biases in the order of the compositions in the training data result in Transformers that fail to compose some combinations of functions; and(4) the attention layers select which capability to apply while the feed-forward layers execute the selected capability. |
2207.02098 | Reliable generalization lies at the heart of safe ML and AI.
However, understanding when and how neural networks generalize remains one of the most important unsolved problems in the field.
In this work, we conduct an extensive empirical study (models, 15 tasks) to investigate whether insights from the theory of computation can predict the limits of neural network generalization in practice.
We demonstrate that grouping tasks according to the Chomsky hierarchy allows us to forecast whether certain architectures will be able to generalize to out-of-distribution inputs.
This includes negative results where even extensive amounts of data and training time never lead to any non-trivial generalization, despite models having sufficient capacity to fit the training data perfectly.
Our results show that, for our subset of tasks, RNNs and Transformers fail to generalize on non-regular tasks, LSTMs can solve regular and counter-language tasks, and only networks augmented with structured memory (such as a stack or memory tape) can successfully generalize on context-free and context-sensitive tasks. |
2404.04237 | The rapid progress of large language models (LLMs) has seen them excel and frequently surpass human performance on standard benchmarks. This has enabled many downstream applications, such as LLM agents, to rely on their sophisticated reasoning to navigate complex task requirements. However, LLMs are known to unexpectedly falter in simple tasks and under seemingly straightforward circumstances - underscoring the need for better and more diverse evaluation setups to measure their true capabilities. To this end, we choose to study compositional and conditional reasoning, two cornerstones of human cognition, and introduce GroundCocoa - a lexically diverse benchmark connecting these reasoning skills to the real-world problem of flight booking. Our task involves aligning detailed user preferences with available flight options presented in a multiple-choice format. Results indicate a significant disparity in performance among current state-of-the-art LLMs with even the best performing model, GPT-4 Turbo, not exceeding 67% accuracy despite advanced prompting techniques. |
1901.05103 | Computer graphics, 3D computer vision and robotics communities have produced multiple approaches to representing 3D geometry for rendering and reconstruction. These provide trade-offs across fidelity, efficiency and compression capabilities. In this work, we introduce DeepSDF, a learned continuous Signed Distance Function (SDF) representation of a class of shapes that enables high quality shape representation, interpolation and completion from partial and noisy 3D input data. DeepSDF, like its classical counterpart, represents a shape’s surface by a continuous volumetric field: the magnitude of a point in the field represents the distance to the surface boundary and the sign indicates whether the region is inside (-) or outside (+) of the shape, hence our representation implicitly encodes a shape’s boundary as the zero-level-set of the learned function while explicitly representing the classification of space as being part of the shapes interior or not. While classical SDF’s both in analytical or discretized voxel form typically represent the surface of a single shape, DeepSDF can represent an entire class of shapes. Furthermore, we show state-of-the-art performance for learned 3D shape representation and completion while reducing the model size by an order of magnitude compared with previous work. |
2307.13702 | Large language models (LLMs) perform better when they produce step-by-step, “Chain-of-Thought” (CoT) reasoning before answering a question, but it is unclear if the stated reasoning is a faithful explanation of the model’s actual reasoning (i.e., its process for answering the question) .
We investigate hypotheses for how CoT reasoning may be unfaithful, by examining how the model predictions change when we intervene on the CoT (e.g., by adding mistakes or paraphrasing it) .
Models show large variation across tasks in how strongly they condition on the CoT when predicting their answer, sometimes relying heavily on the CoT and other times primarily ignoring it.
CoT’s performance boost does not seem to come from CoT’s added test-time compute alone or from information encoded via the particular phrasing of the CoT.
As models become larger and more capable, they produce less faithful reasoning on most tasks we study.
Overall, our results suggest that CoT can be faithful if the circumstances such as the model size and task are carefully chosen. |
2301.13379 | While Chain-of-Thought (CoT) prompting boosts Language Models’ (LM) performance on a gamut of complex reasoning tasks, the generated reasoning chain does not necessarily reflect how the model arrives at the answer (aka.faithfulness) . We proposeFaithful CoT, a reasoning framework involving two stages:Translation(Natural Language querysymbolic reasoning chain) andProblem Solving(reasoning chainanswer) , using an LM and a deterministic solver respectively. This guarantees that the reasoning chain provides a faithful explanation of the final answer. Aside from interpretability, Faithful CoT also improves empirical performance: it outperforms standard CoT on 9 of 10 benchmarks from 4 diverse domains, with a relative accuracy gain of 6.3% on Math Word Problems (MWP) , 3.4% on Planning, 5.5% on Multi-hop Question Answering (QA) , and 21.4% on Relational Inference. Furthermore, with GPT-4 and Codex, it sets the new state-of-the-art few-shot performance on 7 datasets (with 95.0+ accuracy on 6 of them) , showing a strong synergy between faithfulness and accuracy.111Our code, data, and prompts are available athttps://github.com/veronica320/Faithful-COT. |
2404.15758 | Chain-of-thought responses from language models improve performance across most benchmarks. However, it remains unclear to what extent these performance gains can be attributed to human-like task decomposition or simply the greater computation that additional tokens allow. We show that transformers can use meaningless filler tokens (e.g., ‘……’) in place of a chain of thought to solve two hard algorithmic tasks they could not solve when responding without intermediate tokens. However, we find empirically thatlearningto use filler tokens is difficult and requires specific, dense supervision to converge. We also provide a theoretical characterization of the class of problems where filler tokens are useful in terms of thequantifier depthof a first-order formula. For problems satisfying this characterization, chain-of-thought tokens need not provide information about the intermediate computational steps involved in multi-token computations. In summary, our results show that additional tokens can provide computational benefits independent of token choice. The fact that intermediate tokens can act as filler tokens raises concerns about large language models engaging in unauditable, hidden computations that are increasingly detached from the observed chain-of-thought tokens.111Code is available athttps://github.com/JacobPfau/fillerTokens |
2402.12354 | In this paper, we show that Low Rank Adaptation (LoRA) as originally introduced inleads to suboptimal finetuning of models with large width (embedding dimension) . This is due to the fact that adapter matricesandin LoRA are updated with the same learning rate. Using scaling arguments for large width networks, we demonstrate that using the same learning rate foranddoes not allow efficient feature learning. We then show that this suboptimality of LoRA can be corrected simply by setting different learning rates for the LoRA adapter matricesandwith a well-chosen fixed ratio. We call this proposed algorithm LoRA. In our extensive experiments, LoRAimproves performance (improvements) and finetuning speed (up toX SpeedUp) , at the same computational cost as LoRA. |
2404.14047 | Meta’sLLaMAfamily has become one of the most powerful open-source Large Language Model (LLM) series. Notably,LLaMA3models have recently been released and achieve impressive performance across various with super-large scale pre-training on over 15T tokens of data. Given the wide application of low-bit quantization for LLMs in resource-limited scenarios, we exploreLLaMA3’s capabilities when quantized to low bit-width. This exploration holds the potential to unveil new insights and challenges for low-bit quantization ofLLaMA3and other forthcoming LLMs, especially in addressing performance degradation problems that suffer in LLM compression. Specifically, we evaluate the 10 existing post-training quantization and LoRA-finetuning methods ofLLaMA3on 1-8 bits and diverse datasets to comprehensively revealLLaMA3’s low-bit quantization performance. Our experiment results indicate thatLLaMA3still suffers non-negligent degradation in these scenarios, especially in ultra-low bit-width. This highlights the significant performance gap under low bit-width that needs to be bridged in future developments. We expect that this empirical study will prove valuable in advancing future models, pushing the LLMs to lower bit-width with higher accuracy for being practical. Our project is released onhttps://github.com/Macaronlin/LLaMA3-Quantizationand quantizedLLaMA3models are released inhttps://huggingface.co/LLMQ. |
2310.01448 | Do large language models (LLMs) genuinely understand the semantics of the language, or just memorize the training data? The recent concern on potential data contamination of LLMs has raised awareness of the community to conduct research on LLMs evaluation. In this paper, we proposeMSTemp, an approach that creates meta semantic templates to evaluate the semantic understanding ability of LLMs. The core ofMSTempis not to perform evaluation directly on existing benchmark datasets, but togeneratenew out-of-distribution (OOD) evaluation sets using existing datasets asseeds. Specifically, for a given sentence,MSTempleverages another language model to generate new samples while preserving its semantics. The new samples are called semantic templates to the original sentence. Then,MSTempgenerates evaluation samples via sentence parsing and random word replacement on the semantic templates.MSTempis highly flexible, dynamic, and cost-effective. Our initial experiments show thatMSTemp-generated samples can significantly reduce the performance of LLMs using existing datasets as seeds. We hope this initial work can shed light on future research of LLMs evaluation. |
2403.18058 | Recently, there have been significant advancements in large language models (LLMs) , particularly focused on the English language.
These advancements have enabled these LLMs to understand and execute complex instructions with unprecedented accuracy and fluency.
However, despite these advancements, there remains a noticeable gap in the development of Chinese instruction tuning.
The unique linguistic features and cultural depth of the Chinese language pose challenges for instruction tuning tasks. Existing datasets are either derived from English-centric LLMs or are ill-suited for aligning with the interaction patterns of real-world Chinese users.
To bridge this gap, we introduce COIG-CQIA, a high-quality Chinese instruction tuning dataset.
Our aim is to build a diverse, wide-ranging instruction-tuning dataset to better align model behavior with human interactions.
To this end, we collect a high-quality human-written corpus from various sources on the Chinese Internet, including Q&A communities, Wikis, examinations, and existing NLP datasets.
This corpus was rigorously filtered and carefully processed to form the COIG-CQIA dataset.
Furthermore, we train models of various scales on different subsets of CQIA, following in-depth evaluation and analyses.
The findings from our experiments offer valuable insights for selecting and developing Chinese instruction-tuning datasets.
We also find that models trained on CQIA-Subset achieve competitive results in human assessment as well as knowledge and security benchmarks. Data are available athttps://huggingface.co/datasets/m-a-p/COIG-CQIA |
2404.05446v1 | Large Language Models (LLMs) have demonstrated remarkable performance across diverse tasks but are constrained by their small context window sizes.
Various efforts have been proposed to expand the context window to accommodate even up to 200K input tokens.
Meanwhile, building high-quality benchmarks with much longer text lengths and more demanding tasks to provide comprehensive evaluations is of immense practical interest to facilitate long context understanding research of LLMs.
However, prior benchmarks create datasets that ostensibly cater to long-text comprehension by expanding the input of traditional tasks, which falls short to exhibit the unique characteristics of long-text understanding, including long dependency tasks and longer text length compatible with modern LLMs’ context window size.
In this paper, we introduce a benchmark for eXtremelyLong context understanding withLong-range dependencies,XL2Bench, which includes three scenarios—Fiction Reading, Paper Reading, and Law Reading—and four tasks of increasing complexity: Memory Retrieval, Detailed Understanding, Overall Understanding, and Open-ended Generation, covering 27 subtasks in English and Chinese.
It has an average length of 100K+ words (English) and 200K+ characters (Chinese) . Evaluating six leading LLMs on XL2Bench, we find that their performance significantly lags behind human levels.
Moreover, the observed decline in performance across both the original and enhanced datasets underscores the efficacy of our approach to mitigating data contamination. |
2404.14367 | Learning from preference labels plays a crucial role in fine-tuning large language models. There are several distinct approaches for preference fine-tuning, including supervised learning, on-policy reinforcement learning (RL) , and contrastive learning.
Different methods come with different implementation tradeoffs and performance differences, and existing empirical findings present different conclusions, for instance, some results show that online RL is quite important to attain good fine-tuning results, while others find (offline) contrastive or even purely supervised methods sufficient. This raises a natural question:what kind of approaches are important for fine-tuning with preference data and why?In this paper, we answer this question by performing a rigorous analysis of a number of fine-tuning techniques on didactic and full-scale LLM problems.
Our main finding is that, in general, approaches that use on-policy sampling or attempt to push down the likelihood on certain responses (i.e., employ a “negative gradient”) outperform offline and maximum likelihood objectives. We conceptualize our insights and unify methods that use on-policy sampling or negative gradient under a notion of mode-seeking objectives for categorical distributions. Mode-seeking objectives are able to alter probability mass on specific bins of a categorical distribution at a fast rate compared to maximum likelihood, allowing them to relocate masses across bins more effectively. Our analysis prescribes actionable insights for preference fine-tuning of LLMs and informs how data should be collected for maximal improvement. |
2306.02928 | This paper presents a new approach to image similarity search in the context of fashion, a domain with inherent ambiguity due to the multiple ways in which images can be considered similar. We introduce the concept of Referred Visual Search (RVS) , where users provide additional information to define the desired similarity. We present a new dataset, LAION-RVS-Fashion, consisting of 272K fashion products with 842K images extracted from LAION, designed explicitly for this task. We then propose an innovative method for learning conditional embeddings using weakly-supervised training, achieving a 6% increase in Recall at one (R@1) against a gallery with 2M distractors, compared to classical approaches based on explicit attention and filtering. The proposed method demonstrates robustness, maintaining similar R@1 when dealing with 2.5 times as many distractors as the baseline methods. We believe this is a step forward in the emerging field of Referred Visual Search both in terms of accessible data and approach111Code, data and models are available athttps://github.com/Simon-Lepage/CondViT-LRVSF. |
2403.09611 | In this work, we discuss building performant Multimodal Large Language Models (MLLMs) . In particular, we study the importance of various architecture components and data choices. Through careful and comprehensive ablations of the image encoder, the vision language connector, and various pre-training data choices, we identified several crucial design lessons. For example, we demonstrate that for large-scale multimodal pre-training using a careful mix of image-caption, interleaved image-text, and text-only data is crucial for achieving state-of-the-art (SOTA) few-shot results across multiple benchmarks, compared to other published multimodal pre-training results. Further, we show that the image encoder together with image resolution and the image token count has substantial impact, while the vision-language connector design is of comparatively negligible importance. By scaling up the presented recipe, we buildMM1, a family of multimodal models, including both dense variants up to 30B and mixture-of-experts (MoE) variants up to 64B, that are SOTA in pre-training metrics and achieve competitive performance after supervised fine-tuning on a range of established multimodal benchmarks. Thanks to large-scale pre-training, MM1 enjoys appealing properties such as enhanced in-context learning, and multi-image reasoning, enabling few-shot chain-of-thought prompting. |
2404.15574 | Despite the recent progress in long-context large language models (LLMs) , it remains elusive how these transformer-based language models acquire the capability to retrieve relevant information from arbitrary locations within the long context. This paper aims to address this question.
Our systematic investigation across 4 model families, 6 model scales, and 3 types of finetuning reveals that a special type of attention heads are largely responsible for retrieving relevant information from long context, which we dubretrieval heads.
We identify important and intriguing properties of retrieval heads:
(1) universal:
all the explored models with long-context capability have a set of retrieval heads;
(2) sparse: only a small portion (less than 5%) of the attention heads are retrieval.
(3) intrinsic: retrieval heads already exist in
models pretrained with short context.
When extending the context length to 32-128K by continual pretraining,
it is still the same set of heads that perform information retrieval.
(4) dynamically activated:
take Llama-2 7B for example, 12
retrieval heads always attend to the required information no matter how the context is changed.
The rest of the retrieval heads are activated in different contexts.
(5) causal:
completely pruning retrieval heads leads to failure in retrieving relevant information and results in hallucination, while pruning random non-retrieval heads does not affect the model’s retrieval ability.
We further show that retrieval heads strongly influence
chain-of-thought (CoT) reasoning, where the model needs to frequently refer back the question and previously-generated context.
Conversely, tasks where the model directly generates the answer using its intrinsic knowledge
are less impacted by masking out retrieval heads.
These observations collectively explain which internal part of the model seeks information from the input tokens.
We believe our insights on retrieval heads foster future research on reducing hallucination, improving reasoning, and compressing the KV cache. |
2404.10642 | We explore the self-play training procedure of large language models (LLMs)
in a two-player adversarial language game calledAdversarial Taboo. In this game, an attacker and a defender communicate
with respect to a target word only visible to the attacker. The attacker aims to induce the defender to utter the target word unconsciously, while the defender tries to infer the target word from the attacker’s utterances. To win the game, both players should have sufficient knowledge about the target word and high-level reasoning ability to infer and express in this information-reserved conversation. Hence, we are curious about whether LLMs’ reasoning ability can be further enhanced bySelf-Play in thisAdversarial languageGame (SPAG) . With this goal, we let LLMs act as the attacker and play with a copy of itself as the defender on an extensive range of target words. Through reinforcement learning on the game outcomes, we observe that the LLMs’ performance uniformly improves on a broad range of reasoning benchmarks. Furthermore, iteratively adopting this self-play
process can continuously promote LLM’s reasoning ability. The code is athttps://github.com/Linear95/SPAG. |
2207.10551 | There have been a lot of interest in the scaling properties of Transformer models. However, not much has been done on the front of investigating the effect of scaling properties of different inductive biases and model architectures. Do model architectures scale differently? If so, how does inductive bias affect scaling behaviour? How does this influence upstream (pretraining) and downstream (transfer) ? This paper conducts a systematic study of scaling behaviour of ten diverse model architectures such as Transformers, Switch Transformers, Universal Transformers, Dynamic convolutions, Performers, and recently proposed MLP-Mixers. Via extensive experiments, we show that (1) architecture is an indeed an important consideration when performing scaling and (2) the best performing model can fluctuate at different scales. We believe that the findings outlined in this work has significant implications to how model architectures are currently evaluated in the community. |
2404.12272 | Due to the cumbersome nature of human evaluation and limitations of code-based evaluation, Large Language Models (LLMs) are increasingly being used to assist humans in evaluating LLM outputs. Yet LLM-generated evaluators simply inherit all the problems of the LLMs they evaluate, requiring further human validation. We present a mixed-initiative approach to "validate the validators"— aligning LLM-generated evaluation functions (be it prompts or code) with human requirements. Our interface, EvalGen, provides automated assistance to users in generating evaluation criteria and implementing assertions. While generating candidate implementations (Python functions, LLM grader prompts), EvalGen asks humans to grade a subset of LLM outputs; this feedback is used to select implementations that better align with user grades. A qualitative study finds overall support for EvalGen but underscores the subjectivity and iterative process of alignment. In particular, we identify a phenomenon we dub *criteria drift*: users need criteria to grade outputs, but grading outputs helps users define criteria. What is more, some criteria appears *dependent* on the specific LLM outputs observed (rather than independent criteria that can be defined *a priori*), raising serious questions for approaches that assume the independence of evaluation from observation of model outputs. We present our interface and implementation details, a comparison of our algorithm with a baseline approach, and implications for the design of future LLM evaluation assistants. |
2404.12358 | Reinforcement Learning From Human Feedback (RLHF) has been a critical to the success of the latest generation of generative AI models. In response to the complex nature of the classical RLHF pipeline, direct alignment algorithms such as Direct Preference Optimization (DPO) have emerged as an alternative approach. Although DPO solves the same objective as the standard RLHF setup, there is a mismatch between the two approaches. Standard RLHF deploys reinforcement learning in a specific token-level MDP, while DPO is derived as a bandit problem in which the whole response of the model is treated as a single arm. In this work we rectify this difference, first we theoretically show that we can derive DPO in the token-level MDP as a general inverse Q-learning algorithm, which satisfies the Bellman equation. Using our theoretical results, we provide three concrete empirical insights. First, we show that because of its token level interpretation, DPO is able to perform some type of credit assignment. Next, we prove that under the token level formulation, classical search-based algorithms, such as MCTS, which have recently been applied to the language generation space, are equivalent to likelihood-based search on a DPO policy. Empirically we show that a simple beam search yields meaningful improvement over the base DPO policy. Finally, we show how the choice of reference policy causes implicit rewards to decline during training. We conclude by discussing applications of our work, including information elicitation in multi-tun dialogue, reasoning, agentic applications and end-to-end training of multi-model systems. |
2303.08128 | Answering visual queries is a complex task that requires both visual processing and reasoning.
End-to-end models, the dominant approach for this task, do not explicitly differentiate between the two, limiting interpretability and generalization.
Learning modular programs presents a promising alternative, but has proven challenging due to the difficulty of learning both the programs and modules simultaneously.
We introduceViperGPT, a framework that leverages code-generation models to compose vision-and-language models into subroutines to produce a result for any query.ViperGPTutilizes a provided API to access the available modules, and
composes them by generating Python code that is later executed.
This simple approach requires no further training, and achieves state-of-the-art results across various complex visual tasks. |
2404.12096 | Embedding models play a pivot role in modern NLP applications such as IR and RAG.
While the context limit of LLMs has been pushed beyond 1 million tokens, embedding models are still confined to a narrow context window not exceeding 8k tokens, refrained from application scenarios requiring long inputs such as legal contracts.
This paper explores context window extension of existing embedding models, pushing the limit to 32k without requiring additional training.
First, we examine the performance of current embedding models for long context retrieval on our newly constructedLongEmbedbenchmark.LongEmbedcomprises two synthetic tasks and four carefully chosen real-world tasks, featuring documents of varying length and dispersed target information. Benchmarking results underscore huge room for improvement in these models.
Based on this, comprehensive experiments show that training-free context window extension strategies like position interpolation can effectively extend the context window of existing embedding models by several folds, regardless of their original context being 512 or beyond 4k.
Furthermore, for models employing absolute position encoding (APE) , we show the possibility of further fine-tuning to harvest notable performance gains while strictly preserving original behavior for short inputs. For models using rotary position embedding (RoPE) , significant enhancements are observed when employing RoPE-specific methods, such as NTK and SelfExtend, indicating RoPE’s superiority over APE for context window extension.
To facilitate future research, we release E5Base-4k and E5-RoPEBase, along with theLongEmbedbenchmark. |
2404.13950 | The late interaction paradigm introduced with ColBERT stands out in the neural Information Retrieval space, offering a compelling effectiveness-efficiency trade-off across many benchmarks. Efficient late interaction retrieval is based on an optimized multi-step strategy, where an approximate search first identifies a set of candidate documents to re-rank exactly. In this work, we introduce SPLATE, a simple and lightweight adaptation of the ColBERTv2 model which learns an “MLM adapter”, mapping itsfrozentoken embeddings to a sparse vocabulary space with a partially learned SPLADE module. This allows us to perform the candidate generation step in late interaction pipelines with traditional sparse retrieval techniques, making it particularly appealing for running ColBERT in CPU environments. Our SPLATE ColBERTv2 pipeline achieves the same effectiveness as the PLAID ColBERTv2 engine by re-ranking 50 documents that can be retrieved under 10. |
2404.00376 | While recent advancements in commercial large language models (LM) have shown promising results in medical tasks, their closed-source nature poses significant privacy and security concerns, hindering their widespread use in the medical field.
Despite efforts to create open-source models, their limited parameters often result in insufficient multi-step reasoning capabilities required for solving complex medical problems.
To address this, we introduce Meerkat-7B, a novel medical AI system with 7 billion parameters.
Meerkat-7B was trained using our new synthetic dataset consisting of high-quality chain-of-thought reasoning paths sourced from 18 medical textbooks, along with diverse instruction-following datasets.
Our system achieved remarkable accuracy across seven medical benchmarks, surpassing GPT-3.5 by 13.1%, as well as outperforming the previous best 7B models such as MediTron-7B and BioMistral-7B by 13.4% and 9.8%, respectively.
Notably, it surpassed the passing threshold of the United States Medical Licensing Examination (USMLE) for the first time for a 7B-parameter model.
Additionally, our system offered more detailed free-form responses to clinical queries compared to existing 7B and 13B models, approaching the performance level of GPT-3.5.
This significantly narrows the performance gap with large LMs, showcasing its effectiveness in addressing complex medical challenges. |
2404.08698 | While Large Language Models (LLMs) have shown remarkable abilities, they are hindered by significant resource consumption and considerable latency due to autoregressive processing. In this study, we introduce Adaptive N-gram Parallel Decoding (ANPD) , an innovative and lossless approach that accelerates inference by allowing the simultaneous generation of multiple tokens. ANPD incorporates a two-stage approach: it begins with a rapid drafting phase that employs an N-gram module, which adapts based on the current interactive context, followed by a verification phase, during which the original LLM assesses and confirms the proposed tokens. Consequently, ANPD preserves the integrity of the LLM’s original output while enhancing processing speed. We further leverage a multi-level architecture for the N-gram module to enhance the precision of the initial draft, consequently reducing inference latency. ANPD eliminates the need for retraining or extra GPU memory, making it an efficient and plug-and-play enhancement. In our experiments, models such as LLaMA and its fine-tuned variants have shown speed improvements up to, validating the effectiveness of our proposed ANPD. |
1811.03115 | Deep autoregressive sequence-to-sequence models have demonstrated impressive performance across a wide variety of tasks in recent years. While common architecture classes such as recurrent, convolutional, and self-attention networks make different trade-offs between the amount of computation needed per layer and the length of the critical path at training time, generation still remains an inherently sequential process. To overcome this limitation, we propose a novel blockwise parallel decoding scheme in which we make predictions for multiple time steps in parallel then back off to the longest prefix validated by a scoring model. This allows for substantial theoretical improvements in generation speed when applied to architectures that can process output sequences in parallel. We verify our approach empirically through a series of experiments using state-of-the-art self-attention models for machine translation and image super-resolution, achieving iteration reductions of up to 2x over a baseline greedy decoder with no loss in quality, or up to 7x in exchange for a slight decrease in performance. In terms of wall-clock time, our fastest models exhibit real-time speedups of up to 4x over standard greedy decoding. |
2307.02973 | Neural network pruning and quantization techniques are almost as old as neural networks themselves. However, to date only ad-hoc comparisons between the two have been published. In this paper, we set out to answer the question on which is better: neural network quantization or pruning? By answering this question, we hope to inform design decisions made on neural network hardware going forward.
We provide an extensive comparison between the two techniques for compressing deep neural networks.
First, we give an analytical comparison of expected quantization and pruning error for general data distributions.
Then, we provide lower bounds for the per-layer pruning and quantization error in trained networks, and compare these to empirical error after optimization.
Finally, we provide an extensive experimental comparison for training 8 large-scale models on 3 tasks.
Our results show that in most cases quantization outperforms pruning. Only in some scenarios with very high compression ratio, pruning might be beneficial from an accuracy standpoint. |
2308.01825 | Mathematical reasoning is a challenging task for large language models (LLMs) , while the scaling relationship of it with respect to LLM capacity is under-explored.
In this paper, we investigate how the pre-training loss, supervised data amount, and augmented data amount influence the reasoning performances of a supervised LLM.
We find that pre-training loss is a better indicator of the model’s performance than the model’s parameter count.
We apply supervised fine-tuning (SFT) with different amounts of supervised data and empirically find a log-linear relation between data amount and model performance, and we find better models improve less with enlarged supervised datasets.
To augment more data samples for improving model performances without any human effort, we propose to apply Rejection sampling Fine-Tuning (RFT) .
RFT uses supervised models to generate and collect correct reasoning paths as augmented fine-tuning datasets.
We find with augmented samples containing more distinct reasoning paths, RFT improves mathematical reasoning performance more for LLMs.
We also find RFT brings more improvement for less performant LLMs.
Furthermore, we combine rejection samples from multiple models which push LLaMA-7B to an accuracy of 49.3% on GSM8K which outperforms the supervised fine-tuning (SFT) accuracy of 35.9% significantly.
We release our codes and rejection sampling augmented data inhttps://github.com/OFA-Sys/gsm8k-ScRel. |
2309.06657 | Improving the alignment of language models with human preferences remains an active research challenge. Previous approaches have primarily utilized online Reinforcement Learning from Human Feedback (RLHF) . Recently, offline methods such as Sequence Likelihood Calibration (SLiC) and Direct Preference Optimization (DPO) have emerged as attractive alternatives, offering improvements in stability and scalability while maintaining competitive performance. SLiC refines its loss function using sequence pairs sampled from a supervised fine-tuned (SFT) policy, while DPO directly optimizes language models based on preference data, foregoing the need for a separate reward model. However, the maximum likelihood estimator (MLE) of the target optimal policy requires labeled preference pairs sampled from that policy. The absence of a reward model in DPO constrains its ability to sample preference pairs from the optimal policy. Meanwhile, SLiC can only sample preference pairs from the SFT policy. To address these limitations, we introduce a novel offline approach calledStatistical Rejection Sampling Optimization(RSO) designed to source preference data from the estimated target optimal policy using rejection sampling, enabling a more accurate estimation of the optimal policy. We also propose a unified framework that enhances the loss functions used in both SLiC and DPO from a preference modeling standpoint. Through extensive experiments across diverse tasks, we demonstrate that RSO consistently outperforms both SLiC and DPO as evaluated by gold reward, Large Language Models (LLMs) and human raters. |
2404.03608 | We present Sailor, a family of open language models ranging from 0.5B to 7B parameters, tailored for South-East Asian (SEA) languages.
These models are continually pre-trained from Qwen1.5, a great language model for multilingual use cases.
From Qwen1.5, Sailor models accept 200B to 400B tokens, primarily covering the languages of English, Chinese, Vietnamese, Thai, Indonesian, Malay, and Lao.
The training leverages several techniques, including BPE dropout for improving the model robustness, aggressive data cleaning and deduplication, and small proxy models to optimize data mixture.
Experimental results on four typical tasks indicate that Sailor models demonstrate strong performance across different benchmarks, including commonsense reasoning, question answering, reading comprehension and examination.
Embracing the open-source spirit, we share our insights through this report to spark a wider interest in developing large language models for multilingual use cases. |
2404.10719 | Reinforcement Learning from Human Feedback (RLHF) is currently the most widely used method to align large language models (LLMs) with human preferences.
Existing RLHF methods can be roughly categorized as eitherreward-basedorreward-free.
Novel applications such as ChatGPT and Claude leveragereward-basedmethods that first learn a reward model and apply actor-critic algorithms, such as Proximal Policy Optimization (PPO) .
However, in academic benchmarks, the state-of-the-art results are often achieved viareward-freemethods, such as Direct Preference Optimization (DPO) .Is DPO truly superior to PPO?Why does PPO perform poorly on these benchmarks?In this paper, we first conduct both theoretical and empirical studies on the algorithmic properties of DPO and show that DPO may have fundamental limitations.
Moreover, we also comprehensively examine PPO and reveal the key factors for the best performances of PPO in fine-tuning LLMs. Finally, we benchmark DPO and PPO across a collection of RLHF testbeds, ranging from dialogue to code generation. Experiment results demonstrate that PPO is able to surpass other alignment methods in all cases and achieve state-of-the-art results in challenging code competitions. |
2308.06394 | Instruction tuned Large Vision Language Models (LVLMs) have significantly advanced in generalizing across a diverse set of multi-modal tasks, especially for Visual Question Answering (VQA) . However, generating detailed responses that are visually grounded is still a challenging task for these models. We find that even the current state-of-the-art LVLMs (InstructBLIP) still contain a staggering 30 percent of the hallucinatory text in the form of non-existent objects, unfaithful descriptions, and inaccurate relationships. To address this, we introduceM-HalDetect, aMultimodalHallucinationDetection Dataset that can be used to train and benchmark models for hallucination detection and prevention. M-HalDetect consists of 16k fine-grained annotations on VQA examples, making it the first comprehensive multi-modal hallucination detection dataset for detailed image descriptions. Unlike previous work that only consider object hallucination, we additionally annotate both entity descriptions and relationships that are unfaithful. To demonstrate the potential of this dataset for hallucination prevention, we optimize InstructBLIP through our novel Fine-grained Direct Preference Optimization (FDPO) . We also train fine-grained multi-modal reward models from InstructBLIP and evaluate their effectiveness with best-of-n rejection sampling (RS) . We perform human evaluation on both FDPO and rejection sampling, and find that they reduce hallucination rates in InstructBLIP by 41% and 55% respectively. We also find that our reward model generalizes to other multi-modal models, reducing hallucinations in LLaVA and mPLUG-OWL by 15% and 57% respectively, and has strong correlation with human evaluated accuracy scores. The dataset is available athttps://github.com/hendryx-scale/mhal-detect. |
2312.15685 | Instruction tuning is a standard technique employed to align large language models to end tasks and user preferences after the initial pretraining phase.
Recent research indicates the critical role of data engineering in instruction tuning – when appropriately selected, only limited data is necessary to achieve superior performance.
However, we still lack a principled understanding of what makes good instruction tuning data for alignment, and how we should select data automatically and effectively.
In this work, we delve deeply into automatic data selection strategies for alignment.
We start with controlled studies to measure data across three dimensions: complexity, quality, and diversity, along which we examine existing methods and introduce novel techniques for enhanced data measurement.
Subsequently, we propose a simple strategy to select data samples based on the measurement.
We presentDeita(short forData-Efficient Instruction Tuning for Alignment) , a series of models fine-tuned from LLaMA and Mistral models using data samples automatically selected with our proposed approach.
Empirically,Deitaperforms better or on par with the state-of-the-art open-source alignment models with only 6K SFT training data samples – over 10x less than the data used in the baselines.
When further trained with direct preference optimization (DPO) ,Deita-Mistral-7B + DPO trained with 6K SFT and 10K DPO samples achieve 7.55 MT-Bench and 90.06% AlpacaEval scores.
We anticipate this work to provide tools on automatic data selection, facilitating data-efficient alignment.
We release our models as well as the selected datasets for future researches to effectively align models more efficiently.111Model checkpoints and data resources are available athttps://github.com/hkust-nlp/deita. |
2404.05961 | Large decoder-only language models (LLMs) are the state-of-the-art models on most of today’s NLP tasks and benchmarks. Yet, the community is only slowly adopting these models for text embedding tasks, which require rich contextualized representations. In this work, we introduce LLM2Vec, a simple unsupervised approach that can transform any decoder-only LLM into a strong text encoder. LLM2Vec consists of three simple steps: 1) enabling bidirectional attention, 2) masked next token prediction, and 3) unsupervised contrastive learning. We demonstrate the effectiveness of LLM2Vec by applying it to 3 popular LLMs ranging from 1.3B to 7B parameters and evaluate the transformed models on English word- and sequence-level tasks. We outperform encoder-only models by a large margin on word-level tasks and reach a new unsupervised state-of-the-art performance on the Massive Text Embeddings Benchmark (MTEB) . Moreover, when combining LLM2Vec with supervised contrastive learning, we achieve state-of-the-art performance on MTEB among models that train only on publicly available data. Our strong empirical results and extensive analysis demonstrate that LLMs can be effectively transformed into universal text encoders in a parameter-efficient manner without the need for expensive adaptation or synthetic GPT-4 generated data. |
2305.14259 | We explore and enhance the ability of neural language models to generate novel scientific directions grounded in literature. Work on literature-based hypothesis generation has traditionally focused on binary link prediction—severely limiting the expressivity of hypotheses. This line of work also does not focus on optimizing novelty.
We take a dramatic departure with a novel setting in which models use as input background contexts (e.g., problems, experimental settings, goals) , and outputnatural language ideasgrounded in literature. We presentSciMON, a modeling framework that uses retrieval of “inspirations” from past scientific papers, and explicitly optimizes for novelty by iteratively comparing to prior papers and updating idea suggestions until sufficient novelty is achieved.
Comprehensive evaluations reveal that GPT-4 tends to generate ideas with overall low technical depth and novelty, while our methods partially mitigate this issue. Our work represents a first step toward evaluating and developing language models that generate new ideas derived from the scientific literature.111Code, data, and resources are publicly available for research purposes:https://github.com/eaglew/clbd. |
2312.08583 | This study examines 4-bit quantization methods like GPTQ in large language models (LLMs) , highlighting GPTQ’s overfitting and limited enhancement in Zero-Shot tasks. While prior works merely focusing on zero-shot measurement, we extend task scope to more generative categories such as code generation and abstractive summarization, in which we found that INT4 quantization can significantly underperform. However, simply shifting to higher precision formats like FP6 has been particularly challenging, thus overlooked, due to poor performance caused by the lack of sophisticated integration and system acceleration strategies on current AI hardware. Our results show that FP6, even with a coarse-grain quantization scheme, performs robustly across various algorithms and tasks, demonstrating its superiority in accuracy and versatility. Notably, with the FP6 quantization, StarCoder-15B model performs comparably to its FP16 counterpart in code generation, and for smaller models like the 406M it closely matches their baselines in summarization. Neither can be achieved by INT4. To better accommodate various AI hardware and achieve the best system performance, we propose a novel 4+2 design for FP6 to achieve similar latency to the state-of-the-art INT4 fine-grain quantization. With our design, FP6 can become a promising solution to the current 4-bit quantization methods used in LLMs.111Code will be released soon as a part ofhttps://github.com/microsoft/DeepSpeed |
2404.09173 | While Transformers have revolutionized deep learning, their quadratic attention complexity hinders their ability to process infinitely long inputs. We propose Feedback Attention Memory (FAM) , a novel Transformer architecture that leverages a feedback loop to enable the network to attend to its own latent representations. This design fosters the emergence of working memory within the Transformer, allowing it to process indefinitely long sequences. TransformerFAM requires no additional weights, enabling seamless integration with pre-trained models. Our experiments show that TransformerFAM significantly improves Transformer performance on long-context tasks across various model sizes (1B, 8B, and 24B) . These results showcase the potential to empower Large Language Models (LLMs) to process sequences of unlimited length. |
2404.08801 | The quadratic complexity and weak length extrapolation of Transformers limits their ability to scale to long sequences, and while sub-quadratic solutions like linear attention and state space models exist, they empirically underperform Transformers in pretraining efficiency and downstream task accuracy.
We introduceMegalodon, an neural architecture for efficient sequence modeling with unlimited context length.Megalodoninherits the architecture ofMega(exponential moving average with gated attention) , and further introduces multiple technical components to improve its capability and stability, includingcomplex exponential moving average (CEMA) ,timestep normalizationlayer,normalized attentionmechanism andpre-norm with two-hop residualconfiguration.
In a controlled head-to-head comparison withLlama2,Megalodonachieves better efficiency than Transformer in the scale of 7 billion parameters and 2 trillion training tokens.Megalodonreaches a training loss of 1.70, landing mid-way betweenLlama2-7B (1.75) and 13B (1.67) .
The improvements ofMegalodonover Transformers are robust throughout a range of benchmarks across different tasks and modalities.Code:https://github.com/XuezheMax/megalodon |
2302.07253 | Our work combines aspects of three promising paradigms in
machine learning, namely, attention mechanism, energy-based models, and
associative memory. Attention is the power-house driving modern deep
learning successes, but it lacks clear theoretical foundations. Energy-based
models allow a principled approach to discriminative and generative
tasks, but the design of the energy functional is not straightforward. At
the same time, Dense Associative Memory models or Modern Hopfield Networks
have a well-established theoretical foundation, and allow an intuitive
design of the energy function. We propose a novel architecture, called the
Energy Transformer (or ET for short) , that uses a sequence of attention
layers that are purposely designed to minimize a specifically engineered
energy function, which is responsible for representing the relationships
between the tokens. In this work, we introduce the theoretical foundations
of ET, explore its empirical capabilities using the image completion task,
and obtain strong quantitative results on the graph anomaly detection and
graph classification tasks. |
2403.20327 | We present Gecko, a compact and versatile text embedding model.
Gecko achieves strong retrieval performance by leveraging a key idea: distilling knowledge from large language models (LLMs) into a retriever.
Our two-step distillation process begins with generating diverse, synthetic paired data using an LLM.
Next, we further refine the data quality by retrieving a set of candidate passages for each query, and relabeling the positive and hard negative passages using the same LLM. The effectiveness of our approach is demonstrated by the compactness of the Gecko. On the Massive Text Embedding Benchmark (MTEB) , Gecko with 256 embedding dimensions outperforms all existing entries with 768 embedding size. Gecko with 768 embedding dimensions achieves an average score of 66.31, competing with 7x larger models and 5x higher dimensional embeddings. |
2402.01035 | Tokenization is an understudied and often neglected component of modern LLMs.
Most published works use a single tokenizer for all experiments, often borrowed from another model, without performing ablations or analysis to optimize tokenization.
Moreover, the tokenizer is generally kept unchanged when fine-tuning a base model.
In this paper, we show that the size, pre-tokenization regular expression, and training data of a tokenizer can significantly impact the model’s generation speed, effective context size, memory usage, and downstream performance.
We train specialized Byte-Pair Encoding code tokenizers, and conduct extensive ablations on the impact of tokenizer design on the performance of LLMs for code generation tasks such as HumanEval and MBPP, and provide recommendations for tokenizer hyper-parameters selection and switching the tokenizer in a pre-trained LLM.
We perform our experiments on models trained from scratch and from pre-trained models, verifying their applicability to a wide range of use-cases.
We find that when fine-tuning on more than 50 billion tokens, we can specialize the tokenizer of a pre-trained LLM to obtain large gains in generation speed and effective context size. |
2404.05728 | Large neural network models have become a mainstay of natural language processing and computer vision, yet their initialization and learning rates are set in a largely heuristic fashion, potentially varying from paper to paper and one model size to the next.
The-Parameterization (P) offers a potential solution to these challenges, yielding scaling rules for model initialization and learning rates, and reportedly enabling zero-shot hyperparameter transfer from small to large models in a variety of cases. Despite the evident promise, theP scaling rules are not yet widely adopted, perhaps due to higher implementation complexity, many variations, or complex theoretical background.
This work investigatesP empirically, focusing on the ubiquitous transformer architecture, and aims to answer a simple question: does-Transfer yield optimal learning rates in practice?
Studying models with up to 10B parameters and training budgets of up to 190B tokens, we find-Transfer works as intended for the majority of important cases, yet also identify a few cases where it may not. |
2403.19588 | This paper revives Densely Connected Convolutional Networks (DenseNets) and reveals the underrated effectiveness over predominant ResNet-style architectures.
We believe DenseNets’ potential was overlooked due to untouched training methods and traditional design elements not fully revealing their capabilities. Our pilot study shows dense connections through concatenation are strong, demonstrating that DenseNets can be revitalized to compete with modern architectures. We methodically refine suboptimal components - architectural adjustments, block redesign, and improved training recipes towards widening DenseNets and boosting memory efficiency while keeping concatenation shortcuts. Our models, employing simple architectural elements, ultimately surpass Swin Transformer, ConvNeXt, and DeiT-III — key architectures in the residual learning lineage. Furthermore, our models exhibit near state-of-the-art performance on ImageNet-1K, competing with the very recent models and downstream tasks, ADE20k semantic segmentation, and COCO object detection/instance segmentation. Finally, we provide empirical analyses that uncover the merits of the concatenation over additive shortcuts, steering a renewed preference towards DenseNet-style designs. Our code is available athttps://github.com/naver-ai/rdnet. |
2209.14958 | Language models are increasingly attracting interest from writers. However, such models lack long-range semantic coherence, limiting their usefulness for longform creative writing. We address this limitation by applying language models hierarchically, in a system we callDramatron. By building structural context via prompt chaining, Dramatron can generate coherent scripts and screenplays complete with title, characters, story beats, location descriptions, and dialogue. We illustrate Dramatron’s usefulness as an interactive co-creative system with a user study oftheatre and film industry professionals. Participants co-wrote theatre scripts and screenplays with Dramatron and engaged in open-ended interviews. We report critical reflections both from our interviewees and from independent reviewers who watched stagings of the works to illustrate how both Dramatron and hierarchical text generation could be useful for human-machine co-creativity. Finally, we discuss the suitability of Dramatron for co-creativity, ethical considerations—including plagiarism and bias—and participatory models for the design and deployment of such tools. |
2404.07647 | Recent advances in language modeling consist in pretraining highly parameterized neural networks on extremely large web-mined text corpora. Training and inference with such models can be costly in practice, which incentivizes the use of smaller counterparts. However, it has been observed that smaller models can suffer from saturation, characterized as a drop in performance at some advanced point in training followed by a plateau. In this paper, we find that such saturation can be explained by a mismatch between the hidden dimension of smaller models and the high rank of the target contextual probability distribution. This mismatch affects the performance of the linear prediction head used in such models through the well-known softmax bottleneck phenomenon. We measure the effect of the softmax bottleneck in various settings and find that models based on less than 1000 hidden dimensions tend to adopt degenerate latent representations in late pretraining, which leads to reduced evaluation performance. |
2402.13228 | Direct Preference Optimisation (DPO) is effective at significantly improving the performance of large language models (LLMs) on downstream tasks such as reasoning, summarisation, and alignment. Using pairs of preferred and dispreferred data, DPO models therelativeprobability of picking one response over another. In this work, first we show theoretically that the standard DPO loss can lead to areductionof the model’s likelihood of the preferred examples, as long as the relative probability between the preferred and dispreferred classes increases. We then show empirically that this phenomenon occurs when fine-tuning LLMs on common datasets, especially datasets in which the edit distance between pairs of completions is low. Using these insights, we design DPO-Positive (DPOP) , a new loss function and training procedure which avoids this failure mode. Surprisingly, we also find that DPOP significantly outperforms DPO across a wide variety of datasets and downstream tasks, including datasets with high edit distances between completions. By fine-tuning with DPOP, we create and release Smaug-34B and Smaug-72B, which achieve state-of-the-art open-source performance. Notably, Smaug-72B is nearly 2% better than any other open-source model on the HuggingFace Open LLM Leaderboard and becomes the first open-source LLM to surpass an average accuracy of 80%. |
2404.07965 | Previous language model pre-training methods have uniformly applied a next-token prediction loss to all training tokens.
Challenging this norm, we posit that“Not all tokens in a corpus are equally important for language model training”.
Our initial analysis delves into token-level training dynamics of language model, revealing distinct loss patterns for different tokens.
Leveraging these insights, we introduce a new language model calledRho-1. Unlike traditional LMs that learn to predict every next token in a corpus,Rho-1employs Selective Language Modeling (SLM) , which selectively trains on useful tokens that aligned with the desired distribution.
This approach involves scoring pretraining tokens using a reference model, and then training the language model with a focused loss on tokens with higher excess loss.
When continual pretraining on 15B OpenWebMath corpus,Rho-1yields an absolute improvement in few-shot accuracy of up to 30% in 9 math tasks.
After fine-tuning,Rho-1-1B and 7B achieved state-of-the-art results of 40.6% and 51.8% on MATH dataset, respectively — matching DeepSeekMath with only 3% of the pretraining tokens.
Furthermore, when pretraining on 80B general tokens,Rho-1achieves 6.8% average enhancement across 15 diverse tasks, increasing both efficiency and performance of the language model pre-training. |
2404.04656 | Aligning Large Language Models (LLMs) to human preferences through preference optimization has been crucial but labor-intensive, necessitating for each prompt a comparison of both a chosen and a rejected text completion by evaluators.
Recently, Kahneman-Tversky Optimization (KTO) has demonstrated that LLMs can be aligned using merely binary ”thumbs-up” or ”thumbs-down” signals on each prompt-completion pair.
In this paper, we present theoretical foundations to explain the successful alignment achieved through these binary signals.
Our analysis uncovers a new perspective: optimizing a binary classifier, whose logit is a reward, implicitly induces minimizing the Direct Preference Optimization (DPO) loss.
In the process of this discovery, we identified two techniques for effective alignment: reward shift and underlying distribution matching.
Consequently, we propose a new algorithm,Binary Classifier Optimization, that integrates the techniques.
We validate our methodology in two settings: first, on a paired preference dataset, where our method performs on par with DPO and KTO; and second, on binary signal datasets simulating real-world conditions with divergent underlying distributions between thumbs-up and thumbs-down data.
Our model consistently demonstrates effective and robust alignment across two base LLMs and three different binary signal datasets, showcasing the strength of our approach to learning from binary feedback. |
2403.15246 | Modern Large Language Models (LLMs) are capable of following long and complex instructions that enable a diverse amount of user tasks.
However, despite Information Retrieval (IR) models using LLMs as the backbone of their architectures, nearly all of them still only take queries as input, with no instructions.
For the handful of recent models that do take instructions, it’s unclear how they use them.
We introduce our datasetFollowIR, which contains a rigorous instruction evaluation benchmark as well as a training set for helping IR models learn to better follow real-world instructions.FollowIRbuilds off the long history of the TREC conferences: as TREC provides human annotators with instructions (also known asnarratives) to determine document relevance, so should IR models be able to understand and decide relevance based on these detailed instructions.
Our evaluation benchmark starts with three deeply judged TREC collections and alters the annotator instructions, re-annotating relevant documents.
Through this process, we can measure how well IR models follow instructions, through a new pairwise evaluation framework.
Our results indicate that existing retrieval models fail to correctly use instructions, using them for basic keywords and struggling to understand long-form information.
However, we show that it is possible for IR models to learn to follow complex instructions: our newFollowIR-7B model has significant improvements (over 13%) after fine-tuning on our training set. |
2404.05405v1 | Scaling laws describe the relationship between the size of language models and their capabilities. Unlike prior studies that evaluate a model’s capability via loss or benchmarks, we estimate the number of knowledgebitsa model stores. We focus on factual knowledge represented as tuples, such as (USA, capital, Washington D.C.) from a Wikipedia page. Through multiple controlled datasets, we establish that language models can and only can store2 bits of knowledge per parameter, even when quantized to int8, and such knowledge can be flexibly extracted for downstream applications. Consequently, a 7B model can store 14B bits of knowledge, surpassing the English Wikipedia and textbooks combined based on our estimation. More broadly, we present 12 resultson how (1) training duration, (2) model architecture, (3) quantization, (4) sparsity constraints such as MoE, and (5) data signal-to-noise ratio affect a model’s knowledge storage capacity. Notable insights include: The GPT-2 architecture, with rotary embedding, matches or even surpasses LLaMA/Mistral architecturesin knowledge storage, particularly over shorter training durations. This arises because LLaMA/Mistral uses GatedMLP, which is less stable and harder to train. Prepending training data with domain names (e.g., wikipedia.org) significantly increases a model’s knowledge capacity. Language models can autonomously identify and prioritize domains rich in knowledge, optimizing their storage capacity. |
2404.02948 | As the parameters of large language models (LLMs) expand, the computational cost of fine-tuning the entire model becomes prohibitive. To address this challenge, we introduce a parameter-efficient fine-tuning (PEFT) method,PrincipalSingular values andSingular vectorsAdaptation (PiSSA) , which optimizes a significantly reduced parameter space while achieving or surpassing the performance of full-parameter fine-tuning. PiSSA is inspired by Intrinsic SAID, which suggests that pre-trained, over-parametrized models inhabit a space of low intrinsic dimension. Consequently, PiSSA represents a matrixwithin the model by the product of two trainable matricesand, where, plus a residual matrixfor error correction. Singular value decomposition (SVD) is employed to factorize, and the principal singular values and vectors ofare utilized to initializeand. The residual singular values and vectors initialize the residual matrix, which keeps frozen during fine-tuning. Notably, PiSSA shares the same architecture with Low-Rank Adaptation (LoRA) , which hypothesizes that changes in model parametersform a low-rank matrix. However, LoRA approximatesthrough the product of two matrices,, initialized with Gaussian noise, and, initialized with zeros, while PiSSA initializesandwith principal singular values and singular vectors of the original matrix. Given that the principal singular values and vectors capture the essence of a low-rank matrix, PiSSA can better approximate the outcomes of full-parameter fine-tuning at the beginning by changing the essential parts while freezing the “noisy” parts. In comparison, LoRA freezes the original matrix and updates the “noise”. This distinction enables PiSSA to convergence much faster than LoRA and also achieve better performance in the end. On five common benchmarks, PiSSA outperforms LoRA on all of them using exactly the same setups except for a different initialization. On GSM8K, Mistral-7B fine-tuned with PiSSA achieves an accuracy of 72.86%, outperforming LoRA’s 67.7% by 5.16%.
Due to the same architecture, PiSSA inherits many of LoRA’s advantages, such as parameter efficiency and compatibility with quantization.
Furthermore, PiSSA reduces the 4-bit quantization error in LLaMA 2-7B by 18.97%, resulting in a substantial improvement in fine-tuning performance. On the GSM8K benchmark, PiSSA achieves an accuracy of 49.13%, surpassing the performances of QLoRA at 39.8% and LoftQ at 40.71%.
Leveraging a fast SVD technique, the initialization of PiSSA takes only a few seconds, inducing negligible cost of switching LoRA to PiSSA. |
2402.11131 | Speculative decoding is a prominent technique to speed up the inference of a large target language model based on predictions of an auxiliary draft model. While effective, in application-specific settings, it often involves fine-tuning both draft and target models to achieve high acceptance rates. As the number of downstream tasks grows, these draft models add significant complexity to inference systems. We propose Speculative Streaming, a single-model speculative decoding method that fuses drafting into the target model by changing the fine-tuning objective from next token prediction to future n-gram prediction. Speculative Streaming speeds up decoding by 1.8 - 3.1X in a diverse set of tasks, such as Summarization, Structured Queries, and Meaning Representation, without sacrificing generation quality. Additionally, Speculative Streaming is parameter-efficient. It achieves on-par/higher speed-ups than Medusa-style architectures while using10000X fewer extra parameters, making it well-suited for resource-constrained devices. |
2209.14958v1 | Language models are increasingly attracting interest from writers. However, such models lack long-range semantic coherence, limiting their usefulness for longform creative writing. We address this limitation by applying language models hierarchically, in a system we callDramatron. By building structural context via prompt chaining, Dramatron can generate coherent scripts and screenplays complete with title, characters, story beats, location descriptions, and dialogue. We illustrate Dramatron’s usefulness as an interactive co-creative system with a user study oftheatre and film industry professionals. Participants co-wrote theatre scripts and screenplays with Dramatron and engaged in open-ended interviews. We report critical reflections both from our interviewees and from independent reviewers who watched stagings of the works to illustrate how both Dramatron and hierarchical text generation could be useful for human-machine co-creativity. Finally, we discuss the suitability of Dramatron for co-creativity, ethical considerations—including plagiarism and bias—and participatory models for the design and deployment of such tools. |
2403.17844 | The development of deep learning architectures is a resource-demanding process, due to a vast design space, long prototyping times, and high compute costs associated with at-scale model training and evaluation. We set out to simplify this process by grounding it in an end-to-endmechanistic architecture design(MAD) pipeline, encompassing small-scale capability unit tests predictive of scaling laws. Through a suite of synthetic token manipulation tasks such as compression and recall, designed to probe capabilities, we identify and test new hybrid architectures constructed from a variety of computational primitives. We experimentally validate the resulting architectures via an extensive compute-optimal and a new state-optimal scaling law analysis, training overlanguage models betweentoparameters. Surprisingly, we findMADsynthetics to correlate with compute-optimal perplexity, enabling accurate evaluation of new architectures via isolated proxy tasks. The new architectures found viaMAD, based on simple ideas such as hybridization and sparsity, outperform state-of-the-art Transformer, convolutional, and recurrent architectures (Transformer++, Hyena, Mamba) in scaling, both at compute-optimal budgets and in overtrained regimes. Overall, these results provide evidence that performance on curated synthetic tasks can be predictive of scaling laws, and that an optimal architecture should leverage specialized layers via a hybrid topology. |
2404.00498 | CIFAR-10 is among the most widely used datasets in machine learning, facilitating thousands of research projects per year.
To accelerate research and reduce the cost of experiments, we introduce training methods for CIFAR-10 which reach 94% accuracy in 3.29 seconds, 95% in 10.4 seconds, and 96% in 46.3 seconds, when run on a single NVIDIA A100 GPU.
As one factor contributing to these training speeds, we propose a derandomized variant of horizontal flipping augmentation, which we show improves over the standard method in every case where flipping is beneficial over no flipping at all.
Our code is released athttps://github.com/KellerJordan/cifar10-airbench. |
2404.02805 | Dense retrieval techniques employ pre-trained large language models to build a high-dimensional representation of queries and passages. These representations compute the relevance of a passage w.r.t. to a query using efficient similarity measures. In this line, multi-vector representations show improved effectiveness at the expense of a one-order-of-magnitude increase in memory footprint and query latency by encoding queries and documents on a per-token level.
Recently, PLAID has tackled these problems by introducing a centroid-based term representation to reduce the memory impact of multi-vector systems. By exploiting a centroid interaction mechanism, PLAID filters out non-relevant documents, thus reducing the cost of the successive ranking stages.
This paper proposes “Efficient Multi-Vector dense retrieval with Bit vectors” (EMVB) , a novel framework for efficient query processing in multi-vector dense retrieval.
First, EMVB employs a highly efficient pre-filtering step of passages using optimized bit vectors. Second, the computation of the centroid interaction happens column-wise, exploiting SIMD instructions, thus reducing its latency. Third, EMVB leverages Product Quantization (PQ) to reduce the memory footprint of storing vector representations while jointly allowing for fast late interaction. Fourth, we introduce a per-document term filtering method that further improves the efficiency of the last step.
Experiments on MS MARCO and LoTTE show that EMVB is up tofaster while reducing the memory footprint bywith no loss in retrieval accuracy compared to PLAID. |
2402.05120 | We find that, simply via a sampling-and-voting method, the performance of large language models (LLMs) scales with the number of agents instantiated. Also, this method is orthogonal to existing complicated methods to further enhance LLMs, while the degree of enhancement is correlated to the task difficulty. We conduct comprehensive experiments on a wide range of LLM benchmarks to verify the presence of our finding, and to study the properties that can facilitate its occurrence. Our code is publicly available at:Git. |
2403.06963 | Can a mere next-token predictor faithfully model human intelligence? We crystallize this intuitive concern, which is fragmented in the literature. As a starting point, we argue that the two often-conflated phases of next-token prediction — autoregressive inference and teacher-forced training — must be treated distinctly.
The popular criticism that errors can compound during autoregressive inference, crucially assumes that teacher-forcing has learned an accurate next-token predictor. This assumption sidesteps a more deep-rooted problem we expose: in certain classes of tasks, teacher-forcing can simply fail to learn an accurate next-token predictor in the first place. We describe a general mechanism of how teacher-forcing can fail, and design a minimal planning task where both the Transformer and the Mamba architecture empirically fail in that manner — remarkably, despite the task being straightforward to learn.
We provide preliminary evidence that this failure can be resolved when training to predict multiple tokens in advance.
We hope this finding can ground future debates and inspire explorations beyond the next-token prediction paradigm. We make our code available underhttps://github.com/gregorbachmann/Next-Token-Failures |
2404.03626 | In this paper, we explore the idea of training large language models (LLMs) over highly compressed text. While standard subword tokenizers compress text by a small factor, neural text compressors can achieve much higher rates of compression. If it were possible to train LLMs directly over neurally compressed text, this would confer advantages in training and serving efficiency, as well as easier handling of long text spans. The main obstacle to this goal is that strong compression tends to produce opaque outputs that are not well-suited for learning. In particular, we find that text naïvely compressed via Arithmetic Coding is not readily learnable by LLMs. To overcome this, we propose Equal-Info Windows, a novel compression technique whereby text is segmented into blocks that each compress to the same bit length. Using this method, we demonstrate effective learning over neurally compressed text that improves with scale, and outperforms byte-level baselines by a wide margin on perplexity and inference speed benchmarks. While our method delivers worse perplexity than subword tokenizers for models trained with the same parameter count,
it has the benefit of shorter sequence lengths.
Shorter sequence lengths require fewer autoregressive generation steps,
and reduce latency.
Finally, we provide extensive analysis of the properties that contribute to learnability, and offer concrete suggestions for how to further improve the performance of high-compression tokenizers. |
2404.02258 | Transformer-based language models spread FLOPs uniformly across input sequences. In this work we demonstrate that transformers can instead learn to dynamically allocate FLOPs (orcompute) to specific positions in a sequence, optimising the allocation along the sequence for different layers across the model depth. Our method enforces a total compute budget by capping the number of tokens () that can participate in the self-attention and MLP computations at a given layer. The tokens to be processed are determined by the network using a top-routing mechanism. Sinceis defineda priori, this simple procedure uses a static computation graph with known tensor sizes, unlike other conditional computation techniques. Nevertheless, since the identities of thetokens are fluid, this method can expend FLOPs non-uniformly across the time and model depth dimensions. Thus, compute expenditure is entirely predictable in sum total, but dynamic and context-sensitive at the token-level. Not only do models trained in this way learn to dynamically allocate compute, they do so efficiently. These models match baseline performance for equivalent FLOPS and wall-clock times to train, but require a fraction of the FLOPs per forward pass, and can be upwards of 50% faster to step during post-training sampling. |
2310.07096 | The Universal Transformer (UT) is a variant of the Transformer that shares parameters across its layers.
Empirical evidence shows that UTs have better compositional generalization than Vanilla Transformers (VTs) in formal language tasks.
The parameter-sharing also affords it better parameter efficiency than VTs.
Despite its many advantages, scaling UT parameters is much more compute and memory intensive than scaling up a VT.
This paper proposes the Sparse Universal Transformer (SUT) , which leverages Sparse Mixture of Experts (SMoE) and a new stick-breaking-based dynamic halting mechanism to reduce UT’s computation complexity while retaining its parameter efficiency and generalization ability.
Experiments show that SUT achieves the same performance as strong baseline models while only using half computation and parameters on WMT’14 and strong generalization results on formal language tasks (Logical inference and CFQ) .
The new halting mechanism also enables around 50% reduction in computation during inference with very little performance decrease on formal language tasks. |
2310.03294 | Increasing the context length of large language models (LLMs) unlocks fundamentally new capabilities, but also significantly increases the memory footprints of training.
Previous model-parallel systems such as Megatron-LM partition and compute different attention heads in parallel, resulting in large communication volumes,
so they cannot scale beyond the number of attention heads, thereby hindering its adoption.
In this paper, we introduce a new approach,LightSeq, for long-context LLMs training.LightSeqhas many notable advantages. First,LightSeqpartitions over the sequence dimension, hence is agnostic to model architectures and readily applicable for models with varying numbers of attention heads, such as Multi-Head, Multi-Query and Grouped-Query attention. Second,LightSeqnot only requires up to 4.7less communication than Megatron-LM on popular LLMs but also overlaps the communication with computation. To further reduce the training time,LightSeqfeatures a novel gradient checkpointing scheme to bypass an forward computation for memory-efficient attention.
We evaluateLightSeqon Llama-7B and its variants with sequence lengths from 32K to 512K. Through comprehensive experiments on single and cross-node training, we show thatLightSeqachieves up to 1.24-2.01end-to-end speedup, and a 2-8longer sequence length on models with fewer heads, compared to Megatron-LM. Codes will be available athttps://github.com/RulinShao/LightSeq. |
2404.00456 | We introduce QuaRot, a newQuantization scheme based onRotations, which is able to quantize LLMs end-to-end, including all weights, activations, and KV cache in 4 bits. QuaRot rotates LLMs in a way that removes outliers from the hidden state without changing the output, making quantization easier. Thiscomputational invarianceis applied to the hidden state (residual) of the LLM, as well as to the activations of the feed-forward components, aspects of the attention mechanism and to the KV cache. The result is a quantized model where all matrix multiplications are performed in 4-bits, without any channels identified for retention in higher precision. Our quantizedLlama2-70Bmodel has losses of at most 0.29 WikiText-2 perplexity and retains 99% of the zero-shot performance. Code is available at:https://github.com/spcl/QuaRot. |
2403.19887 | We present Jamba, a new base large language model based on a novel hybrid Transformer-Mamba mixture-of-experts (MoE) architecture. Specifically, Jamba interleaves blocks of Transformer and Mamba layers, enjoying the benefits of both model families. MoE is added in some of these layers to increase model capacity while keeping active parameter usage manageable. This flexible architecture allows resource- and objective-specific configurations.
In the particular configuration we have implemented, we end up with a powerful model that fits in a single 80GB GPU.
Built at large scale, Jamba provides high throughput and small memory footprint compared to vanilla Transformers, and at the same time state-of-the-art performance on standard language model benchmarks and long-context evaluations. Remarkably, the model presents strong results for up to 256K tokens context length.
We study various architectural decisions, such as how to combine Transformer and Mamba layers, and how to mix experts, and show that some of them are crucial in large scale modeling. We also describe several interesting properties of these architectures which the training and evaluation of Jamba have revealed, and plan to release checkpoints from various ablation runs, to encourage further exploration of this novel architecture.
We make the weights of our implementation of Jamba publicly available under a permissive license.Model:https://huggingface.co/ai21labs/Jamba-v0.1 |
2403.19888 | Recent advances in deep learning have mainly relied on Transformers due to their data dependency and ability to learn at scale. The attention module in these architectures, however, exhibit quadratic time and space in input size, limiting their scalability for long-sequence modeling. Despite recent attempts to design efficient and effective architecture backbone for multi-dimensional data, such as images and multivariate time series, existing models are either data independent, or fail to allow inter- and intra-dimension communication. Recently, State Space Models (SSMs) , and more specifically Selective State Space Models (S6) , with efficient hardware-aware implementation, have shown promising potential for long sequence modeling. Motivated by the recent success of SSMs, we present MambaMixer block, a new architecture with data dependent weights that uses a dual selection mechanism across tokens and channels–called Selective Token and Channel Mixer. MambaMixer further connects the sequential selective mixers using a weighted averaging mechanism, allowing layers to have direct access to different layers’ input and output. As a proof of concept, we design Vision MambaMixer (ViM2) and Time Series MambaMixer (TSM2) architectures based on MambaMixer block and explore their performance in various vision and time series forecasting tasks. Our results underline the importance of selectively mixing across both tokens and channels. In ImageNet classification, object detection, and semantic segmentation tasks, ViM2 achieves competitive performance with well-established vision models, i.e., ViT, MLP-Mixer, ConvMixer, and outperforms SSM-based vision models, i.e., ViM and VMamba. In time series forecasting, TSM2, an attention and MLP-free architecture, achieves outstanding performance compared to state-of-the-art methods while demonstrating significantly improved computational cost. These results show that while Transformers, cross-channel attention, and cross-channel MLPs are sufficient for good performance in practice, neither is necessary. |
2403.17297 | The evolution of Large Language Models (LLMs) like ChatGPT and GPT-4 has sparked discussions on the advent of Artificial General Intelligence (AGI) . However, replicating such advancements in open-source models has been challenging. This paper introduces InternLM2, an open-source LLM that outperforms its predecessors in comprehensive evaluations across 6 dimensions and 30 benchmarks, long-context modeling, and open-ended subjective evaluations through innovative pre-training and optimization techniques. The pre-training process of InternLM2 is meticulously detailed, highlighting the preparation of diverse data types including text, code, and long-context data. InternLM2 efficiently captures long-term dependencies, initially trained on 4k tokens before advancing to 32k tokens in pre-training and fine-tuning stages, exhibiting remarkable performance on the 200k “Needle-in-a-Haystack” test. InternLM2 is further aligned using Supervised Fine-Tuning (SFT) and a novel Conditional Online Reinforcement Learning from Human Feedback (COOL RLHF) strategy that addresses conflicting human preferences and reward hacking. By releasing InternLM2 models in different training stages and model sizes, we provide the community with insights into the model’s evolution. |
2103.13262 | Mixture-of-Expert (MoE) presents a strong potential in enlarging the size of language model to trillions of parameters.
However, training trillion-scale MoE requires
algorithm and system co-design for a well-tuned high performance distributed training system.
Unfortunately,
the only existing platform that meets the requirements
strongly depends on Google’s hardware (TPU) and software (Mesh Tensorflow) stack, and is
not open and available to the public, especially GPU and PyTorch communities. In this paper, we presentFastMoE, a distributed MoE training system based on PyTorch with common accelerators.
The system provides a hierarchical interface for both flexible model design and easy adaption to different applications, such as Transformer-XL and Megatron-LM.
Different from direct implementation of MoE models using PyTorch, the training speed is highly optimized inFastMoEby sophisticated high-performance acceleration skills.
The system supports placing different experts on multiple GPUs across multiple nodes,
enabling enlarging the number of experts linearly against the number of GPUs.
The source ofFastMoEis available athttps://github.com/laekov/fastmoeunder Apache-2 license. |
2202.09368 | Sparsely-activated Mixture-of-experts (MoE) models allow the number of parameters to greatly increase while keeping the amount of computation for a given token or a given sample unchanged. However, a poor expert routing strategy can cause certain experts to be under-trained, leading to an expert being under or over-specialized. Prior work allocates a fixed number of experts to each token using a top-function regardless of the relative importance of different tokens. To address this, we propose a heterogeneous mixture-of-experts employing an expert choice method. Instead of letting tokens select the top-experts, we have experts selecting the top-tokens. As a result, each token can be routed to a variable number of experts and each expert can have a fixed bucket size. We systematically study pre-training speedups using the same computational resources of the Switch Transformer top-1 and GShard top-2 gating of prior work and find that our method improves training convergence time by more than 2. For the same computational cost, our method demonstrates higher performance in fine-tuning 11 selected tasks in the GLUE and SuperGLUE benchmarks. For a smaller activation cost, our method outperforms the T5 dense model in 7 out of the 11 tasks. |
2106.04426 | We investigate the training of sparse layers that use different parameters for different inputs based on hashing in large Transformer models. Specifically, we modify the feedforward layer to hash to different sets of weights depending on the current token, over all tokens in the sequence. We show that this procedure either outperforms or is competitive with learning-to-route mixture-of-expert methods such as Switch Transformers and BASE Layers, while requiring no routing parameters or extra terms in the objective function such as a load balancing loss, and no sophisticated assignment algorithm. We study the performance of different hashing techniques, hash sizes and input features, and show that balanced and random hashes focused on the most local features work best, compared to either learning clusters or using longer-range context. We show our approach works well both on large language modeling and dialogue tasks, and on downstream fine-tuning tasks. |
2103.16716 | We introduce a new balanced assignment of experts (BASE) layer for large language models that greatly simplifies existing high capacity sparse layers.
Sparse layers can dramatically improve the efficiency of training and inference by routing each token to specialized expert modules that contain only a small fraction of the model parameters.
However, it can be difficult to learn balanced routing functions that make full use of the available experts; existing approaches typically use routing heuristics or auxiliary expert-balancing loss functions.
In contrast, we formulate token-to-expert allocation as a linear assignment problem, allowing an optimal assignment in which each expert receives an equal number of tokens.
This optimal assignment scheme improves efficiency by guaranteeing balanced compute loads, and also simplifies training by not requiring any new hyperparameters or auxiliary losses. Code is publicly released.111https://github.com/pytorch/fairseq/ |
2202.08906 | Scale has opened new frontiers in natural language processing – but at a high cost.
In response, Mixture-of-Experts (MoE) and Switch Transformers have been proposed as an energy efficient path to even larger and more capable language models.
But advancing the state-of-the-art across a broad set of natural language tasks has been hindered by training instabilities and uncertain quality during fine-tuning.
Our work focuses on these issues and acts as a design guide.
We conclude by scaling a sparse model to 269B parameters, with a computational cost comparable to a 32B dense encoder-decoder Transformer (Stable and Transferable Mixture-of-Experts or ST-MoE-32B) .
For the first time, a sparse model achieves state-of-the-art performance in transfer learning, across a diverse set of tasks including reasoning (SuperGLUE, ARC Easy, ARC Challenge) , summarization (XSum, CNN-DM) , closed book question answering (WebQA, Natural Questions) , and adversarially constructed tasks (Winogrande, ANLI R3) .111Code for our models is available athttps://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/moe.py |
2403.12031 | As the range of applications for Large Language Models (LLMs) continues to grow, the demand for effective serving solutions becomes increasingly critical. Despite the versatility of LLMs, no single model can optimally address all tasks and applications, particularly when balancing performance with cost. This limitation has led to the development of LLM routing systems, which combine the strengths of various models to overcome the constraints of individual LLMs. Yet, the absence of a standardized benchmark for evaluating the performance of LLM routers hinders progress in this area. To bridge this gap, we presentRouterBench, a novel evaluation framework designed to systematically assess the efficacy of LLM routing systems, along with a comprehensive dataset comprising
over 405k inference outcomes from representative LLMs to support the development of routing strategies. We further propose a theoretical framework for LLM routing, and deliver a comparative analysis of various routing approaches throughRouterBench, highlighting their potentials and limitations within our evaluation framework. This work not only formalizes and advances the development of LLM routing systems but also sets a standard for their assessment, paving the way for more accessible and economically viable LLM deployments. The code and data are available at https://github.com/withmartian/routerbench. |
2306.06805 | Feature visualization has gained substantial popularity, particularly after the influential work by Olah et al. in 2017, which established it as a crucial tool for explainability.
However, its widespread adoption has been limited due to a reliance on tricks to generate interpretable images, and corresponding challenges in scaling it to deeper neural networks.
Here, we describeMACO, a simple approach to address these shortcomings.
The main idea is to generate images by optimizing the phase spectrum while keeping the magnitude constant to ensure that generated explanations lie in the space of natural images. Our approach yields significantly better results – both qualitatively and quantitatively – and unlocks efficient and interpretable feature visualizations for large state-of-the-art neural networks.
We also show that our approach exhibits an attribution mechanism allowing us to augment feature visualizations with spatial importance.
We validate our method on a novel benchmark for comparing feature visualization methods, and release its visualizations for all classes of the ImageNet dataset onLens. Overall, our approach unlocks, for the first time, feature visualizations for large, state-of-the-art deep neural networks without resorting to any parametric prior image model. |
2305.14739 | Language models (LMs) often struggle to pay enough attention to the input context, and generate texts that are unfaithful or contain hallucinations. To mitigate this issue, we present context-aware decoding (CAD) , which follows a contrastive output distribution that amplifies the difference between the output probabilities when a model is used with and without context.
Our experiments show that CAD, without additional training, significantly improves the faithfulness of different LM families, including OPT, GPT, LLaMA and FLAN-T5 for summarization tasks (e.g., 14.3% gain for LLaMA in factuality metrics) . Furthermore, CAD is particularly effective in overriding a model’s prior knowledge when it contradicts the provided context, leading to substantial improvements in tasks where resolving the knowledge conflict is essential. |
2305.09857 | We introduceCoEdIT, a state-of-the-art text editing system for writing assistance.CoEdITtakes instructions from the user specifying the attributes of the desired text, such as "Make the sentence simpler" or "Write it in a more neutral style," and outputs the edited text. We present a large language model fine-tuned on a diverse collection of task-specific instructions for text editing (a total of 82K instructions) . Our model (1) achieves state-of-the-art performance on various text editing benchmarks, (2) is competitive with publicly available largest-sized LLMs trained on instructions while being60x smaller, (3) is capable of generalizing to unseen edit instructions, and (4) exhibits abilities to generalize to composite instructions containing different combinations of edit actions.
Through extensive qualitative and quantitative analysis, we show that writers prefer the edits suggested byCoEdIT, relative to other state-of-the-art text editing models111Code, data, and models available athttps://github.com/vipulraheja/coedit. |
2402.01306 | Kahneman & Tversky’sprospect theorytells us that humans perceive random variables in a biased but well-defined manner; for example, humans are famously loss-averse.
We show that objectives for aligning LLMs with human feedback implicitly incorporate many of these biases—the success of these objectives (e.g., DPO) over cross-entropy minimization can partly be ascribed to them beinghuman-aware loss functions(HALOs) .
However, the utility functions these methods attribute to humans still differ from those in the prospect theory literature.
Using a Kahneman-Tversky model of human utility, we propose a HALO that directly maximizes the utility of generations instead of maximizing the log-likelihood of preferences, as current methods do.
We call this approach Kahneman-Tversky Optimization (KTO) , and it matches or exceeds the performance of preference-based methods at scales from 1B to 30B.
Crucially, KTO does not need preferences—only a binary signal of whether an output is desirable or undesirable for a given input.
This makes it far easier to use in the real world, where preference data is scarce and expensive. |
2401.00908 | Enterprise documents such as forms, invoices, receipts, reports, contracts, and other similar records, often carry rich semantics at the intersection of textual and spatial modalities. The visual cues offered by their complex layouts play a crucial role in comprehending these documents effectively. In this paper, we presentDocLLM, a lightweight extension to traditional large language models (LLMs) for reasoning over visual documents, taking into account both textual semantics and spatial layout. Our model differs from existing multimodal LLMs by avoiding expensive image encoders and focuses exclusively on bounding box information to incorporate the spatial layout structure. Specifically, the cross-alignment between text and spatial modalities is captured by decomposing the attention mechanism in classical transformers to a set of disentangled matrices. Furthermore, we devise a pre-training objective that learns to infill text segments. This approach allows us to address irregular layouts and heterogeneous content frequently encountered in visual documents. The pre-trained model is fine-tuned using a large-scale instruction dataset, covering four core document intelligence tasks. We demonstrate that our solution outperforms SotA LLMs on 14 out of 16 datasets across all tasks, and generalizes well to 4 out of 5 previously unseen datasets. |
2111.12958 | State-of-the-art frameworks in self-supervised learning have recently shown that fully utilizing transformer-based models can lead to performance boost compared to conventional CNN models. Striving to maximize the mutual information of two views of an image, existing works apply a contrastive loss to the final representations. Motivated by self-distillation in the supervised regime, we further exploit this by allowing theintermediate representationsto learn from the final layer via the contrastive loss. Through self-distillation, the intermediate layers are better suited for instance discrimination, making the performance of an early-exited sub-network not much degraded from that of the full network. This renders the pretext task easier also for the final layer, leading to better representations. Our method, Self-Distilled Self-Supervised Learning (SDSSL) , outperforms competitive baselines (SimCLR, BYOL and MoCo v3) using ViT on various tasks and datasets. In the linear evaluation and k-NN protocol, SDSSL not only leads to superior performance in the final layers, but also in most of the lower layers. Furthermore, qualitative and quantitative analyses show how representations are formed more effectively along the transformer layers. Code is available at https://github.com/hagiss/SDSSL. |
2401.09566 | Advancements in large language models (LLMs) have demonstrated remarkable capabilities across a diverse range of applications. These models excel in generating text completions that are contextually coherent and cover an extensive array of subjects. However, the vast datasets required for their training make aligning response styles during the pretraining and instruction tuning phases challenging. Consequently, an additional alignment phase is typically employed, wherein the model is further trained with human preference data to better align its outputs with human expectations. While this process doesn’t introduce new capabilities per se, it does accentuate generation styles innate to the model. This paper explores the utilization of counterfactual prompting within the framework of Direct Preference Optimization (DPO) to align the model’s style without relying on human intervention. We demonstrate that this method effectively instils desirable behaviour, mitigates undesirable ones, and encourages the model to disregard inappropriate instructions. Our findings suggest that counterfactual prompting with DPO presents a low-resource way to fine-tune LLMs to meet the demands for responsible and ethically aligned AI systems. |
2401.01301 | Large language models (LLMs) have the potential to transform the practice of law, but this potential is threatened by the presence of legal hallucinations—responses from these models that are not consistent with legal facts. We investigate the extent of these hallucinations using an original suite of legal queries, comparing LLMs’ responses to structured legal metadata and examining their consistency. Our work makes four key contributions: (1) We develop a typology of legal hallucinations, providing a conceptual framework for future research in this area.
(2) We find that legal hallucinations are alarmingly prevalent, occurring between 69% of the time with ChatGPT 3.5 and 88% with Llama 2, when these models are asked specific, verifiable questions about random federal court cases.
(3) We illustrate that LLMs often fail to correct a user’s incorrect legal assumptions in a contra-factual question setup.
(4) We provide evidence that LLMs cannot always predict, or do not always know, when they are producing legal hallucinations.
Taken together, these findings caution against the rapid and unsupervised integration of popular LLMs into legal tasks. Even experienced lawyers must remain wary of legal hallucinations, and the risks are highest for those who stand to benefit from LLMs the most—pro selitigants or those without access to traditional legal resources.111All our code, raw data, prompts, and results are available at:https://github.com/reglab/legal_hallucinations. |
2311.06720 | Large language models (LLMs) such as T0, FLAN, and OPT-IML, excel in multi-tasking under a unified instruction-following paradigm, where they also exhibit remarkable generalization abilities to unseen tasks. Despite their impressive performance, these LLMs, with sizes ranging from several billion to hundreds of billions of parameters, demand substantial computational resources, making their training and inference expensive and inefficient. Furthermore, adapting these models to downstream applications, particularly complex tasks, is often unfeasible due to the extensive hardware requirements for finetuning, even when utilizing parameter-efficient approaches such as prompt tuning. Additionally, the most powerful multi-task LLMs, such as OPT-IML-175B and FLAN-PaLM-540B, are not publicly accessible, severely limiting their customization potential. To address these challenges, we introduce a pretrained small scorer,Cappy, designed to enhance the performance and efficiency of multi-task LLMs. With merely 360 million parameters, Cappy functions either independently on classification tasks or serve as an auxiliary component for LLMs, boosting their performance. Moreover, Cappy enables efficiently integrating downstream supervision without requiring LLM finetuning nor the access to their parameters. Our experiments demonstrate that, when working independently on 11 language understanding tasks from PromptSource, Cappy outperforms LLMs that are several orders of magnitude larger. Besides, on 45 complex tasks from BIG-Bench, Cappy boosts the performance of the advanced multi-task LLM, FLAN-T5, by a large margin. Furthermore, Cappy is flexible to cooperate with other LLM adaptations, including finetuning and in-context learning, offering additional performance enhancement.111Code and model available athttps://github.com/tanyuqian/cappyandhttps://huggingface.co/btan2/cappy-large, respectively. |
2401.14489 | While GPUs are responsible for training the vast majority of state-of-the-art deep learning models, the implications of their architecture are often overlooked when designing new deep learning (DL) models. As a consequence, modifying a DL model to be more amenable to the target hardware can significantly improve the runtime performance of DL training and inference. In this paper, we provide a set of guidelines for users to maximize the runtime performance of theirtransformermodels. These guidelines have been created by carefully considering the impact of various model hyperparameters controlling model shape on the efficiency of the underlying computation kernels executed on the GPU. We find the throughput of models with “efficient” model shapes is up to 39% higher while preserving accuracy compared to models with a similar number of parameters but with unoptimized shapes. |
2403.01081 | This work introduces LAB (Large-scale Alignment for chatBots) , a novel methodology designed to overcome the scalability challenges in the instruction-tuning phase of large language model (LLM) training. Leveraging a taxonomy-guided synthetic data generation process and a multi-phase tuning framework, LAB significantly reduces reliance on expensive human annotations and proprietary models like GPT-4. We demonstrate that LAB-trained models can achieve competitive performance across several benchmarks compared to models trained with traditional human-annotated or GPT-4 generated synthetic data. Thus offering a scalable, cost-effective solution for enhancing LLM capabilities and instruction-following behaviors without the drawbacks of catastrophic forgetting, marking a step forward in the efficient training of LLMs for a wide range of applications. |
1905.13727 | We study lossy gradient compression methods to alleviate the communication bottleneck in data-parallel distributed optimization.
Despite the significant attention received, current compression schemes either do not scale well, or fail to achieve the target test accuracy.
We propose a new low-rank gradient compressor based on power iteration that can i) compress gradients rapidly, ii) efficiently aggregate the compressed gradients using all-reduce, and iii) achieve test performance on par with SGD.
The proposed algorithm is the only method evaluated that achieves consistent wall-clock speedups when benchmarked against regular SGD using highly optimized off-the-shelf tools for distributed communication.
We demonstrate reduced training times for convolutional networks as well as LSTMs on common datasets.
Our code is available athttps://github.com/epfml/powersgd. |
2001.08885 | Training machine learning models on mobile devices has the potential of improving both privacy and accuracy of the models. However, one of the major obstacles to achieving this goal is the memory limitation of mobile devices.
Reducing training memory enables models with high-dimensional weight matrices, like automatic speech recognition (ASR) models, to be trained on-device.
In this paper, we propose approximating the gradient matrices of deep neural networks using a low-rank parameterization as an avenue to save training memory.
The low-rank gradient approximation enables more advanced, memory-intensive optimization techniques to be run on device.
Our experimental results show that we can reduce the training memory by about 33.0% for Adam optimization. It uses comparable memory to momentum optimization and achieves a 4.5% relative lower word error rate on an ASR personalization task. |
2403.12014 | Recent state-of-the-art approaches for embodied learning via interaction directly employ large language models (LLMs) as agents to determine the next steps in an environment. Due to their world knowledge and reasoning capabilities, LLM agents achieve stronger performance than previous smaller agents based on reinforcement learning (RL) ; however, frequently calling LLMs is slow and expensive.
This begs an interesting question:Instead of directly employing LLMs as embodied agents, can we use LLMs’ reasoning capabilities to adaptively create training environments to help smaller embodied RL agents learn useful skills that they are weak at?In this work, we proposeEnvGen, a novel framework to address this question.
First, we prompt an LLM to generate training environments that allow agents to quickly learn different tasks in parallel. Concretely, the LLM is given the task description and environment simulator objectives that the agents should learn and is then asked to generate a set of environment configurations (e.g., different terrains, items initially given to agents, chance of finding certain objects,etc.) . Next, we train a small RL agent in a mixture of the original and LLM-generated environments. Then, we enable the LLM tocontinuously adaptthe generated environments to progressively improve the skills that the agent is weak at, by providing feedback to the LLM in the form of the agent’s performance.
We demonstrate the usefulness of EnvGen with comprehensive experiments in Crafter and Heist game environments. We find that a small RL agent trained with EnvGen can outperform SOTA methods, including a GPT-4 agent, and learns long-horizon tasks significantly faster. We also show qualitatively how the LLM adapts training environments to help improve RL agents’ weaker skills over time. Additionally, EnvGen is substantially more efficient as it only uses a small number of LLM calls (e.g., 4 in total) , whereas LLM agents require one or more LLM calls per step (resulting in thousands of LLM calls per episode) . Lastly, we present detailed ablation studies for EnvGen’s design choices. |
Subsets and Splits