Spaces:

---

huggingface

/

HuggingDiscussions

like 57

Running

https://arxiv.org/abs/2406.01954

We are excited to share our recent work on MLLM architecture design titled "Ovis: Structural Embedding Alignment for Multimodal Large Language Model".

Paper: https://arxiv.org/abs/2405.20797
Github: https://github.com/AIDC-AI/Ovis
Model: https://huggingface.co/AIDC-AI/Ovis-Clip-Llama3-8B
Data: https://huggingface.co/datasets/AIDC-AI/Ovis-dataset

@Yiwen-ntu for now we support only videos as paper covers in the Daily.

we are excited to share our work titled "Hierarchical Prompting Taxonomy: A Universal Evaluation Framework for Large Language Models" : https://arxiv.org/abs/2406.12644

doi:10.57967/hf/4791
🚀 ScrapeGoat Music Models: Now Available! 🎵
After our recent work on HCF-Aware Training to optimize memory efficiency in language models (especially beneficial for specialized models like ours), we're excited to announce that all training and weights for the ScrapeGoat Music project are now available in our Hugging Face repositories:

👉 https://lnkd.in/gVrZ56JS

👉 https://lnkd.in/g3wS3rS4

ALL training python codes and weights available in the repo on HF.

Dear AK and HF Team,

Exciting Update! We're thrilled to share WritingBench: A Comprehensive Framework for Evaluating Generative Writing

📃 [Paper] • 🚀 [Github Repo] • 📏 [Critic Model] • ✍️ [Writing Model]

💡WritingBench is a comprehensive benchmark for evaluating LLMs' writing capabilities across 1,239 real-world queries, spanning:

• 6 primary domains and 100 fine-grained subdomains
• 3 core writing requirements: Style / Format / Length
• 1,546 avg. tokens per query, integrating diverse sources of materials
• Each query is paired with 5 instance-specific criteria

Regards,
Yuning

Hello AK and HF Team,

We would to add our recent paper "Simple Semi-supervised Knowledge Distillation from Vision-Language Models via Dual-Head Optimization" in HF daily papers.
I'm putting request here because I don't have paper claimed in HF daily papers yet.

Paper: https://arxiv.org/pdf/2503.02424

Regards,
Seongjae

@erjui kudos on great paper: https://huggingface.co/papers/2505.07675

Excited to share our new work: Time-R1 is a framework designed to endow Language Models (LLMs) with comprehensive temporal reasoning capabilities, enabling them to progressively cultivate sophisticated temporal logic from past events, predict future occurrences, and creatively generate plausible future scenarios. Our 3B parameter language model trained with a novel three-stage reinforcement learning curriculum and dynamic rewards, outperforming state-of-the-art models over 200 times its size, including DeepSeek-R1, on challenging future-oriented tasks.

Hello AK and HF Team,

We would to add our recent paper "MM-Agent: LLM as Agents for Real-world Mathematical Modeling Problem"

Mathematical modeling is more than reasoning – it requires open-ended analysis, abstraction, and principled formulation. This work introduces MM-Agent, a large language model (LLM)-powered agent framework designed to tackle real-world mathematical modeling tasks end-to-end.

Key highlights:

📊 Proposes MM-Bench: 111 curated problems from MCM/ICM (2000–2025), across physics, biology, economics, etc.

🧩 MM-Agent decomposes modeling into 4 expert-inspired stages:
Problem Analysis → Model Formulation → Problem Solving → Report Generation

🚀 Outperforms baselines by 11.88% over expert-written solutions using GPT-4o, while costing just $0.88 and 15 minutes per task.

🏆 Helped two undergrad teams win Finalist Award (top 2%) in MCM/ICM 2025.

📄 Paper: https://arxiv.org/abs/2505.14148

💻 Code: https://github.com/usail-hkust/LLM-MM-Agent

Hi, everyone! We are happy to share with you our work SVD-Free Low-Rank Adaptive Gradient Optimization for Large Language Models.

We focus on the low-rank compression of optimizer states and propose replacing the expensive SVD decomposition with a fixed orthogonal matrix that comes from the Discrete Consine Transformation (DCT).

In our work we couple the DCT matrix with a theoretically-justified approach to choose the most appropriate columns from the DCT matrix that minimize the reconstruction error for each gradient matrix G and obtain a dynamic projection matrix tailored to each gradient G.

Our numerical results show that DCT matrix not only recovers the performance of existing low-rank optimizers, but also reduces the running time by 20% and memory usage for large models, both for pretraining and finetuning.

📜 Paper: https://arxiv.org/pdf/2505.17967

🐍 Code: soon to appear in https://github.com/IST-DASLab/ISTA-DASLab-Optimizers via pip