Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
50 篇论文 · 按点赞排序
Petr Sychev, Andrey Goncharov, Daniil Vyazhev +2 authors
Token-wise entropy and model-as-judge (MASJ) are evaluated as uncertainty estimation methods for multiple-choice question-answering tasks, showing that entropy performs better in knowledge-dependent domains but requires reasoning, while MASJ needs refinement.
Kechen Li, Wenqi Zhu, Coralia Cartis +2 authors
LLMs are investigated for solving nonnegative polynomial problems, showing significant improvement with reasoning instructions, and suggesting potential for tackling NP-hard mathematical challenges.
Zhuo Chen, Oriol Mayné i Comas, Zhuotao Jin +2 authors
A scaling law for bipartite mutual information in natural language helps understand long-context language modeling and guides the development of large language models with longer context lengths.
Toby Simonds, Akira Yoshiyama
LADDER, a framework for autonomous difficulty-driven problem-solving, and TTRL, a test-time reinforcement learning approach, enhance Large Language Models' performance on mathematical integration tasks without human supervision.
Qiuchen Wang, Ruixue Ding, Zehui Chen +4 authors
ViDoRAG, a multi-agent RAG framework using GMM-based hybrid strategy, enhances performance on visually rich documents requiring complex reasoning by addressing limitations in feature integration and reasoning capacity.
Benjamin Schneider, Florian Kerschbaum, Wenhu Chen
ABC is a multimodal embedding model using a vision-language backbone that achieves top performance on MSCOCO, classification, and VQA tasks, offering high-quality representations with flexible natural language control.
Yuzhe Gu, Wenwei Zhang, Chengqi Lyu +2 authors
Mask-DPO, a fine-grained factuality alignment method based on Direct Preference Optimization, improves the factuality of LLM responses by learning from factually correct sentences and generalizes to unseen questions and topics.
Disen Lan, Weigao Sun, Jiaxi Hu +2 authors
Liger, a novel method, converts pretrained large language models into gated linear recurrent structures using pretrained weights and lightweight fine-tuning, recovering performance and improving efficiency.
Karolina Stańczak, Nicholas Meade, Mehar Bhatia +14 authors
Improving large language model alignment requires incorporating insights from societal, economic, and contractual alignment frameworks to address the underspecified nature of alignment objectives.
Yisen Li, Lingfeng Yang, Wenxuan Shen +4 authors
The paper explores using a multi-dimensional approach, informed by multiple LLMs and a reward model, to select high-quality instruction data, improving smaller models' performance in instruction-following tasks.
Rui Zhao, Weijia Mao, Mike Zheng Shou
DoraCycle adapts generative models to specific domains using unpaired data through text-to-image and image-to-text cycles, enabling both independent and paired knowledge tasks.
Xinyi Wan, Penghui Qi, Guangxing Huang +2 authors
Memory offload in pipeline parallelism effectively reduces activation memory and improves throughput for large language models.
Ziyi Yang, Fanqi Wan, Longguang Zhong +3 authors
FuseChat-3.0, a suite of large language models, integrates diverse source models into smaller targets using supervised fine-tuning and Direct Preference Optimization, achieving significant performance improvements across various tasks.
Zhenhua Liu, Lijun Li, Ruizhe Chen +4 authors
A novel framework, Iterative Value Function Optimization, improves value-guided decoding in reinforcement learning by enhancing value function estimation accuracy, aligning language models and reducing computational costs.
Chengsong Huang, Langlin Huang, Jixuan Leng +2 authors
Self-Calibration enhances the efficiency of test-time scaling in Large Language Models by using model confidence to dynamically adjust sampling strategies.
Jiantao Lin, Xin Yang, Meixi Chen +7 authors
Kiss3DGen leverages a 2D diffusion model to generate high-quality 3D models by fine-tuning for 3D bundle images and using normal maps for mesh reconstruction.
Toru Lin, Kartik Sachdev, Linxi Fan +2 authors
This work introduces techniques to enhance reinforcement learning for humanoid dexterous manipulation through real-to-sim tuning, generalized reward design, divide-and-conquer distillation, and mixed object representations, achieving promising performance on manipulation tasks.
Jun Li, Che Liu, Wenjia Bai +3 authors
A new approach decomposes medical knowledge into fundamental attributes and visual patterns to enhance visual language models in medical abnormality detection and localization, achieving comparable performance to larger models with less data.
Shuang Li, Yihuai Gao, Dorsa Sadigh +1 authors
The Unified Video Action (UVA) model integrates video and action predictions through a joint latent representation and decoupled decoding for high accuracy and efficient action inference in robotics.
Hamish Ivison, Muru Zhang, Faeze Brahman +2 authors
A study on the scalability of data selection methods in instruction-tuning large language models reveals that a simpler, compute-efficient variant of representation-based data selection outperforms more complex methods.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号