大模型推理能力 - arXivDaily 专题

2509.24299 2025-09-30 cs.CV 82%

SVGThinker: Instruction-Aligned and Reasoning-Driven Text-to-SVG Generation

Hanqi Chen, Zhongyin Zhao, Ye Chen, Zhujin Liang, Bingbing Ni

机构 * Shanghai Jiao Tong University（上海交通大学）； SJTU Paris Elite Institute of Technology（上海交通大学巴黎精英理工学院）

专题命中其他推理：reasoning(title,abstract)；chain-of-thought(abstract)

2507.12948 2025-09-30 cs.LG cs.CL 81%

Probabilistic Soundness Guarantees in LLM Reasoning Chains

Weiqiu You, Anton Xue, Shreya Havaldar, Delip Rao, Helen Jin, Chris Callison-Burch, Eric Wong

机构 * University of Pennsylvania（宾夕法尼亚大学）； University of Texas at Austin（德克萨斯大学奥斯汀分校）

专题命中其他推理：reasoning(title,abstract)；分类 cs.CL、cs.LG

Comments EMNLP 2025 camera ready

2506.04695 2025-09-30 cs.LG cs.AI stat.ML 81%

Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection

Xingwu Chen, Tianle Li, Difan Zou

机构 * School of Computing & Data Science（计算与数据科学学院）； Institute of Data Science（数据科学研究院）； The University of Hong Kong（香港大学）

专题命中其他推理：reasoning(title,abstract)；分类 cs.AI、cs.LG

Comments 28 pages, 5 figures, 2 table

2509.24776 2025-09-30 cs.CV cs.AI 79%

VTPerception-R1: Enhancing Multimodal Reasoning via Explicit Visual and Textual Perceptual Grounding

Yizhuo Ding, Mingkang Chen, Zhibang Feng, Tong Xiao, Wanying Qu, Wenqi Shao, Yanwei Fu

机构 * Fudan University（复旦大学）； Shanghai AI Laboratory（上海人工智能实验室）； The University of Hong Kong（香港大学）； Shenzhen University（深圳大学）； University of Science and Technology of China（中国科学技术大学）

专题命中其他推理：reasoning(title,abstract)；分类 cs.AI

2506.06820 2025-09-30 cs.CL cs.SD eess.AS 79%

Beyond Classification: Towards Speech Emotion Reasoning with Multitask AudioLLMs

Wenyu Zhang, Yingxu He, Geyu Lin, Zhuohan Liu, Shuo Sun, Bin Wang, Xunlong Zou, Jeremy H. M. Wong, Qiongqiong Wang, Hardik B. Sailor, Nancy F. Chen, Ai Ti Aw

机构 * Institute for Infocomm Research (I 2 {}^{\text{2}} R), Agency for Science, Technology and Research (A*STAR)（信息与通信研究机构（I2R），科技研究局（A*STAR））； Centre for Frontier AI Research (CFAR), Agency for Science, Technology and Research (A*STAR)（前沿人工智能研究中心（CFAR），科技研究局（A*STAR））

专题命中其他推理：reasoning(title,abstract)；分类 cs.CL

2503.15952 2025-09-30 cs.CL 79%

Adaptive Group Policy Optimization: Towards Stable Training and Token-Efficient Reasoning

Chen Li, Nazhou Liu, Kai Yang

机构 * HPC-AI Tech（HPC-AI技术）

专题命中其他推理：reasoning(title,abstract)；分类 cs.CL

2508.04117 2025-09-30 cs.CL 79%

Unveiling Over-Memorization in Finetuning LLMs for Reasoning Tasks

Zhiwen Ruan, Yun Chen, Yutao Hou, Peng Li, Yang Liu, Guanhua Chen

机构 * Southern University of Science and Technology（南方科技大学）； Tsinghua University（清华大学）； Shanghai University of Finance and Economics（上海财经大学）

专题命中其他推理：reasoning(title,abstract)；分类 cs.CL

2509.25100 2025-09-30 cs.LG cs.AI cs.CL 76%

ORPO-Distill: Mixed-Policy Preference Optimization for Cross-Architecture LLM Distillation

Aasheesh Singh, Vishal Vaddina, Dagnachew Birru

机构 * Phi Labs, Quantiphi（Phi Labs，Quantiphi）

专题命中其他推理：reasoning(abstract,comments)；CoT(abstract)；分类 cs.CL、cs.AI、cs.LG

Comments Accepted at NeurIPS 2025, Efficient Reasoning Workshop

2509.24663 2025-09-30 cs.CL cs.AI cs.LG 75%

InfLLM-V2: Dense-Sparse Switchable Attention for Seamless Short-to-Long Adaptation

Weilin Zhao, Zihan Zhou, Zhou Su, Chaojun Xiao, Yuxuan Li, Yanghao Li, Yudi Zhang, Weilun Zhao, Zhen Li, Yuxiang Huang, Ao Sun, Xu Han, Zhiyuan Liu

机构 * Tsinghua University（清华大学）； OpenBMB（开放大脑实验室）； Harbin Institute of Technology（哈尔滨工业大学）

专题命中其他推理：reasoning(abstract)；chain-of-thought(abstract)；分类 cs.CL、cs.AI、cs.LG

2410.05362 2025-09-30 cs.CL cs.AI cs.LG 67%

LLMs Are In-Context Bandit Reinforcement Learners

Giovanni Monea, Antoine Bosselut, Kianté Brantley, Yoav Artzi

专题命中其他推理：reasoning(abstract)；分类 cs.CL、cs.AI、cs.LG

Comments Published at COLM 2025

2509.25176 2025-09-30 cs.LG cs.CL 62%

SIRI: Scaling Iterative Reinforcement Learning with Interleaved Compression

Haoming Wen, Yushi Bai, Juanzi Li, Jie Tang

机构 * Tsinghua University（清华大学）

专题命中其他推理：reasoning(abstract)；分类 cs.CL、cs.LG

Comments In submission

2509.25131 2025-09-30 cs.SD cs.AI cs.CL cs.CV cs.MM 62%

MGM-Omni: Scaling Omni LLMs to Personalized Long-Horizon Speech

Chengyao Wang, Zhisheng Zhong, Bohao Peng, Senqiao Yang, Yuqi Liu, Haokun Gui, Bin Xia, Jingyao Li, Bei Yu, Jiaya Jia

专题命中其他推理：reasoning(abstract)；分类 cs.CL、cs.AI

Comments Code is available at https://github.com/dvlab-research/MGM-Omni

2509.24356 2025-09-30 cs.CL cs.AI 62%

Beyond Repetition: Text Simplification and Curriculum Learning for Data-Constrained Pretraining

Matthew Theodore Roque, Dan John Velasco

机构 * Samsung R&D Institute Philippines（三星菲律宾研发中心）

专题命中其他推理：reasoning(abstract)；分类 cs.CL、cs.AI

Comments To be published in BabyLM Workshop at EMNLP 2025

2407.03779 2025-09-30 cs.CL cs.LG 62%

Sheaf Discovery with Joint Computation Graph Pruning and Flexible Granularity

Lei Yu, Jingcheng Niu, Zining Zhu, Xi Chen, Gerald Penn

机构 * University of Toronto（多伦多大学）； UKP Lab（UKP实验室）； Stevens Institute of Technology（斯蒂文斯理工学院）

专题命中其他推理：reasoning(abstract)；分类 cs.CL、cs.LG

2509.23755 2025-09-30 cs.CL cs.AI 62%

Understanding Textual Capability Degradation in Speech LLMs via Parameter Importance Analysis

Chao Wang, Rui-Chen Zheng, Yang Ai, Zhen-Hua Ling

机构 * National Engineering Research Center of Speech and Language Information Processing（语音与语言信息处理国家级工程研究中心）； University of Science and Technology of China（中国科学技术大学）

专题命中其他推理：reasoning(abstract)；分类 cs.CL、cs.AI

2509.23412 2025-09-30 cs.CL cs.LG 62%

Comparison of Scoring Rationales Between Large Language Models and Human Raters

Haowei Hua, Hong Jiao, Dan Song

机构 * Princeton University（普林斯顿大学）； University of Maryland（马里兰大学）； University of Iowa（爱荷华大学）

专题命中其他推理：reasoning(abstract)；分类 cs.CL、cs.LG

Comments 23 Pages, 4 Tables, 13 Figures

2509.18521 2025-09-30 cs.LG cs.AI 62%

APRIL: Active Partial Rollouts in Reinforcement Learning to Tame Long-tail Generation

Yuzhen Zhou, Jiajun Li, Yusheng Su, Gowtham Ramesh, Zilin Zhu, Xiang Long, Chenyang Zhao, Jin Pan, Xiaodong Yu, Ze Wang, Kangrui Du, Jialian Wu, Ximeng Sun, Jiang Liu, Qiaolin Yu, Hao Chen, Zicheng Liu, Emad Barsoum

机构 * Advanced Micro Devices, Inc. (AMD)（先进微器件公司）； Carnegie Mellon University (CMU)（卡内基梅隆大学）； LMSYS Org（LMSYS组织）； University of California, Los Angeles (UCLA)（加州大学洛杉矶分校）

专题命中其他推理：reasoning(abstract)；分类 cs.AI、cs.LG

2507.14783 2025-09-30 cs.LG cs.AI 62%

Omni-Thinker: Scaling Multi-Task RL in LLMs with Hybrid Reward and Task Scheduling

Derek Li, Jiaming Zhou, Leo Maxime Brunswic, Abbas Ghaddar, Qianyi Sun, Liheng Ma, Yu Luo, Dong Li, Mark Coates, Jianye Hao, Yingxue Zhang

机构 * Huawei Noah’s Ark Lab, Montréal, Canada（华为诺亚实验室，加拿大蒙特利尔）； McGill University and Mila - Québec AI Institute（麦吉尔大学和魁北克人工智能研究所）； Huawei Noah’s Ark Lab, Beijing, China（华为诺亚实验室，中国北京）

专题命中其他推理：reasoning(abstract)；分类 cs.AI、cs.LG

2505.24357 2025-09-30 cs.LG cs.AI 62%

ReCalKV: Low-Rank KV Cache Compression via Head Reordering and Offline Calibration

Xianglong Yan, Zhiteng Li, Tianao Zhang, Haotong Qin, Linghe Kong, Yulun Zhang, Xiaokang Yang

机构 * Shanghai Jiao Tong University（上海交通大学）； ETH Zürich（苏黎世联邦理工学院）

专题命中其他推理：reasoning(abstract)；分类 cs.AI、cs.LG

2509.24771 2025-09-30 cs.CL 57%

LatentEvolve: Self-Evolving Test-Time Scaling in Latent Space

Guibin Zhang, Fanci Meng, Guancheng Wan, Zherui Li, Kun Wang, Zhenfei Yin, Lei Bai, Shuicheng Yan

机构 * NUS（新加坡国立大学）； USTC（中国科学技术大学）； UCLA（美国大学洛杉矶分校）； NTU（国立科技大学）； Shanghai AI Lab（上海人工智能实验室）

专题命中其他推理：reasoning(abstract)；分类 cs.CL

2509.23793 2025-09-30 cs.CL 57%

Transformer Tafsir at QIAS 2025 Shared Task: Hybrid Retrieval-Augmented Generation for Islamic Knowledge Question Answering

Muhammad Abu Ahmad, Mohamad Ballout, Raia Abu Ahmad, Elia Bruni

机构 * QIAS 2025 Shared Task（QIAS 2025 共同任务）

专题命中其他推理：reasoning(abstract)；分类 cs.CL

Comments Accepted at ArabicNLP 2025, co-located with EMNLP 2025

2509.23109 2025-09-30 cs.AI cs.CV 57%

AttAnchor: Guiding Cross-Modal Token Alignment in VLMs with Attention Anchors

Junyang Zhang, Tianyi Zhu, Thierry Tambe

机构 * California Institute of Technology（加州理工学院）； Stanford University（斯坦福大学）

专题命中其他推理：reasoning(abstract)；分类 cs.AI

Comments 31 pages, 17 figures

2509.19128 2025-09-30 cs.LG 57%

PipelineRL: Faster On-policy Reinforcement Learning for Long Sequence Generation

Alexandre Piché, Ehsan Kamalloo, Rafael Pardinas, Xiaoyin Chen, Dzmitry Bahdanau

机构 * ServiceNow AI Research（ServiceNow人工智能研究）

专题命中其他推理：reasoning(abstract)；分类 cs.LG

2506.12027 2025-09-30 cs.CC cs.LG 57%

Constant Bit-size Transformers Are Turing Complete

Qian Li, Yuyi Wang

机构 * Shenzhen International Center For Industrial And Applied Mathematics（深圳国际工业与应用数学中心）； Shenzhen Research Institute of Big Data（深圳大数据研究院）； CRRC Zhuzhou Institute（CRRC 长沙研究所）； Tengen Intelligence Institute（腾格智能研究院）

专题命中其他推理：reasoning(abstract)；分类 cs.LG

Comments Accepted by NeurIPS 2025

2509.25177 2025-09-30 cs.CV 50%

Mitigating Hallucination in Multimodal LLMs with Layer Contrastive Decoding

Bingkui Tong, Jiaer Xia, Kaiyang Zhou

机构 * Mohamed bin Zayed University of Artificial Intelligence（莫扎德·本·扎耶德人工智能大学）； Hong Kong Baptist University（香港 Baptist大学）

专题命中其他推理：reasoning(abstract)

2504.12083 2025-09-30 cs.CV 50%

Self-alignment of Large Video Language Models with Refined Regularized Preference Optimization

Pritam Sarkar, Ali Etemad

机构 * Queen’s University（女王大学）； Vector Institute（向量研究所）

专题命中其他推理：reasoning(abstract)

Comments NeurIPS 2025

1. 其他推理 26 篇

SVGThinker: Instruction-Aligned and Reasoning-Driven Text-to-SVG Generation

Probabilistic Soundness Guarantees in LLM Reasoning Chains

Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection

VTPerception-R1: Enhancing Multimodal Reasoning via Explicit Visual and Textual Perceptual Grounding

Beyond Classification: Towards Speech Emotion Reasoning with Multitask AudioLLMs

Adaptive Group Policy Optimization: Towards Stable Training and Token-Efficient Reasoning

Unveiling Over-Memorization in Finetuning LLMs for Reasoning Tasks

ORPO-Distill: Mixed-Policy Preference Optimization for Cross-Architecture LLM Distillation

InfLLM-V2: Dense-Sparse Switchable Attention for Seamless Short-to-Long Adaptation

LLMs Are In-Context Bandit Reinforcement Learners

SIRI: Scaling Iterative Reinforcement Learning with Interleaved Compression

MGM-Omni: Scaling Omni LLMs to Personalized Long-Horizon Speech

Beyond Repetition: Text Simplification and Curriculum Learning for Data-Constrained Pretraining

Sheaf Discovery with Joint Computation Graph Pruning and Flexible Granularity

Understanding Textual Capability Degradation in Speech LLMs via Parameter Importance Analysis

Comparison of Scoring Rationales Between Large Language Models and Human Raters

APRIL: Active Partial Rollouts in Reinforcement Learning to Tame Long-tail Generation

Omni-Thinker: Scaling Multi-Task RL in LLMs with Hybrid Reward and Task Scheduling

ReCalKV: Low-Rank KV Cache Compression via Head Reordering and Offline Calibration

LatentEvolve: Self-Evolving Test-Time Scaling in Latent Space

Transformer Tafsir at QIAS 2025 Shared Task: Hybrid Retrieval-Augmented Generation for Islamic Knowledge Question Answering

AttAnchor: Guiding Cross-Modal Token Alignment in VLMs with Attention Anchors

PipelineRL: Faster On-policy Reinforcement Learning for Long Sequence Generation

Constant Bit-size Transformers Are Turing Complete

Mitigating Hallucination in Multimodal LLMs with Layer Contrastive Decoding

Self-alignment of Large Video Language Models with Refined Regularized Preference Optimization