语言大模型 / LLM - arXivDaily 专题

2510.06652 2025-10-09 cs.CL 91%

Aligning Large Language Models via Fully Self-Synthetic Data

Shangjian Yin, Zhepei Wei, Xinyu Zhu, Wei-Lin Chen, Yu Meng

机构 * University of California, Riverside（加州大学河滨分校）； University of Virginia（弗吉尼亚大学）

专题命中后训练与偏好优化：large language model(title,abstract)；language model(title,abstract)；LLM(abstract)；RLHF(abstract)

2510.06988 2025-10-09 cs.CV 78%

No MoCap Needed: Post-Training Motion Diffusion Models with Reinforcement Learning using Only Textual Prompts

Girolamo Macaluso, Lorenzo Mandelli, Mirko Bicchierai, Stefano Berretti, Andrew D. Bagdanov

机构 * University of Florence（佛罗伦萨大学）

专题命中后训练与偏好优化：post-training(title,abstract)

2505.22296 2025-10-09 cs.CL cs.LG 76%

360-LLaMA-Factory: Plug & Play Sequence Parallelism for Long Post-Training

Haosheng Zou, Xiaowei Lv, Shousheng Jia, Lin Li, Xiaochun Gong, Xiangzheng Zhang

机构 * Qiyuan Tech（启元科技）； Renmin University（中国人民大学）

专题命中后训练与偏好优化：post-training(title)；分类 cs.CL、cs.LG

Comments v2: sp for vlms; code at https://github.com/Qihoo360/360-LLaMA-Factory

2510.06391 2025-10-09 cs.CL cs.AI 73%

Reward Model Perspectives: Whose Opinions Do Reward Models Reward?

Elle

机构 * Elle University of Oxford（埃勒奥克舍大学）

专题命中后训练与偏好优化：language model(abstract)；prompting(abstract)；分类 cs.CL、cs.AI

Comments Published at EMNLP 2025 under the full author name "Elle"

2510.06878 2025-10-09 cs.AI 70%

TGPR: Tree-Guided Policy Refinement for Robust Self-Debugging of LLMs

Daria Ozerova, Ekaterina Trofimova

机构 * HSE University（俄罗斯莫斯科国立高等经济学院）

专题命中后训练与偏好优化：large language model(abstract)；language model(abstract)；分类 cs.AI

2510.06866 2025-10-09 cs.CL 70%

Unlocking Latent Discourse Translation in LLMs Through Quality-Aware Decoding

Wafaa Mohammed, Vlad Niculae, Chrysoula Zerva

专题命中后训练与偏好优化：large language model(abstract)；language model(abstract)；分类 cs.CL

2508.03905 2025-10-09 cs.CL 70%

Sotopia-RL: Reward Design for Social Intelligence

Haofei Yu, Zhengyang Qi, Yining Zhao, Kolby Nottingham, Keyang Xuan, Bodhisattwa Prasad Majumder, Hao Zhu, Paul Pu Liang, Jiaxuan You

机构 * University of Illinois Urbana-Champaign（伊利诺伊大学厄巴纳-香槟分校）； University of California Irvine（加州大学尔湾分校）； Allen Institute for Artificial Intelligence（人工智能艾伦研究所）； Carnegie Mellon University（卡内基梅隆大学）； Stanford University（斯坦福大学）； Massachusetts Institute of Technology（麻省理工学院）

专题命中后训练与偏好优化：large language model(abstract)；language model(abstract)；分类 cs.CL

Comments 10 pages

2510.06629 2025-10-09 cs.CR cs.CV cs.LG 57%

Unsupervised Backdoor Detection and Mitigation for Spiking Neural Networks

Jiachen Li, Bang Wu, Xiaoyu Xia, Xiaoning Liu, Xun Yi, Xiuzhen Zhang

机构 * RMIT University（皇家墨尔本理工大学）

专题命中后训练与偏好优化：post-training(abstract)；分类 cs.LG

Comments To appear in The 28th International Symposium on Research in Attacks, Intrusions and Defenses (RAID 2025)

1. 后训练与偏好优化 8 篇

Aligning Large Language Models via Fully Self-Synthetic Data

No MoCap Needed: Post-Training Motion Diffusion Models with Reinforcement Learning using Only Textual Prompts

360-LLaMA-Factory: Plug & Play Sequence Parallelism for Long Post-Training

Reward Model Perspectives: Whose Opinions Do Reward Models Reward?

TGPR: Tree-Guided Policy Refinement for Robust Self-Debugging of LLMs

Unlocking Latent Discourse Translation in LLMs Through Quality-Aware Decoding

Sotopia-RL: Reward Design for Social Intelligence

Unsupervised Backdoor Detection and Mitigation for Spiking Neural Networks