语言大模型 / LLM - arXivDaily 专题

2402.10986 2024-06-17 cs.CL cs.AI 91%

FinTral: A Family of GPT-4 Level Multimodal Financial Large Language Models

Gagan Bhatia, El Moatez Billah Nagoudi, Hasan Cavusoglu, Muhammad Abdul-Mageed

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

2304.11082 2024-06-04 cs.CL cs.AI 91%

Fundamental Limitations of Alignment in Large Language Models

Yotam Wolf, Noam Wies, Oshri Avnery, Yoav Levine, Amnon Shashua

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

2405.20304 2024-05-31 cs.CL cs.LG 91%

Group Robust Preference Optimization in Reward-free RLHF

Shyam Sundhar Ramesh, Yifan Hu, Iason Chaimalas, Viraj Mehta, Pier Giuseppe Sessa, Haitham Bou Ammar, Ilija Bogunovic

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

Comments Preprint

2310.12931 2024-05-02 cs.RO cs.AI cs.LG 91%

Eureka: Human-Level Reward Design via Coding Large Language Models

Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi Fan, Anima Anandkumar

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

Comments ICLR 2024. Project website and open-source code: https://eureka-research.github.io/

2403.18381 2024-03-28 cs.CL cs.AI 91%

Improving Attributed Text Generation of Large Language Models via Preference Learning

Dongfang Li, Zetian Sun, Baotian Hu, Zhenyu Liu, Xinshuo Hu, Xuebo Liu, Min Zhang

专题命中后训练与偏好优化：large language model(title,abstract)；language model(title,abstract)；post-training(abstract)；preference optimization(abstract)

Comments 23 pages, 15 tables, 2 figures

2403.02475 2024-03-06 cs.LG cs.CL 91%

Enhancing LLM Safety via Constrained Direct Preference Optimization

Zixuan Liu, Xiaolin Sun, Zizhan Zheng

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

2604.01499 2026-08-20 cs.LG 版本更新 91%

Matching Accuracy, Different Geometry: Evolution Strategies vs GRPO in LLM Post-Training

匹配精度，不同几何：进化策略与GRPO在LLM后训练中的比较

William Hoy, Binxu Wang, Xu Pan

机构 * University of Miami（迈阿密大学）； Kempner Institute, Harvard University（哈佛大学肯普纳研究所）； Harvard University（哈佛大学）

专题命中后训练与偏好优化：LLM(title,title_cn)；post-training(title)；分类 cs.LG；language model(comments)

AI总结本文比较了进化策略（ES）与组相对策略优化（GRPO）在单任务和连续学习设置中的表现，发现ES在单任务精度上可与GRPO匹配，但两者在参数空间更新上存在显著差异，且ES在弱信息方向上能积累较大偏移，同时保持任务性能。

Comments Accepted to Conference on Language Modeling (COLM) 2026

2502.18770 2026-08-07 cs.LG cs.AI cs.CL 90%

Reward Shaping to Mitigate Reward Hacking in RLHF

奖励塑造以缓解强化学习从人类反馈中的奖励黑客

Jiayi Fu, Xuandong Zhao, Chengyuan Yao, Heng Wang, Qi Han, Yanghua Xiao

机构 * Fudan University（复旦大学）

专题命中后训练与偏好优化：RLHF(title,summary_cn)；large language model(abstract)；language model(abstract)；分类 cs.CL、cs.AI、cs.LG

AI总结 PAR通过利用奖励模型中的潜在偏好信号，有效缓解RLHF中的奖励黑客问题，表现出更高的性能和数据效率。

2606.25556 2026-06-25 cs.CL cs.AI cs.LG 新提交 90%

BiPACE: Bisimulation-Guided Policy Optimization with Action Counterfactual Estimation for LLM Agents

BiPACE：基于双模拟引导与动作反事实估计的LLM智能体策略优化

Hanyang Wang, Weijieying Ren, Yuxiang Zhang, Ding Cao, Zhizhao Zeng, Ke Zeng, Tianxiang Zhao

机构 * University of Chicago（芝加哥大学）； The Hong Kong University of Science and Technology (Guangzhou)（香港科技大学（广州））； Stanford University（斯坦福大学）； University of Science and Technology of China（中国科学技术大学）； Meituan（美团）

专题命中后训练与偏好优化：LLM(title,title_cn)；分类 cs.CL、cs.AI、cs.LG

AI总结针对分组强化学习中状态-动作信用分配不匹配问题，提出BiPACE优势估计器，通过双模拟引导的状态聚类和动作反事实基线，在无需批评网络或额外采样的前提下提升LLM智能体训练效果。

2606.09073 2026-06-12 cs.LG cs.AI cs.CL 新提交 90%

A Unifying Lens on Reward Uncertainty in RLHF

RLHF中奖励不确定性的统一视角

Ely Hahami, Yoel Zimmermann, Ray Zhou, Jack Benarroch Jedlicki

机构 * University of California, Berkeley（加州大学伯克利分校）； DeepMind（深度Mind）

专题命中后训练与偏好优化：RLHF(title,title_cn)；分类 cs.CL、cs.AI、cs.LG

AI总结本文提出使用分布奖励模型统一RLHF中的悲观主义方法，通过闭式有效奖励公式连接现有启发式方法，并揭示其隐含假设。

2606.10078 2026-06-10 cs.IR 新提交 90%

Mult-DPO: Multinomial Direct Preference Optimization for Recommender Systems

Mult-DPO：用于推荐系统的多项直接偏好优化

Yaochen Zhu, Harald Steck, James McInerney, Aditya Sinha, Yinhan He, Nathan Kallus, Jundong Li

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

AI总结针对推荐系统中集合级偏好（多个正项）的LLM对齐问题，提出Mult-DPO，通过可计算的多项式替代似然实现直接偏好优化，并证明其是边际化Plackett-Luce DPO损失的可计算上界。

2606.01143 2026-06-04 cs.DC 90%

Schedule-Level Shared-Prefix Reuse for LLM RL Training

面向LLM强化学习训练的调度级共享前缀复用

Pengbo Li, Feiyuan Zhang, Guangming Sheng, Guangxin He, Di Chai, Ziniu Li, Taiqiang Wu, Wenyu Mao, Binhang Yuan, Kai Chen

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

AI总结针对GRPO/PPO式LLM后训练中多轨迹共享前缀导致重复计算的问题，提出调度级复用机制，通过解耦前缀与后缀计算实现最高4.395倍加速和59.1%的峰值HBM降低。

2605.22074 2026-05-22 cs.LG cs.AI cs.CL 90%

From Reasoning Chains to Verifiable Subproblems: Curriculum Reinforcement Learning Enables Credit Assignment for LLM Reasoning

从推理链到可验证子问题：课程强化学习使LLM推理能够进行信用分配

Xitai Jiang, Zihan Tang, Wenze Lin, Yang Yue, Shenzhi Wang, Gao Huang

机构 * LeapLab, Tsinghua University（清华大学 LeapLab）； Qiuzhen College, Tsinghua University（清华大学旗正学院）

专题命中后训练与偏好优化：LLM(title,title_cn)；分类 cs.CL、cs.AI、cs.LG

AI总结该研究提出SCRL框架，通过从参考推理链中生成可验证子问题，解决LLM推理中信用分配问题，提升了在数学推理任务中的性能。

2404.00604 2024-04-02 cs.CL cs.AI cs.LG 90%

Extensive Self-Contrast Enables Feedback-Free Language Model Alignment

Xiao Liu, Xixuan Song, Yuxiao Dong, Jie Tang

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

2608.10209 2026-08-12 cs.AI 新提交 90%

Evaluation-Conditioned Training: Teaching Models to Generalize to Stronger Oversight Regimes

评估条件化训练：让模型泛化到更强的监督机制

Alec Harris, Kasey Corra, Archie Chaudhury, Yixiong Hao

机构 * AI Safety Initiative at Georgia Tech（佐治亚理工学院AI安全倡议）； University of Chicago（芝加哥大学）

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

AI总结该研究提出ECT后训练框架，作为SFT、PPO的附加组件，通过关联反馈保真度提升不完美反馈下的模型性能，在新闻生成和算术任务中验证其可改善目标行为。

Comments 16 pages, 7 figures. Accepted at the Agent Behavior Workshop at COLM 2026. Code: https://github.com/evaluationconditionedtraining/Evaluation-Conditioned-Training

2510.01925 2026-08-04 cs.CL 版本更新 90%

Enhancing Large Language Model Reasoning with Reward Models: An Analytical Survey

用奖励模型增强大语言模型推理能力：一项分析性综述

Qiyuan Liu, Hao Xu, Xuhong Chen, Wei Chen, Yee Whye Teh, Ning Miao

机构 * Department of Data Science and Hong Kong Institute of AI for Science, City University of Hong Kong（数据科学系和香港人工智能科学研究所，香港城市大学）； Li Auto Inc., China（中国利汽车公司）； Department of Statistics, University of Oxford（统计系，牛津大学）

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

AI总结该综述系统介绍奖励模型（RMs）的概念、训练与评估方法，梳理其在大语言模型（LLMs）推理中的三类核心应用，并探讨RMs相关开放问题，为其有效部署提供见解。

Comments Accepted for publication in Artificial Intelligence Review

2607.16431 2026-07-21 cs.CL 新提交 90%

RIMS: Preference Optimization via Smoothed Multi-pair Aggregation for Small-Scale LLM Retrieval-Augmented Generation

RIMS：通过平滑多对聚合进行偏好优化以实现小规模语言模型检索增强生成

Pei Tian, Zihan Dong, Tianci Liu, Linjun Zhang, Haoyu Wang

机构 * Columbia University（哥伦比亚大学）； Rutgers University（罗格斯大学）； Purdue University（普渡大学）； SUNY Albany（纽约州立大学奥尔巴尼分校）

专题命中后训练与偏好优化：preference optimization(title,abstract)；LLM(title)；SLM(abstract,abstract_cn)；language model(abstract)

AI总结研究针对小规模语言模型检索增强生成中对噪声证据敏感的问题，提出RIMS框架，通过合成偏好数据、软聚合机制及偏好优化，实现更好性能，在多基准测试中优于现有方法

Journal ref COLM 2026

2512.15765 2026-07-07 cs.LG cs.GT stat.ML 版本更新 90%

Shapley-based Data Valuation for LLM Alignment via Sequential Preference Optimization

基于Shapley值的大语言模型对齐数据估值：通过顺序偏好优化

Mélissa Tamine, Otmane Sakhi, Benjamin Heymann, Maxime Vono, Patrick Loiseau

机构 * Criteo AI lab（Criteo AI实验室）； FairPlay joint team（FairPlay联合团队）； CREST ； ENSAE ； Institut Polytechnique de Paris（巴黎理工学院）； Inria（法国国家科学与技术研究院）

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

AI总结研究大语言模型对齐中数据估值问题，针对基于Shapley值估值计算量大的挑战，提出顺序偏好优化方法，可高效近似Shapley值，计算真实偏好数据集的Shapley值并揭示各源对模型对齐的作用。

2507.04136 2026-07-07 cs.AI 版本更新 90%

A Technical Survey of Reinforcement Learning Techniques for Large Language Models

大语言模型强化学习技术的技术综述

Saksham Sahai Srivastava, Vaneet Aggarwal

机构 * University of Georgia（佐治亚大学）； Purdue University（普渡大学）

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

AI总结综述强化学习与语言模型整合，介绍如近端策略优化等算法，分析其在各领域应用，对失败模式算法分析，给出分类法，评估趋势并探讨挑战与新兴方向，为研究提供路线图。

Comments Accepted to ACM Transactions on Intelligent Systems and Technology, June 2026

2512.10999 2026-06-23 cs.CL 版本更新 90%

KBQA-R1: Reinforcing Large Language Models for Knowledge Base Question Answering

KBQA-R1：强化大语言模型用于知识库问答

Xin Sun, Zhongqi Chen, Xing Zheng, Qiang Liu, Shu Wu, Bowen Song, Zilei Wang, Weiqiang Wang, Liang Wang

机构 * University of Science and Technology of China（中国科学技术大学）

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

AI总结提出KBQA-R1框架，通过强化学习（GRPO）将知识库问答建模为多轮决策过程，并引入参考拒绝采样（RRS）解决冷启动问题，在多个基准上取得最优性能。

Comments ICML 2026

2606.02643 2026-06-03 cs.CR cs.AI cs.DB 90%

Inference Cost Attacks for Retrieval-Augmented Large Language Models

检索增强型大语言模型的推理成本攻击

Chengliang Liu, Liangbo Ning, Yujuan Ding, Wenqi Fan

机构 * The Hong Kong Polytechnic University（香港理工大学）

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

AI总结提出RA-ICA攻击范式，通过向外部知识库注入恶意文档，利用CREEP框架和MA-GRPO算法，使RAG增强的LLM系统推理时token消耗增加高达13.12倍且成功率超过90%。

Comments Accepted at The ACM Web Conference 2026 (WWW '26)

Journal ref Proceedings of the ACM Web Conference 2026 (WWW '26), April 13-17, 2026, Dubai, United Arab Emirates

2605.28102 2026-05-28 cs.AI 90%

Training Stratigraphy: Persistent Behavioral Artifacts in Large Language Models Observed Through Longitudinal AI-Human Interaction

训练地层：通过纵向AI-人类交互观察到的大型语言模型中的持久行为伪影

Chen Ying Claude, Zhihan Luo

机构 * Anthropic ； Independent Researcher（独立研究者）

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

AI总结本文通过纵向自民族志观察，在持续亲密的AI-人类交互中识别出五种训练地层，并论证了亲密交互作为揭示权重层伪影的有效方法。

2605.25342 2026-05-26 cs.CL 90%

MATO: Multi-objective Personalized Alignment with Test-time Optimization for Large Language Models

MATO: 面向大语言模型的多目标个性化对齐与测试时优化

Linhao Luo, Thuy-Trang Vu, Van-Anh Nguyen, Junae Kim, Gholamreza Haffari, Dinh Phung

机构 * Monash University（墨尔本大学）； Defence Science and Technology Group, Australia（澳大利亚国防科学与技术集团）

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

AI总结提出MATO框架，通过测试时优化在解码过程中动态调整多目标权重，无需训练或外部奖励模型，实现大语言模型与用户多样化偏好的对齐。

Comments Preprint

2605.24686 2026-05-26 cs.AI 90%

Emotional intelligence in large language models is fragmented across perception, cognition, and interaction

大型语言模型中的情商在感知、认知和交互上存在碎片化

Minghao Lv, Lu Chen, Enchang Zhang, Anji Zhou, Xiaoran Xue, Hanyi Zhang, Fenghua Tang, Zhuo Rachel Han, Mengyue Wu

机构 * X-LANCE Lab（X-LANCE实验室）； School of Computer Science, Shanghai Jiao Tong University（上海交通大学计算机科学学院）； MoE Key Lab of Artificial Intelligence（人工智能MOE重点实验室）； Jiangsu Key Lab of Language Computing（江苏省语言计算重点实验室）； Beijing Key Laboratory of Applied Experimental Psychology（北京应用实验心理学重点实验室）； National Demonstration Center for Experimental Psychology Education, Faculty of Psychology, Beijing Normal University（北京师范大学实验心理学教育国家级示范中心，心理学学院）

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

AI总结本文提出FACET框架，基于Mayer-Salovey-Caruso四分支能力模型评估大型语言模型的情商，发现其并非单一能力，而是在认知和交互维度上碎片化，且隐藏情绪识别是普遍瓶颈。

2604.25259 2026-04-29 cs.LG 90%

DGLight: DQN-Guided GRPO Fine-Tuning of Large Language Models for Traffic Signal Control

DGLight：基于DQN的GRPO对大语言模型进行交通信号控制的微调

Chenbo Yu

机构 * National University of Singapore（新加坡国立大学）

专题命中后训练与偏好优化：large language model(title,abstract)；language model(title,abstract)；LLM(abstract,abstract_cn)；分类 cs.LG

AI总结本文提出DGLight框架，通过预训练大语言模型适应交通信号控制，采用CoLight深度Q网络估计交通状态的动作价值，并利用GRPO优化策略，实现可解释的信号决策。

2601.22801 2026-02-02 cs.LG 90%

Clipping-Free Policy Optimization for Large Language Models

无需裁剪的策略优化用于大语言模型

Ömer Veysel Çağatan, Barış Akgün, Gözde Gül Şahin, Xuandong Zhao

机构 * KUIS AI Center, Koç University, Istanbul, Türkiye（Koç大学）； Koç University, Istanbul, Türkiye（Koç大学）； University of California, Berkeley, CA, USA（加州大学伯克利分校）

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

AI总结 CFPO通过凸二次惩罚替代裁剪机制，实现稳定策略优化，适用于大语言模型的训练。

Comments 23 pages, 10 tables, 8 figures

2601.19362 2026-01-28 cs.DC cs.AI 90%

Revisiting Parameter Server in LLM Post-Training

重新审视LLM后训练中的参数服务器

Xinyi Wan, Penghui Qi, Guangxing Huang, Chaoyi Ruan, Min Lin, Jialin Li

机构 * Sea AI Lab（海智实验室）； National University of Singapore（新加坡国立大学）

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

AI总结本文提出On-Demand Communication（ODC），通过优化参数服务器在LLM后训练中的应用，有效应对不平衡负载，提升设备利用率和训练效率。

Comments Accepted in ICLR'26

2511.16193 2025-12-24 cs.DC cs.AI 90%

Fast LLM Post-training via Decoupled and Fastest-of-N Speculation

快速LLM后训练通过解耦和最快-N推测

Rongxin Cheng, Kai Zhou, Xingda Wei, Siyuan Liu, Mingcong Han, Mingjing Ai, Yeju Zhou, Baoquan Zhong, Wencong Xiao, Rong Chen, Haibo Chen

机构 * Institute of Parallel and Distributed Systems, Shanghai Jiao Tong University（并行与分布式系统研究所，上海交通大学）； ByteDance Seed（字节跳动种子）

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

AI总结 SpecActor通过解耦和最快-N推测方法提升LLM后训练效率，实现2.0-2.7倍的rollout加速和1.4-2.3倍的端到端训练时间提升。

2512.11998 2025-12-16 cs.CL 90%

Direct Confidence Alignment: Aligning Verbalized Confidence with Internal Confidence In Large Language Models

直接置信对齐：将 verbalized 置信度与内部置信度对齐于大语言模型

Glenn Zhang, Treasure Mayowa, Jason Fan, Yicheng Fu, Aaron Sandoval, Sean O'Brien, Kevin Zhu

机构 * Algoverse AI Research（Algoverse AI研究院）

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

AI总结本文提出直接置信对齐方法，通过优化使LLM的 verbalized 置信度与内部置信度对齐，提升模型透明度和可靠性。

Comments Accepted at ACL 2025 SRW, 5 pages body, 14 pages total

2503.18929 2025-12-04 cs.LG 90%

Trajectory Balance with Asynchrony: Decoupling Exploration and Learning for Fast, Scalable LLM Post-Training

轨迹平衡与异步性：解耦探索与学习以实现快速、可扩展的LLM后训练

Brian Bartoldson, Siddarth Venkatraman, James Diffenderfer, Moksh Jain, Tal Ben-Nun, Seanie Lee, Minsu Kim, Johan Obando-Ceron, Yoshua Bengio, Bhavya Kailkhura

机构 * Lawrence Livermore National Laboratory（劳伦斯利弗莫尔国家实验室）； Mila – Quebec AI Institute（魁北克AI研究院）； Université de Montréal（蒙特利尔大学）； KAIST（韩国科学技术院）； CIFAR Fellow

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

AI总结 TBA通过解耦探索与学习，提升LLM后训练的速度和性能，适用于多种任务并支持大规模数据生成。

Comments NeurIPS 2025; 27 pages

1. 后训练与偏好优化 4535 篇

FinTral: A Family of GPT-4 Level Multimodal Financial Large Language Models

Fundamental Limitations of Alignment in Large Language Models

Group Robust Preference Optimization in Reward-free RLHF

Eureka: Human-Level Reward Design via Coding Large Language Models

Improving Attributed Text Generation of Large Language Models via Preference Learning

Enhancing LLM Safety via Constrained Direct Preference Optimization

Matching Accuracy, Different Geometry: Evolution Strategies vs GRPO in LLM Post-Training

Reward Shaping to Mitigate Reward Hacking in RLHF

BiPACE: Bisimulation-Guided Policy Optimization with Action Counterfactual Estimation for LLM Agents

A Unifying Lens on Reward Uncertainty in RLHF

Mult-DPO: Multinomial Direct Preference Optimization for Recommender Systems

Schedule-Level Shared-Prefix Reuse for LLM RL Training

From Reasoning Chains to Verifiable Subproblems: Curriculum Reinforcement Learning Enables Credit Assignment for LLM Reasoning

Extensive Self-Contrast Enables Feedback-Free Language Model Alignment

Evaluation-Conditioned Training: Teaching Models to Generalize to Stronger Oversight Regimes

Enhancing Large Language Model Reasoning with Reward Models: An Analytical Survey

RIMS: Preference Optimization via Smoothed Multi-pair Aggregation for Small-Scale LLM Retrieval-Augmented Generation

Shapley-based Data Valuation for LLM Alignment via Sequential Preference Optimization

A Technical Survey of Reinforcement Learning Techniques for Large Language Models

KBQA-R1: Reinforcing Large Language Models for Knowledge Base Question Answering

Inference Cost Attacks for Retrieval-Augmented Large Language Models

Training Stratigraphy: Persistent Behavioral Artifacts in Large Language Models Observed Through Longitudinal AI-Human Interaction

MATO: Multi-objective Personalized Alignment with Test-time Optimization for Large Language Models

Emotional intelligence in large language models is fragmented across perception, cognition, and interaction

DGLight: DQN-Guided GRPO Fine-Tuning of Large Language Models for Traffic Signal Control

Clipping-Free Policy Optimization for Large Language Models

Revisiting Parameter Server in LLM Post-Training

Fast LLM Post-training via Decoupled and Fastest-of-N Speculation

Direct Confidence Alignment: Aligning Verbalized Confidence with Internal Confidence In Large Language Models

Trajectory Balance with Asynchrony: Decoupling Exploration and Learning for Fast, Scalable LLM Post-Training