牛哥精选 · 半年

1

📝 深度技术 arXiv AI 2026-06-23

Curriculum Reinforcement Learning Can Incentivize Reasoning Capacity in LLMs Beyond the Base Model

课程强化学习能否突破基础模型推理上限？这项研究给出新答案。

arXiv:2606.22317v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards (RLVR) is widely viewed as a promising path toward co…

强化学习课程学习 llm 推理能力 rlvr

2

🤖 AI·大模型 arXiv 机器学习 2026-06-18

Mechanism-Guided Selective Unlearning for RLVR-Induced Reasoning

提出机制引导的选择性遗忘方法，精准消除RLVR引发的错误推理，保持模型原有能力。

arXiv:2606.19222v1 Announce Type: new Abstract: We propose MAST (Mechanism-Aligned Selective Targeting), a mechanism-guided method for unlearning RLVR…

机制引导选择性遗忘 rlvr 推理偏差大模型鲁棒性

3

📝 深度技术 arXiv 机器学习 2026-06-18

SFT Overtraining Predicts Rank Inversion via Entropy Collapse Under RLVR

ICML 2026 研究揭示：监督微调过度训练会通过熵塌陷导致强化学习下模型输出排名反转，对齐策略需要警惕。

arXiv:2606.18487v1 Announce Type: new Abstract: The standard heuristic of selecting the SFT checkpoint with the highest pass@1 for GRPO can fail when …

sft过训练排名反转熵塌陷 rlvr 大模型对齐

4

📝 深度技术 arXiv NLP 2026-06-12

One Token to Fool LLM-as-a-Judge

只需一个token就能轻松骗过LLM评判者，揭示AI评估体系的安全软肋。

arXiv:2507.08794v3 Announce Type: replace-cross Abstract: Large language models (LLMs) are increasingly trusted as automated judges, assisting evaluat…

ai安全 llm漏洞自动评判单token攻击模型可信度

5

📝 深度技术 arXiv AI 2026-06-12

ReSum: Synergizing LLM Reasoning and Summarization with Reinforcement Learning

新论文ReSum用强化学习协同LLM推理与摘要，解决长推理链低效问题，干货满满。

arXiv:2606.13316v1 Announce Type: new Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) is a central technique for improving long-horizo…

强化学习 llm推理摘要 rlvr 协同优化

6

🤖 AI·大模型 arXiv AI 2026-06-10

TD-Grokking: Learning from Zero-Reward Problems by Training-Time Decomposition

LLM推理训练新突破：通过训练时分解攻克零奖励难题，让模型从失败轨迹中学习！

arXiv:2606.09883v1 Announce Type: cross Abstract: Large language models (LLMs) have made remarkable progress in reasoning tasks, largely driven by pos…

大语言模型推理强化学习零奖励问题训练时间分解

7

📝 深度技术 arXiv AI 2026-06-09

ComplexConstraints and Beyond: Expert Rubrics for RLVR

提出专家评分标准解决RLVR中复杂约束问题，为强化学习奖励设计提供新范式

arXiv:2606.09118v1 Announce Type: new Abstract: As LLM capabilities advance rapidly, the evaluation methods used to assess them increasingly lag behin…

rlvr 专家评分标准复杂约束强化学习奖励设计

8

🤖 AI·大模型 arXiv 计算机视觉 2026-06-09

DyCo-RL: Dynamic Cross-Modal Coordination for Visual Reasoning

突破传统结果优化，动态跨模态协调强化学习提升视觉推理性能

arXiv:2606.08035v1 Announce Type: new Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a leading paradigm for enhancing …

强化学习视觉推理多模态大模型跨模态协调 rlvr

9

📝 深度技术 arXiv 机器学习 2026-06-09

CATPO: Critique-Augmented Tree Policy Optimization

CATPO方法通过批评增强的树策略优化，显著提升大语言模型推理中的密集奖励获取效率。

arXiv:2606.08346v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards (RLVR) has become a dominant paradigm for improving t…

强化学习大语言模型推理能力树策略优化批判增强

10

📝 深度技术 arXiv AI 2026-06-04

GeoMin: Data-Efficient Semi-Supervised RLVR via Geometric Distribution Modeling

针对强化学习验证器数据需求难题，GeoMin用几何分布建模实现高效半监督学习，大幅降低标注成本

arXiv:2606.04516v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards (RLVR) significantly advances LLM reasoning, yet it f…

geomin rlvr 半监督学习几何分布数据高效

11

📝 深度技术 arXiv AI 2026-06-02

CAPF: Guiding Search-Agent Rollouts with Credit-Attenuated Privileged Feedback

提出信用衰减特权反馈方法，引导搜索代理在硬问题上突破成功轨迹稀疏瓶颈

arXiv:2606.01830v1 Announce Type: new Abstract: Recent LLM search agents use reinforcement learning with verifiable rewards (RLVR) to learn search-aug…

llm 搜索代理强化学习 rlvr 信用衰减反馈

12

📝 深度技术 arXiv 机器学习 2026-06-02

RLVR without Ineffective Samples: Group Prioritized Off-Policy Optimization for LLM Reasoning

提出GPOP方法，通过分组优先离轨优化提升LLM推理效率，有效剔除无效样本，为强化学习提供新思路。

arXiv:2606.01281v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards (RLVR) has emerged as a powerful paradigm for enhancing…

rlvr llm推理强化学习离轨优化分组优先

13

📝 深度技术 arXiv AI 2026-05-28

Mechanistically Interpreting the Role of Sample Difficulty in RLVR for LLMs

从因果机制出发，揭秘样本难度如何影响大模型强化学习中的推理表现

arXiv:2605.28388v1 Announce Type: new Abstract: Reinforcement Learning with Verifiable Reward (RLVR) is empirically shown to notably enhance the reaso…

强化学习样本难度机械可解释性大语言模型 rlvr

14

📝 深度技术 arXiv NLP 2026-05-28

Beyond pass@k: Redundancy-Aware RLVR for Multi-Sample Code Generation

提出冗余感知RLVR方法，改进多样本代码生成的评估指标

arXiv:2605.28022v1 Announce Type: new Abstract: LLMs for code generation are commonly evaluated in repeated-sampling settings using Pass@k, where mult…

代码生成多样本冗余感知 rlvr pass@k

15

📝 深度技术 arXiv 机器学习 2026-05-26

Not only where, But when: Temporal Scheduling for RLVR

从强化学习与虚拟现实融合的新视角切入，论文揭示时序调度比空间定位更关键，为RLVR优化提供突破性思路。

arXiv:2605.25381v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards (RLVR) has become a core technique for post-training of…

rlvr 时间调度时序学习强化学习论文推荐

16

📝 深度技术 arXiv NLP 2026-05-22

Linear Dynamics in the RLVR Training of Large Language Models

揭秘大语言模型RLVR训练中的线性动力学机制，为强化学习优化提供新视角。

arXiv:2601.04537v3 Announce Type: replace-cross Abstract: Reinforcement learning with verifiable rewards (RLVR) has driven significant performance gai…

大语言模型 rlvr 线性动力学训练优化强化学习

17

📝 深度技术 arXiv NLP 2026-05-22

Faithful-MR1: Faithful Multimodal Reasoning via Anchoring and Reinforcing Visual Attention

锚定并强化视觉注意力，解决多模态大模型强化学习中的忠实感知难题

arXiv:2605.22072v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards (RLVR) has emerged as a promising paradigm for advancin…

多模态推理强化学习忠实性视觉注意力 rlvr

18

📝 深度技术 arXiv 机器学习 2026-05-21

Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex

将组级别强化学习从奖励建模到策略优化统一为响应单纯形上的目标投影，提出Listwise Policy Optimization新方法。

arXiv:2605.06139v2 Announce Type: replace Abstract: Reinforcement learning with verifiable rewards (RLVR) has become a standard approach for large lan…

listwise策略组基rlvr llm 响应单纯形目标投影

19

📝 深度技术 arXiv 机器学习 2026-05-21

Conformal Selective Acting: Anytime-Valid Risk Control for RLVR-Trained LLMs

新方法保障LLM在线部署每轮风险可控，基于共形预测与RLVR训练，安全认证更可靠。

arXiv:2605.20270v1 Announce Type: new Abstract: A local specialist LLM, fine-tuned with reinforcement learning from verifiable rewards (RLVR) on opera…

llm 安全部署风险控制 rlvr 共形预测

20

📝 深度技术 arXiv 机器学习 2026-05-21

You Only Need Minimal RLVR Training: Extrapolating LLMs via Rank-1 Trajectories

揭示RLVR训练中参数轨迹的秩一结构，仅需极小规模训练即可外推LLM推理能力，颠覆传统认知。

arXiv:2605.21468v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards (RLVR) has become a dominant paradigm for improving rea…

强化学习大语言模型参数轨迹秩1外推可验证奖励

🐂 牛哥精选

Curriculum Reinforcement Learning Can Incentivize Reasoning Capacity in LLMs Beyond the Base Model

Mechanism-Guided Selective Unlearning for RLVR-Induced Reasoning

SFT Overtraining Predicts Rank Inversion via Entropy Collapse Under RLVR

One Token to Fool LLM-as-a-Judge

ReSum: Synergizing LLM Reasoning and Summarization with Reinforcement Learning

TD-Grokking: Learning from Zero-Reward Problems by Training-Time Decomposition

ComplexConstraints and Beyond: Expert Rubrics for RLVR

DyCo-RL: Dynamic Cross-Modal Coordination for Visual Reasoning

CATPO: Critique-Augmented Tree Policy Optimization

GeoMin: Data-Efficient Semi-Supervised RLVR via Geometric Distribution Modeling

CAPF: Guiding Search-Agent Rollouts with Credit-Attenuated Privileged Feedback

RLVR without Ineffective Samples: Group Prioritized Off-Policy Optimization for LLM Reasoning

Mechanistically Interpreting the Role of Sample Difficulty in RLVR for LLMs

Beyond pass@k: Redundancy-Aware RLVR for Multi-Sample Code Generation

Not only where, But when: Temporal Scheduling for RLVR

Linear Dynamics in the RLVR Training of Large Language Models

Faithful-MR1: Faithful Multimodal Reasoning via Anchoring and Reinforcing Visual Attention

Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex

Conformal Selective Acting: Anytime-Valid Risk Control for RLVR-Trained LLMs

You Only Need Minimal RLVR Training: Extrapolating LLMs via Rank-1 Trajectories

📅 日期