美团提出 DynaMO：面向RLVR的动态Rollout分配与优势调节策略
论文标题：How to Allocate, How to Learn? Dynamic Rollout Allocation and Advantage
...
你的推理模型其实知道何时停止：解决 Long CoT 中的“过度思考”
论文标题：Does Your Reasoning Model Implicitly Know When to Stop Thinking?
论文链接
...
小红书提出 VESPO 变分序列级软策略优化，从测度变换视角重构重要性采样
论文标题：VESPO: Variational Sequence-Level Soft Policy Optimization for Stable O
...
Qwen 团队新作：统一视角解读 Transformer 中的 Attention 与 Residual Sinks
论文标题：A Unified View of Attention and Residual Sinks: Outlier-Driven Rescalin
...
Less is Enough——在大模型特征空间中合成多样化数据
论文标题：Less is Enough: Synthesizing Diverse Data in Feature Space of LLMs
论文
...
腾讯混元提出 G-OPD：超越教师模型的广义在线蒸馏与奖励外推
论文标题：Learning beyond Teacher: Generalized On-Policy Distillation with Reward
...
腾讯混元提出 Composition-RL：通过合成可验证Prompt提升大模型强化学习效率
论文标题：Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learn
...
上交 & 千问提出 OPUS：论大模型预训练中数据选择与优化器几何的对齐
论文标题：OPUS: Towards Efficient and Principled Data Selection in Large Language
...
大语言模型强化微调中的熵动力学分析
论文标题：On the Entropy Dynamics in Reinforcement Fine-Tuning of Large Language
...
Sea AI Lab 提出 DPPO：重新审视 PPO 算法中的信任域
论文标题：Rethinking the Trust Region in LLM Reinforcement Learning
论文链接：https:
...
