Tag
该论文用信号处理工具审计了 DeepConf 等基于 token 置信度的推理链筛选方法，发现滑动窗口最小置信度分数主要度量的是链的长度，剔除长度因素后该分数对正确性几乎不携带额外信息，按长度排序效果相同甚至更好。
本文提出条件对比外推在线蒸馏（CCE-OPD），通过让教师模型对完整问题与掩码条件版本进行对比评估，降低教师在学生生成前缀上偏好无效续写的风险，在六个数学基准上相较标准OPD提升avg@8约1.60-1.95个百分点。
The paper introduces TIPS (Thinking-Induced Process Supervision), an outcome-only reinforcement learning framework that trains generative process reward models using only outcome-level rewards, achieving 85.2 F1 on ProcessBench with just 3.2K outcome-labeled trajectories and outperforming existing trained PRMs and prompt-only judges like GPT-5.4-Instruct.
The paper introduces Self-Play Search Distillation (SPSD), a framework that uses self-play of MuZero-like networks to generate superhuman synthetic data for training LLMs, significantly improving reasoning performance on mathematics benchmarks.
This paper uses a causal mediation framework to audit thinking leakage in hybrid reasoning models, revealing that post-training gains in NoThink mode substantially rely on invoking existing Think behavior, with leakage ratios between 42% and 79%.
This paper introduces Self-Play Search Distillation (SPSD), a framework that uses self-play in board games to generate synthetic data for improving large language model reasoning, with demonstrated improvements on mathematical benchmarks.
The paper proposes Informed Masking (IM), a structure-aware perturbation method for reinforcement learning in diffusion large language models that prioritizes downstream tokens, improving performance on math and planning benchmarks.
When2Think is a post-training framework that dynamically allocates computation in large reasoning models based on problem difficulty, improving accuracy-efficiency trade-offs on mathematical benchmarks.
The paper identifies the Matthew Effect in reinforcement learning for large language models, where easy problems improve more than hard ones, and introduces Never Give Up (NGU), an adaptive sampling method to allocate more compute to hard problems, demonstrating performance gains on math and coding benchmarks.
This paper introduces Lightning Weave, a framework that combines separately trained reasoning capabilities into one efficient model using on-policy distillation, enhancing accuracy and reducing token usage in math and code tasks.
This paper proposes Concretized Proposition Prompting (CPP), a framework that resolves the composition-knowledge dichotomy in LLMs by explicitly concretizing propositions relevant to questions, significantly enhancing reasoning performance especially in medical and math benchmarks.
This paper introduces DASH, a method that uses intermediate answer commitments within reasoning traces to assign segment-level credit, reducing overthinking behaviors and improving accuracy on competition-level math benchmarks.
This paper proposes Trajectory-Augmented Policy Optimization (TAPO), which constructs micro-reflective correction trajectories using the model's own correct and incorrect rollouts to improve reasoning in large language models, outperforming standard self-distillation methods on math benchmarks.
This paper formalizes reasoning redundancy in LLMs as the fraction of trailing steps that can be truncated without affecting correctness, quantifying 61-93% redundancy across frontier models and proving that redundancy is a structural consequence of length-agnostic outcome rewards.
Single-stream Policy Optimization (SPO) is a new policy-gradient training method for Large Language Models that eliminates group-based issues like degenerate groups and synchronization barriers, providing a stable, low-variance learning signal and improving performance on hard math benchmarks compared to GRPO.