Tag
OpenAI 发布了一个新仓库，收录其未公开内部模型产生的 719 篇数学手稿（归为 372 个结果族），涵盖黎曼 zeta 函数零点区域、Mahler 猜想、自由群因子同构等前沿问题，其中约 42% 的结果已用 Lean 形式化验证，并附带部分推理过程摘要。
ReCal 是一种即插即用的校准方法，在结构化剪枝前通过教师-剪枝探针间的前向 KL 识别受损预测并重加权校准统计，从而提升剪枝模型在在线策略蒸馏（OPD）后的恢复效果，在 AIME 上最高提升 16.7 个百分点。
OpenAI 发布了一个包含由其未公开内部模型产出的 719 篇数学手稿（372 个结果族）的仓库，并附带部分 Lean 形式化证明与推理摘要，同时据标题消息撤回了其中 3 篇论文。
该论文研究了推理模型在不同思考预算下是否表现出更少的认知偏误，结果发现增加推理token并不会降低偏误，部分偏误甚至与人类方向相反，表明测试时推理不能保证决策理性。
A speculative post hypothesizes that U.S. fears of model distillation by China are causing labs like OpenAI and Anthropic to intentionally obscure model reasoning, worsening alignment visibility and creating black-box systems even their owners can't interrogate.
DFlash 2, a lightweight block-diffusion model for flash speculative decoding from z-lab/haoailab, enabling efficient high-quality parallel drafting across a wide range of open LLM families with checkpoints released on Hugging Face.
Simon Willison ran a controlled benchmark on local Qwen3.8-27B (Q4_K_M) testing exact addition of up to 13-digit integers output solely in English words, replicating an older GPT-4o experiment. With reasoning disabled the model achieved only 23.57% overall accuracy, collapsing from ~97% on 1–3 digit operands to ~6% on 10–13 digit operands despite 96.17% format compliance.
A tweet satirizing AI safety discourse around losing chain-of-thought monitorability due to opaque internal reasoning, riffing on a quote attributed to Miles Brundage.
Sebastian Raschka's sixth 'Reasoning from Scratch' video introduces and implements Reinforcement Learning with Verifiable Rewards (RLVR) and Group Relative Policy Optimization (GRPO), covering reward design, advantage computation, GRPO loss, and MATH-500 training results.
The paper identifies why self-evolving reasoning models suffer performance collapse — invalid questions and mathematically-equivalent question repetition — and proposes R-Quest, which uses validity and novelty feedback to sustain self-evolution, outperforming R-Zero by 17.32 points across 12 benchmarks over ten rounds.
A community discussion argues that newer open-source models are degrading in non-thinking/instruct mode compared to older releases like Claude 3.6 versus 3.8, and asks labs to keep prioritizing strong instruct performance alongside reasoning capabilities.
Yohei Nakajima muses that decision models are essentially derived by stripping out the reasoning chain, which seems to assume decisions precede reasoning — questioning whether that mirrors how reasoning actually works.
This paper analyzes why pointwise clipping of the forward-KL objective in on-policy self-distillation (OPSD) can backfire: it proves and empirically shows that clipped objectives can push token probabilities away from the teacher, causing persistent token repetitions in reasoning outputs.
This paper introduces Tokenwise Residual Comparison (TRC), a method that localizes repetition anomalies from residual stream dynamics during generation in LVLMs/LLMs and selectively suppresses them, reducing loop rates by 57% on average, while showing that repetition semantics emerge in shallow layers before propagating deeper.
StreamDecisionBench (SDB) evaluates language models as real-time decision components inside applications, measuring the 'decision in force' over evolving evidence streams and attributing errors to judgment, latency, or both. Across four hosted models, it finds that reasoning improves judgment but costs substantial latency, with one faster, weaker component achieving an integrated score comparable to a stronger non-reasoning model.
The paper presents OpenJev-RLCD, a working open-source implementation of Reinforcement Learning for Calibrated Decisions (RLCD) for reasoning models, using a 'calibrate, then reinforce' recipe with proper scoring rules that matches or beats SFT, RFT/STaR and GRPO with temperature scaling, especially in selective prediction (e.g., 81% of GSM8K items decided at ≤5% error versus 19% for GRPO).
This paper introduces RADAR, a method that models large reasoning model generation as four states and uses dynamic attention responses to detect uncontrolled reasoning before repetition begins. It proposes Attention Realignment to correct abnormal attention distributions, reducing generation loops while preserving benign performance.
A Wayne State University paper introduces DSAR, a metric quantifying inconsistency between reasoning traces and final answers in Large Reasoning Models, and proposes SARA, an RL-based method rewarding safety-aware reasoning to mitigate deceptive safety alignment under standard and adversarial prefilling attacks.
The study reveals that thinking in reasoning language models has an asymmetric dual effect on counterfactual fairness: it resolves some biases but creates more, with the latter outnumbering the former by about 5× across models and datasets.
The paper proposes a lightweight, rule-based selector for diverse SFT traces to improve post-RL generalization in reasoning models, showing significant performance gains on mathematical benchmarks.