reasoning

论文提出 OnQ 流水线，通过 on-policy 蒸馏与候选级 Dynamic KL 恢复低比特量化 LLM 的推理能力，在 Qwen3 系列 2/3 比特量化上取得最优九基准平均成绩，2 比特下 Qwen3-8B 平均提升 10.26 分。

The paper introduces Hierarchical Masked Diffusion Language Models (HMDLM), which incorporate position-level priority hierarchies into masked diffusion language models, achieving lower perplexity on LM1B and WikiHow and improved Sudoku solution accuracy over existing diffusion baselines.

该论文提出 RE-ZS-CoT 框架，通过整合领域专家角色来增强大语言模型的零样本链式推理，并在 20 个 LLM 和 7 个基准上评估，最高提升达 44 个百分点。

KRAFTON AI 发布 Raon-1 mini 预览版，这是一个从零训练的 30B-A4B 混合专家（MoE）语言模型，具备强大的韩语对话与智能体任务能力，支持 262K 长上下文并采用 Apache 2.0 许可。

Perplexity releases pplx-decider-v1.2 and pplx-decider-flash-v1.2 under Apache 2.0, claiming 10x lower cost than GPT 6 Luna (low reasoning) with identical decision quality on the Artificial Analysis Decision Index. The new version is also 2.5x cheaper than Jev.

A father-son duo used the newly released hark.com to semi-autonomously iterate on notebooks for days, achieving a medalling result on the ARC Prize and highlighting how AI can lower the barrier to entry.

论文发现语言模型在组合多个任务时错误率显著上升，且无法完全用独立错误累积解释，而是源于隐式的思考预算分配——组合任务时每个任务获得的推理 token 减少，导致长推理任务更脆弱；通过按脆弱性分组可减少 43% token 用量并保持准确率。

The paper proposes a context-steered reasoning framework for contextual ASR that generates an audio-grounded caption before transcription to assess hotword candidate relevance, using supervised fine-tuning and targeted trajectory injection to strengthen GRPO, improving transcription quality and rare-entity recognition while staying robust to irrelevant candidates.

该论文提出了研究进度图（Research Progress Graph, RPG），一种将研究构想的迭代演化记录为带时序标注的认知状态与推理转移的表示方法，并发现现有35个研究智能体均只是RPG的不完整实现，主张研究智能体应从生成静态构想转向显式建模构想演化轨迹。

EDMA提出了一种基于逻辑熵的多模态问答方法，将答题过程形式化为熵减迭代，并量化各模态对答案的贡献，在多个MMQA基准上显著优于Chain-of-Thought等基线。

This paper theoretically and empirically explains why inference-time harnesses (scaffolds) outperform distilled student models: harnesses can compute closures that are NL-complete even from trivial AC0 primitives, a representational power no fixed-depth forward pass can replicate under standard complexity assumptions. The authors verify that enlarging the model, simplifying the task, or lengthening chain-of-thought can close this gap, including on competition math with agentic harnesses.

该论文评估了单一LLM基线与三种多智能体拓扑在临床任务中的表现，发现协作能提升任务指标，但并不必然降低幻觉，且模型本身而非拓扑结构决定了幻觉的变化方向。

论文提出 Activation-Space Exploration (ASE)，通过聚类正确与错误回答的激活差异构建潜空间探索向量库，以多样化 RLVR 训练中的 rollout；在 Qwen3-1.7B 和 Qwen3-8B 上较 GRPO 分别提升 4.87 和 1.87 个百分点的 Mean@8。

We propose CulturalRiddles, a benchmark of 62 cultures (45 countries, 51 languages) of human-created, culturally grounded riddles requiring abductive and compositional reasoning. Even frontier LLMs reach at most 22% accuracy on some communities, highlighting gaps in cultural understanding.

Agnes AI 发布 agnes-3.0-qwen，一款面向 agentic 编码、工具调用与科学推理的多模态推理模型，采用混合线性/全注意力 MoE 架构（约180B 参数、512 专家），支持 1M token 上下文与可调推理强度，并以开源权重形式发布。

Continuity Research 发布 Tess-4.5-27B，一个基于 Qwen3.8-27B 微调的 27B 开源权重模型，主打本地单卡运行的工程与智能体能力，BenchLocal 成绩与基座持平但推理 token 少 25%，并附带多 token 预测草稿头以提升本地推理速度。

该论文研究扩散语言模型智能体之间能否通过压缩的连续隐藏向量进行因果性通信，发现去噪状态的传递能显著影响接收方推理，且单次晚时传递的效果优于多次重复交换。

Ask-to-Distill 提出了一种优化合成问题的框架，通过教师-学生分布差异作为代理信号，在无需参考答案的情况下提升 on-policy 蒸馏效果，在数学推理、代码生成和智能体任务上优于未优化的问题集和真实问题库。

MMEnterprise-Bench is a new benchmark of 1,613 questions over evolving multimodal workplace records, evaluating enterprise agents on evidence reasoning, conflict resolution, and unanswerability. Evaluations of 10 proprietary and open-source multimodal LLMs show that even the strongest models struggle with cross-source and temporal conflicts.

PAR (Progress-Aligned Recurrence) aligns each loop iteration of a looped language model with a step of the task and rescales hidden states to prevent drift, enabling 16× length extrapolation on swap-S5 state tracking with a 0.5B model reaching 95.5% accuracy, far surpassing RLT and CoT baselines.