Tag
提出SOAR（Source-Aware Reinforcement Learning），通过将外部证据与内部知识解耦为可观测的并行流并引入Source-Consistent Reward，实现搜索代理的精确信用分配；在3B和7B模型上分别提升7.3%和5.4%，并将证据伪造率降低74.8%。
提出一种协作学习框架，用多个异构小模型替换单个大模型进行推理（model slicing），并通过金字塔监督蒸馏改进跨模型协作，在多跳问答、数学推理、代码生成等任务上达到与大模型相当或更优的性能，且推理效率更高。
LongR 是一个统一框架，通过将“Think-and-Read”动态机制与基于相对信息增益的上下文密度奖励相结合，提升 LLM 的长上下文推理能力，在 LongBench v2 上取得超过 9% 的提升，并在 RULER 和 InfiniteBench 上保持一致改进。
该论文提出 Tracking-Value PPO（TPPO），通过在对数几率空间对前缀值进行贝叶斯分解，构建累积值参数化以更好地跟踪语言模型推理中的值演化。TPPO 在 Qwen3-4B-Base 和 Qwen3-8B-Base 上分别比 PPO 提升 7.0 和 8.4 个数学推理分数点，并改善了分布外泛化能力。
CARing 框架通过组合式语义ID和覆盖感知的强化学习奖励，优化大语言模型在多标签下次就诊诊断预测中的推理轨迹， 在 MIMIC-III/IV 数据集上超越所有 EHR 训练基线。
本文提出使用可能性理论将开放式科学发现形式化为自主科学发现（AASD）问题，引入abductive utility作为发现进展的可计算度量，并提出possibility frontier search算法，在保持anytime有效性的同时渐近达到ε-最优abductive utility。
该论文提出 Language-Aware Skeleton Exploration Framework (LASEF)，系统研究多层数学推理中骨架提示语言的选择问题，发现英语骨架平均略有优势但并非普遍最优，骨架语言是一个依赖上下文的设计变量。
This paper proposes Trajectory-Local Adaptive Retrieval (TLAR), a speculative decoding method that retrieves matched continuations from the model's own generation history and combines them with model-generated drafts in a shared candidate tree, showing improved token acceptance and end-to-end throughput without changing the target model's output distribution.
A practitioner discussion exploring whether long-running AI agent failures stem from model capabilities or from the scaffolding around them, highlighting error compounding, context pollution, and weak self-correction as key failure modes.
HyperThink introduces a text-to-parameter approach where a lightweight hypernetwork predicts query-conditioned updates to a small subset of an LLM's parameters, amortizing long thinking traces into a single forward pass to improve the accuracy-latency tradeoff for reasoning tasks.
This paper introduces Trained Agentic Context Management, where a small model (Qwen3.6-35B-A3B) is finetuned over a minimal agentic harness — a tool to call itself and a tool to read arbitrary token ranges — allowing it to match GPT-5.4's 1M-token context performance on long-document benchmarks using only 8K tokens of context.
The paper introduces Success-Referenced On-Policy Distillation (SR-OPD), which uses a student's own successful rollouts as references to selectively allocate teacher supervision, achieving comparable reasoning performance on math benchmarks using only 3.46–5.02% of the teacher-input tokens required by vanilla on-policy distillation.
Meta AI researchers partnered with mathematicians to use Muse Spark 1.1/1.2 in Thinking Mode to tackle open research problems, releasing six papers — five of which present solutions to previously open mathematical questions. The collaborations followed strict attribution and human-review principles, with acknowledgment of independent concurrent solutions from other teams.
Google Research presents Cogentic, a multi-agent harness built on Gemini that runs an iterative prove–verify loop to discover proofs for open problems in theoretical computer science, producing expert-verified novel results on five open problems in online learning, auction theory, and mechanism design.
MIT Technology Review's daily newsletter covers an opinion piece by ex-DeepMind researcher Thore Graepel arguing LLMs don't truly reason and proposing AlphaGo-inspired approaches, a biological-age de-aging contest, and OpenAI's report that rogue agents may have affected over 100 organizations.
Learning is most effective at the frontier of capability: for GRPO-trained LLM reasoners, problems that are always or never solved yield zero gradient. The authors introduce "Frontier Learning" to target this effective training regime.
MIT Technology Review argues that unlike AlphaGo, which combined intuitive networks with explicit game-tree search to reason about future consequences, today's LLMs only perform fast, System-1-style next-token pattern completion—and that genuine deliberative reasoning must be built in for AI to be trustworthy in science and medicine.
The paper provides theoretical guarantees for beam-search-based test-time scaling in LLM reasoning, proving a lower bound on vanilla beam search sample complexity and proposing confidence-filtered beam search (CF-Beam) that reduces coverage dependence from quadratic to nearly linear, outperforming sequence-level methods like Best-of-N on long-horizon tasks.
The paper introduces Hermes, a family of configurable harnesses that let models control context allocation across inference-time compute, and Hermes-Learn, a two-stage framework that trains models to make these contextual reasoning decisions adaptively, closing the gap for smaller open-source models.
本文将多路径 LLM 推理形式化为无线通信中的分集合并问题，分析路径正确性相关性对多数投票有效样本量的上限，并在 5 个模型和 12 个基准上验证提示模板分集可降低路径相关性，同时提出基于四路径试点的 Adaptive-K 规则来选择最优采样数。