Tag
SketchDirector 是一个用于复杂场景草图生成的智能体框架,通过 LLM 分解文本提示为矢量资产、基于自演化记忆的迭代布局推理以及扩散先验细节优化,并发布了 MCSketch8K 数据集以支持后续研究。
A researcher-curated reading list of 3 essential papers for getting started with agentic RL: Search-R1 for fundamentals, Agent Lightning v1.0 for practical frontier, and a survey on the landscape of agentic RL for LLMs. The post includes the Search-R1 paper, which trains LLMs to autonomously generate search queries during step-by-step reasoning with real-time retrieval.
OA-RAG is an agentic framework that combines formal ontology-based reasoning with LLM reasoning to improve faithfulness in retrieval-augmented generation, reporting grounding levels (entailed, supported, related, unrelated) and achieving top performance on medical QA benchmarks derived from large-scale ontologies.
该论文提出一个基于本体引导的多智能体差分诊断框架,将诊断推理锚定在结构化的医学知识库上,通过显式、可审计的推理步骤减少幻觉并提升可解释性,同时自动生成追问以降低不确定性。
Researchers introduce Solvent, an 8B LLM that extends long-horizon reasoning to multi-step retrosynthesis through a two-round post-training curriculum combining RL and iterative route revision, achieving 61.86% across four benchmarks and rivaling frontier trillion-parameter models.
提出一个基于语料库的多智能体LLM系统通信失败分类体系,基于3211条失败轨迹分析发现失败主要源于恢复/校正与共识/决策问题,而非首轮错误,并建议加强分歧处理与少数信号保留。
该论文分析了结果奖励的强化学习(Outcome RL)对大模型数学与物理推理能力的影响,发现 pass@1 的提升主要来自更快完成回答(finishing)和更高可靠性(reliability),而极少带来真正解决新问题的能力(reach)。
S3AC 提出一种稳定的 Actor-Critic 训练方法与难度感知提示采样(DAPS),在不依赖分组 rollout 的情况下改进大模型的长链推理,性能优于 GRPO-X 等现代基线,并可在 35B 模型上维持单 rollout 训练。
ReAPO 提出一种基于随机最短路径视角的重试感知策略优化方法,对 GRPO 进行改进,将额外生成分配给失败轨迹以实现更有原则的信用分配,在 Qwen3 和 OLMo-2 系列模型的数学与科学推理基准上显著提升首次尝试性能并减少生成 token 数。
该论文提出两阶段诊断评估方法,检验以图结构表示的LLM推理链的结构可靠性,发现图的可靠性随拓扑、提示设计和模型族而变化,且图结构对不确定性估计的校准和节点级诊断有帮助,但作为提示上下文时收益有限。
本文提出了一种面向软思维链(soft CoT)的稀疏松弛强化学习方法 SpaRLax,利用 entmax 自适应稀疏度在不同时间步上传播梯度,并在九个推理基准上取得优于基线的平均性能。
提出 Self-Assessment 强化学习框架(SARL),让检索增强的 LLM 搜索智能体在训练中学会感知自身搜索状态、识别缺失信息,从而提升搜索效率而无需额外推理时评估,在多跳推理与深度研究基准上取得显著相对提升。
A user points out the irony that AI models have solved some of the hardest unsolved math problems while many people still believe AI struggles with basic multiplication, sparked by a complaint about ChatGPT getting multiplication wrong.
The paper proposes the Reliance Hypothesis: LLM chain-of-thought faithfulness increases in lower-resource languages on knowledge-intensive tasks, because models rely more on reasoning when they cannot recall the answer. The effect is replicated across four models and several faithfulness measures, with model confidence partially explaining resource-level effects.
A exploration of using LLM chain-of-thought reasoning to tackle the final step of a Navier-Stokes mathematical argument, blending AI-driven reasoning with open problems in fluid dynamics and the Millennium Problems.
This paper proposes amortizing sequential Bayesian experimental design into LLMs by training the reasoning that guides action selection with information-gain rewards. Trained on Twenty Questions, the resulting policies outperform training- and inference-time BED baselines at lower cost, generalize to unseen categories, and transfer to the Medi-Q clinical dialogue benchmark.
论文提出 LePhilBench 基准,用于评估大语言模型的法律哲学推理能力,包含法理学画像与法理学感知的判决预测两项任务,测试集由 220 份美国最高法院判决构成,并由法律专业标注者生成 9,240 条法律哲学标签。
该论文提出 Librarian Agents 框架,利用强化训练 builder agent 将原始文档重组为可复用的层级化工作区,以服务未来的多样化查询。在 BrowseComp-Plus 上比最强检索与预计算基线高出 33 个百分点,并在 BRIGHT 上零样本泛化提升 31 个百分点。
该论文研究在智能体循环中,System One 模型选择错误动作的原因,发现大多数错误动作是在正确读取状态事实后仍行动错误,而非误解输入,并揭示模型无法组合已知事实来确定下一步动作。
SHELDON 是一个科学推理框架,通过在假设空间和观测空间中进行联合且独立的搜索来提升 LLM 的推理能力,在诊断、抽象推理和优化基准上较基线最高提升 26.7 分,并生成多达 4.5 倍的新颖解决方案。