今日速览
| 序号 | 标题 | 来源 | 日期 | 主题 | 推荐等级 |
|---|---|---|---|---|---|
| 1 | NeutronGym: Physics-Graded Neutron Instrument Design for LLM Agents | arXiv | 2026-10-02 | AI-Agent | 高 |
| 2 | HazardWeaver: Scientific Route Selection for Hazard Analysis Agents | arXiv | 2026-10-02 | RAG | 高 |
重点论文与技术动态
1. NeutronGym: Physics-Graded Neutron Instrument Design for LLM Agents
- 来源:arXiv
- 日期:2026-10-02
- 作者/机构:Lijie Ding, Changwoo Do
- 主题标签:
AI-Agent,arXiv - 推荐等级:高
- 分类:cs.AI, physics.ins-det
一句话结论
NeutronGym是首个可执行中子仪器设计环境,LLM智能体通过验证工具构建仪器,McStas射线追踪,分级梯子无LLM评判评估语法、运行时、结构和科学,程序化家族提供无限实例,并可RL训练。
核心内容
- 智能体通过验证工具构建仪器,McStas射线追踪,分级梯子无LLM评判评估语法运行时结构科学。
- 程序化家族无限固定布局实例,智能体设参数含held-out;McStasBench 16已发表任务带探针沙盒。
- 七模型最多复现7/16,无检索参考无达目标。RL使Qwen3-8B 11%至77%(69%另一种子),超Qwen3-32B,配方适用于三家族。
方法与数据
- 使用验证工具、McStas射线追踪和分级梯子;McStasBench含16任务,评估七模型,对Qwen3-8B进行RL训练。
- 摘要未明确。
价值判断
- 值得关注:首个环境,RL提升小模型超大未训模型,分析梯子部分信用关键(无则崩60点)。
- 可复用点:分级梯子、程序化家族、记忆探针、发现失败任务的探针。
- 局限/待核查:训练模型仅给闭式物理时达经典优化器(77%对81%);前沿模型98-99%;四个任务设计失败,释放探针。
摘要
Designing a scientific instrument tests whether language-model agents can do physics rather than recall it, provided the grading cannot be argued with. We introduce NeutronGym, to our knowledge the first executable environment for neutron instrument design: agents build instruments through validating tools, McStas ray-traces what they build, and a level-resolved ladder grades syntax, runtime, structure and science with no LLM judge. Procedural families supply unlimited instances of a fixed layout whose design parameters the agent must set, with held-out parameter regimes; a curated slice, McStasBench, adds 16 tasks from published instruments behind memorization probes and a sandbox. Seven models reproduce at most 7 of the 16, none retrieves a reference, and none meets an improvement target. The environment also trains. Reinforcement learning on its reward takes Qwen3-8B from 11% to 77% of held-out instances of a family whose targets come from a hidden design (69% at a second seed), past an untrained Qwen3-32B, and the recipe holds, at one seed each, on three further gated families. The analysis says what that gain is. Without the ladder's partial credit it collapses by 60 points. From reward alone the trained model reaches what a classical optimizer reaches, at the agent's simulation budget, only when handed the closed-form physics (77% against 81%, a gap that does not separate at this size), while frontier models still solve 98-99%. Getting a trustworthy result meant failing four task designs that no-model baselines could solve, and we release the probes that found them.2. HazardWeaver: Scientific Route Selection for Hazard Analysis Agents
- 来源:arXiv
- 日期:2026-10-02
- 作者/机构:Wangshu Zhu, Xueqi Cheng, Liang Wu, Yushun Dong
- 主题标签:
RAG,arXiv - 推荐等级:高
- 分类:cs.AI
一句话结论
HazardWeaver将自然灾害分析中科学方法选择形式化为状态依赖路径选择,使AI代理能根据事件、数据和工具动态确定适用且可执行的工作流,并随新证据修订决策。
核心内容
- 有效自动化要求代理判断科学方法对给定事件的适用性及与可用数据和工具的可执行性,条件会随新证据和执行结果变化。
- 系统先用Hazard Knowledge Compiler提取证据链接的适用性条件,再用Hazard Capability Graph表示可执行科学能力并检查输入输出兼容性。
- Hazard Weaver Agent据此选择适用可执行路径、执行工作流,并随分析状态变化修订决策。
方法与数据
- 引入Hazard Weaver Benchmark,含141个实例,覆盖七个单灾害领域和四个多灾害交互类别,允许多种有效科学路径,评估输出正确性、路径有效性和合理弃权。
价值判断
- 值得关注:实验显示HazardWeaver优于现有代理系统,最大增益在具有多个合格科学路径的任务上。
- 可复用点:代码公开可用。
- 局限/待核查:摘要未明确具体比较系统和详细数值结果。