智能体
x.com ·
@omarsar0
Google 发布面向长时程数学研究的多智能体 harness 论文
X 用户转述 Google Research 的一篇论文，提出名为 Stellar Colosseum 的多智能体 harness，用于长时程数学证明：分阶段探索多种证明策略，在 readiness gate 后把路线拆成章节级子问题，并将验证器发现回传到对应章节；阶段内并行生成候选、进行定向证伪并合并批评意见。据该帖称，配合 Gemini 3.1 Pro 与 Gemini 3.7 Flash 在 TCS-Bench（FOCS/STOC/SODA 研究级定理证明任务）上达到 71.0%，带执行反馈时解出 222 道 Codeforces 题中的 218 道。帖中给出 arXiv 编号 2609.15983，具体结论需以论文原文核实。
AI推荐理由 该帖转述 Google Research 一篇关于长时程数学研究多智能体 harness 的论文（Stellar Colosseum），给出了具体机制（分阶段探索、readiness gate、子问题拆分、并行候选生成与定向证伪、验证反馈回传）和量化结果（TCS-Bench 71.0%，Codeforces 222 题解出 218 题），并附 arXiv 编号，信息密度较高。但内容为第三方转述，未提供论文原文细节，指标与结论无法在输入中独立核实，且模型名（Gemini 3.1 Pro / 3.7 Flash）与 arXiv 编号均来自该帖单方陈述，需以原文为准。互动量仅作传播信号，不参与评分。
72
浏览 1,817
点赞 19
评论 5
转发 3
正文
Banger paper from Google on agent harnesses for long-horizon tasks.
(bookmark it)
Google Research built a many-agent harness for long mathematical proofs, and it produced new results on open problems from FOCS and JMLR papers.
Stellar Colosseum works in stages.
It explores several proof strategies, waits for a readiness gate before breaking a route into section-level subproblems, and sends each verifier finding back to the section it affects.
Inside each stage, candidates are generated in parallel, attacked with targeted falsification, and merged together with their critiques.
With Gemini 3.1 Pro and Gemini 3.7 Flash it reaches 71.0% on TCS-Bench, a set of research-level theorem-proving tasks from FOCS, STOC and SODA papers. With execution feedback it solves 218 of 222 Codeforces problems.
Paper: [arxiv.org/abs/2609.15983](https://arxiv.org/abs/2609.15983)
Chat with Paper: [academy.dair.ai/papers/stellar…](https://academy.dair.ai/papers/stellar-colosseum-a-many-agent-harness-for-long-horizon-research-in-mathematics-2609.15983)
