跳转至

文章背景与核心概要

尽管现代大语言模型在生成简短数学证明方面表现出色，但在面对需要长程决策与精密推演的前沿科学研究时，往往容易因链条脆弱和不确定性累积而遭遇瓶颈。为了攻克这一难题，来自学术界与 Google 等机构的研究团队联合推出了 Stellar Colosseum —— 一个与模型无关的多智能体协同与推理竞技框架，专为数学和理论计算机科学 (Theoretical Computer Science, TCS) 等高难度长程科研任务量身打造。该框架创新性地引入了备选策略探索、成熟度门控、结构化证明图、针对性证伪反馈闭环以及重叠随机采样树聚合机制，让多智能体在协同博弈中不断锤炼严密的学术论证。在涵盖 FOCS、STOC 等顶级学术会议的 TCS-Bench 定理证明评测中，Stellar Colosseum 斩获了高达 71.0% 的准确率，并在复杂算法挑战赛中攻克了绝大部分难题，为 AI 辅助前沿科学探索迈出了坚实一步。

Stellar Colosseum：面向数学与理论计算机科学长程研究的多智能体竞技框架

Stellar Colosseum: A Many-Agent Harness for Long-Horizon Research in Mathematics and Theoretical Computer Science

作者： Honghao Lin, David P. Woodruff, Yuan Deng, Jieming Mao, Song Zuo, Vahab Mirrokni
arXiv 预印本： arXiv:2609.15983 [cs.AI] (提交于 2026 年 9 月 14 日，最后修订于 2026 年 9 月 15 日)
主分类： 人工智能 (cs.AI)
其他分类： 计算与语言 (cs.CL)、机器学习 (cs.LG)

Authors: Honghao Lin, David P. Woodruff, Yuan Deng, Jieming Mao, Song Zuo, Vahab Mirrokni
arXiv: arXiv:2609.15983 [cs.AI] (Submitted on 14 Sep 2026, last revised 15 Sep 2026)
Primary Subject: Artificial Intelligence (cs.AI)
Additional Subjects: Computation and Language (cs.CL), Machine Learning (cs.LG)

摘要

Summary

虽然现代语言模型在给出看似严密自洽的简短证明方面表现抢眼，但在面对长程科研难题时却屡屡碰壁。这类长程探索的成败，往往完全取决于一连串环环相扣、充满不确定性且高度脆弱的连锁决策——稍有不慎，推演便前功尽弃。

While modern language models excel at producing plausible short proofs, they frequently stumble on long-horizon research problems where success hinges on a fragile sequence of uncertain and interdependent decisions.

为了突破这一瓶颈，研究团队推出了 Stellar Colosseum —— 一个专为数学与理论计算机科学 (Theoretical Computer Science, TCS) 复杂长程研究打造的模型无关推理框架。

To bridge this gap, the authors introduce Stellar Colosseum, a model-agnostic inference framework designed for complex research in mathematics and theoretical computer science (TCS).

Colosseum 框架的核心创新点：

Key Innovations of the Colosseum Framework:

  1. 备选策略探索 (Alternative Strategy Exploration) ：在正式构建严谨证明之前，先行探索多种潜在的解题路线与突破策略。
  2. 成熟度门控 (Readiness Gates) ：动态评估某条探索路线是否足够成熟，能否进一步拆解为子问题。
  3. 结构化证明方案 (Structured Proof Plans) ：将推演中的证明抽象表示为由相互依赖、章节级子问题构成的网络图谱。
  4. 针对性证伪与反馈闭环 (Targeted Falsification & Feedback Loops) ：并行生成多个候选推论方案，通过激进的对抗性证伪展开极限施压，并将验证器的批评意见精准反馈至论证受影响的具体环节。
  5. 重叠随机采样树聚合 (Overlapping Random-Sample Tree Aggregation) ：将不同路线的候选论断与审查批评深度综合，融汇成结构自洽、高质量的完整学术研究成果。
  1. Alternative Strategy Exploration: Explores multiple solution pathways prior to formal proof construction.
  2. Readiness Gates: Dynamically evaluates whether a particular route is mature enough to be broken down into subproblems.
  3. Structured Proof Plans: Represents the ongoing proof as a network of interdependent, section-level subproblems.
  4. Targeted Falsification & Feedback Loops: Generates parallel candidates, attacks them with aggressive falsification, and routes verifier critiques directly back to the affected areas of the argument.
  5. Overlapping Random-Sample Tree Aggregation: Synthesizes candidates and critiques into a cohesive, high-quality research artifact.

目前，Colosseum 工作流已作为 Long Proof pattern 成功集成至 Google Antigravity 的 Teamwork framework 中。

The Colosseum workflow has been successfully integrated into Google Antigravity's Teamwork framework as the Long Proof pattern.

关键性能与基准测试

Key Performance & Benchmarks

  • 开放式前沿科研 (Open-Ended Research) ：在由 Gemini 3.1 Pro 驱动的 Colosseum 框架加持下，作者针对源自 FOCS 与 JMLR 等顶级学术会议与期刊中论文提出的多个未解开放问题，成功生成了数个全新的解决方案。
  • TCS-Bench 理论计算机科学基准 (TCS-Bench) ：在由 FOCS、STOC 和 SODA 顶会论文提炼出的科研级定理证明任务构成的严苛基准测试中，Colosseum 联合 Gemini 3.1 Pro 与 Gemini 3.7 Flash 实现了 71.0% 的准确率。
  • Codeforces 算法竞赛评测 (Codeforces Evaluation) ：在利用 Gemini 3.1 Pro 并融入执行反馈的面向证明流水线中，该系统在 222 道难题中成功攻克了 218 道。
  • Open-Ended Research: Utilizing Colosseum powered by Gemini 3.1 Pro, the authors generated several novel solutions addressing open problems originating from papers published at premier venues such as FOCS and JMLR.
  • TCS-Bench: On this rigorous benchmark consisting of research-level theorem-proving tasks drawn from FOCS, STOC, and SODA papers, Colosseum achieved 71.0% accuracy using a combination of Gemini 3.1 Pro and Gemini 3.7 Flash.
  • Codeforces Evaluation: In a proof-oriented pipeline enhanced with execution feedback using Gemini 3.1 Pro, the system successfully solved 218 out of 222 problems.