Stellar Colosseum: A Many-Agent Harness for Long-Horizon Research
Paper Overview
- Field: NLP
- Authors: Honghao Lin, David P. Woodruff, Yuan Deng, Jieming Mao, Song Zuo, Vahab Mirrokni
- Published: 2026-09-14
- arXiv: 2609.15983
- Explores alternative strategies before proof construction
- Uses a readiness gate to decide when a route is mature enough to decompose
- Represents the proof plan as interdependent section-level subproblems
- Routes verifier findings back to the affected part of the argument
- Generates candidates in parallel, attacks them with targeted falsification, and combines candidates with their critiques into a single research artifact via overlapping stochastic sample trees
- Using Gemini 3.1 Pro, the system achieved several new results on open problems from top-venue papers (FOCS, JMLR)
- On TCS-Bench (research-level theorem proving benchmark from FOCS, STOC, and SODA papers), Colosseum reached 71.0% accuracy with Gemini 3.1 Pro and Gemini 3.7 Flash
- On a Codeforces evaluation with Gemini 3.1 Pro, the proof-oriented pipeline with execution feedback solved 218 of 222 problems
Summary
Language models can produce plausible short proofs, but may still be unreliable on long-horizon research problems, where progress depends on a sequence of uncertain and interdependent decisions. Stellar Colosseum is a model-agnostic harness for allocating inference across research in mathematics and theoretical computer science.
Key mechanisms of the Colosseum workflow:
Integration and Results
The Colosseum workflow is integrated into Google Antigravity's Teamwork framework as a "Long Proof Mode."
Evaluation highlights:
Abstract (original)
Language models can produce plausible short proofs, but may still be unreliable on long-horizon research problems, where progress depends on a sequence of uncertain and interdependent decisions. We introduce Stellar Colosseum, a model-agnostic harness for allocating inference across research in mathematics and theoretical computer science...
--- *Auto-collected on 2026-09-16*