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

Details

Stellar Colosseum is a model-agnostic orchestration harness from Google that allocates inference across long-horizon mathematical and theoretical computer science research tasks. The system addresses several known open problems from top venues (FOCS, JMLR) and achieves 71% accuracy on TCS-Bench, a benchmark of research-level theorem-proving tasks drawn from FOCS, STOC, and SODA papers. In a Codeforces evaluation, it solves 218 of 222 competitive programming problems using a proof-oriented pipeline with execution feedback. The architecture is distinctive: before constructing a proof, Colosseum explores alternative strategies and uses a readiness gate to decide when a route is mature enough to decompose into interdependent section-level subproblems. It generates candidates in parallel, attacks them with targeted falsification, and aggregates outputs via overlapping random-sample tree aggregation. Verifier findings are routed back to affected parts of the argument, enabling self-correction at the section level. The results mark a qualitative shift: AI is solving not just graduate-level problems but actual open research problems — tasks that remained unsolved at publication time in premier venues. The workflow has been integrated into Google Antigravity's Teamwork framework as the Long Proof pattern. Gemini 3.7 Flash is also used alongside 3.1 Pro for the TCS-Bench evaluation.