Researchers have introduced Stellar Colosseum, a model-agnostic harness designed to coordinate long-horizon research efforts in mathematics and theoretical computer science. Rather than asking a single model to produce a proof in one pass, the system first explores alternative strategies and uses a readiness gate to decide when a given approach is mature enough to be broken down into interdependent, section-level subproblems.
The harness generates multiple candidate solutions in parallel, subjects them to targeted falsification, and then merges the proposals along with their critiques into a single research artifact through what the authors call overlapping random-sample tree aggregation. Feedback from verification steps is routed back to the specific part of the argument it affects, allowing errors to be corrected locally rather than forcing a full restart of the reasoning process.
The authors note that this workflow has also been integrated into Google Antigravity's Teamwork framework as a pattern called Long Proof, pointing to an ambition to move the method beyond a purely experimental setting. Testing combined open-ended research exploration with quantitative benchmarks: using Gemini 3.1 Pro, the team reports several new results addressing open problems drawn from papers published at venues such as FOCS and in the journal JMLR.
On TCS-Bench, a set of research-level theorem-proving tasks sourced from FOCS, STOC, and SODA papers, Colosseum reaches 71.0% accuracy when combining Gemini 3.1 Pro and Gemini 3.7 Flash. In a separate Codeforces-style evaluation, the proof-oriented pipeline with execution feedback solves 218 out of 222 problems. If these figures hold up under independent scrutiny, they would suggest meaningful progress in getting AI systems to sustain structured, multi-step reasoning — an area where language models have historically struggled once proofs extend beyond a few steps.