Stellar Colosseum agent hits 71% on TCS-Bench, 218/222 Codeforces

Stellar Colosseum, described by its authors as a "model-agnostic harness for allocating inference across research in mathematics and theoretical computer science," reaches 71.0% accuracy on TCS-Bench when paired with Gemini 3.1 Pro and Gemini 3.7 Flash, according to a preprint posted to arXiv on September 14 by Honghao Lin, David P. Woodruff, Yuan Deng, Jieming Mao, Song Zuo, and Vahab Mirrokni. TCS-Bench, the paper says, is a set of "research-level theorem-proving tasks drawn from papers published at FOCS, STOC, and SODA."

The harness "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, and routes verifier findings back to the affected part of the argument." Candidate proofs are generated in parallel, attacked with "targeted falsification," and merged through "overlapping random-sample tree aggregation." Two researchers on our tracker were passing the preprint around the same day it landed.

On a separate Codeforces evaluation with Gemini 3.1 Pro, the paper reports that "the proof-oriented pipeline with execution feedback solves 218 of 222 problems." The workflow, the authors write, "has been integrated into Google Antigravity's Teamwork framework as the Long Proof pattern."

The authors also claim "several new results that address open problems arising from papers published at top venues such as FOCS and JMLR" from running Colosseum with Gemini 3.1 Pro. Those results ride on an unrefereed preprint; the arXiv abstract does not enumerate which open problems were resolved or name the target papers.