Stellar Colosseum: A Many-Agent Harness for Long-Ho... | AI Research

Stellar Colosseum: A Many-Agent Harness for Long-Horizon Research in Mathematics and Theoretical Computer Science introduces a new framework designed to help AI models tackle complex, multi-step research problems. While language models are often capable of solving short, isolated tasks, they frequently struggle with long-horizon research where a single error can propagate through a long chain of interdependent logical steps. Colosseum addresses this by organizing the research process into a structured pipeline that manages strategy, decomposition, and verification, allowing the system to explore multiple paths and repair specific parts of an argument without discarding the entire effort.

A Structured Research Pipeline

Colosseum moves beyond simple prompt-response cycles by breaking research into distinct stages: strategy exploration, proof decomposition, subproblem solving, and global verification. It begins by exploring various potential strategies, using a "readiness gate" to determine when a path is stable enough to be broken down into a formal plan. Once a strategy is chosen, the system decomposes it into a series of interdependent subproblems. These sections can be solved in parallel, and if a specific section fails, the system can attempt to repair it locally rather than restarting the entire proof from scratch.

Adversarial Inference and Aggregation

To ensure the quality of its work, Colosseum employs an adversarial approach to inference. Within each stage, the system generates a population of candidate solutions and subjects them to targeted falsification—a process where reviewers specifically hunt for counterexamples, logical gaps, or invalid assumptions. These candidates and their critiques are then combined using an "overlapping random-sample tree aggregation" method. This technique allows the system to synthesize the best ideas from multiple candidates while keeping track of unresolved objections, ensuring that evidence of failure is not lost during the aggregation process. To see google in practice, Gemini's now Generates Files! walks through a concrete example.

Performance and Results

The researchers tested Colosseum on challenging benchmarks in theorem-proving and competitive programming. Using Gemini 3.1 Pro and Gemini 3.7 Flash, the system achieved 71.0% accuracy on TCS-Bench, a collection of research-level tasks drawn from top-tier computer science venues. In a separate evaluation using competitive programming problems from Codeforces, the system successfully solved 218 out of 222 problems. The framework has also been integrated into Google Antigravity’s Teamwork framework as the "Long Proof" pattern, demonstrating its utility in practical, open-ended research environments.

Managing Research Knowledge

A key feature of Colosseum is its ability to maintain a "shared research state." When a proof attempt is rejected, the system does not simply discard the work. Instead, it retains the draft and the specific feedback provided by the verifier, passing this information into the next round of development. By keeping a record of both the current strategy and a directory of reusable findings, the system ensures that previous insights—and the lessons learned from failed attempts—remain available to guide future iterations of the research. The ai agents story also surfaces in Google AI Introduces EnvHarness for Adaptive..., adding another angle. as detailed in the full paper on Arxiv