A new AI system just helped researchers chip away at real unsolved problems in math and computer science, not just benchmark trivia.
Researchers built Stellar Colosseum, a harness that coordinates many instances of a language model to work through long, multi-step proofs instead of asking one model to write an entire proof in one pass. It explores several strategies before committing to one, uses a "readiness gate" to judge when an approach is solid enough to break into subproblems, and routes any errors a verifier flags back to the specific section of the proof that caused them. It's now built into Google Antigravity's Teamwork framework as the "Long Proof" pattern. Paired with Gemini 3.1 Pro, the system produced new results addressing open problems raised in papers published at FOCS and JMLR, two of the more selective venues in theoretical computer science and machine learning.
This is a step beyond AI systems that ace exam-style problems but stall once progress depends on a chain of earlier, uncertain decisions, which is closer to how real research actually works. The system's own numbers show where that gap still sits: 71.0% accuracy on TCS-Bench's research-level theorem-proving tasks, versus 218 of 222 solved in a Codeforces evaluation with execution feedback.
That gap is the tell. Competitive programming problems get scored automatically and instantly; open research problems don't, which is exactly why the harness leans so hard on internal verifiers and falsification passes to make up the difference.