Google's Stellar Colosseum: many-agent harness proves new math theorems
IgorCarron · x · 2026-09-18
Google (incl. Vahab Mirrokni et al.) published "Stellar Colosseum: A Many-Agent Harness for Long-Horizon Research in Mathematics and Theoretical Computer Science" (arXiv:2609.15983), already integrated into Antigravity's Teamwork framework as the Long Proof pattern (/teamwork-preview).
- Motivation: LLMs produce plausible short proofs but remain unreliable on long-horizon problems depending on chains of uncertain, interdependent decisions.
- Architecture: a model-agnostic harness separating strategy exploration → a readiness gate deciding when a route is mature enough to decompose → proof plans as interdependent section-level subproblems → verifier findings routed back to the affected part; stages generate candidates in parallel, attack them with targeted falsification, and merge candidates plus critiques via overlapping random-sample tree aggregation.
- Results: with Gemini 3.1 Pro it obtained several new results on open problems, evaluated on theorem-proving and competitive programming benchmarks.
- Igor Carron's quip: traditional researchers 'will fall because of strategies like these.'
More from coding & agent
- Anthropic engineer: the longer your agent runs, the costlier its mistakes — build verification in — Roger_M_Taylor · 2026-09-18
- Dev builds LLM-free chatbot wired to tools, sparking debate on what LLMs add beyond speed — jasonkneen · 2026-09-18
- claude-code-best-practice: a GitHub index mapping Claude Code concepts to real setups — Roger_M_Taylor · 2026-09-18
- Free 1-hour Grok Bot agents course: from one prompt to a 24/7 autonomous system — Roger_M_Taylor · 2026-09-18
- Fine-tune 500+ open source models for free in Google Colab with Unsloth Studio — Roger_M_Taylor · 2026-09-18
- Loops vs. Graphs: the loop lives inside a node, the graph lives between them — Roger_M_Taylor · 2026-09-18