Researchers from Stanford and Together AI developed the system by allowing one agent to review past exchanges and rewrite the teamwork strategy, including discussion order and role assignment. The agents learned these cooperation methods from only 15 AIME 2024 problems. Once established, these strategies were applied without further changes to both held-out AIME 2024 tests and four additional benchmarks.
Wednesday, October 21, 2026
Story
Stanford Self Organizing Agent Team Beats Solo AI in Math
AI27 days ago
A group of three Large Language Models recorded an average score of 66.7% on several technical assessments after coordinating their own roles. This coalition, featuring o3-mini, Claude Sonnet 4 and DeepSeek-V3, outperformed the strongest single model's score of 48.8% and the 59.0% achieved by a perfect router that selects the best independent answer. On the AIME 2026 benchmark, the group reached 71.2%, which is 13.4 points higher than the router's performance.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.