Stanford study finds self-organizing AI agent teams outperform oracle routers on math benchmarks

A Stanford-led study demonstrates that AI agent teams capable of learning their own collaboration structure significantly outperform both individual agents and optimal routing strategies. On a suite of five mathematics and physics benchmarks, these self-organizing teams achieved an average accuracy of 66.7%, compared to 48.8% for the strongest individual member and 59.0% for an oracle router that always selects the best member's independent answer.

Collaborative Computation Mechanism

The researchers attribute this performance gain to a mechanism they term collaborative computation, where agents exchange, challenge, repair, and synthesize partial reasoning steps to generate solutions that no single member could produce alone. This suggests that the organizational structure of the agent team itself acts as a distinct capability, rather than just a method for aggregating existing skills.

Performance on AIME 2026

On the specific AIME 2026 benchmark, the self-organizing teams exceeded the oracle router's performance by 13.4 percentage points. The paper, led by Aneesh Pappu, has been accepted as a poster at the Context Beyond the Window workshop at COLM 2026 and the DocInsights workshop at EMNLP 2026.