Stanford agent teams beat oracle router by 13.4 points on AIME

On a suite of five mathematics and physics benchmarks, AI agent teams that learned their own collaboration structure averaged 66.7% accuracy, versus 48.8% for the strongest individual member and 59.0% for a "perfect router" that always picks the best member's independent answer, according to a Stanford-led paper on arXiv.

On AIME 2026 specifically, the self-organizing teams beat the oracle router by 13.4 percentage points. Lead author Aneesh Pappu and colleagues call the mechanism "collaborative computation": "agents exchange, challenge, repair, and synthesize partial reasoning into solutions no member produced independently."

The teamwork strategies transferred unchanged to unseen benchmarks after training on just 15 mathematics and 25 graduate-level knowledge problems. Improvements did not hold everywhere. Across eight benchmarks the paper reports a Spearman ρ=0.90 (p=0.005) correlation between gains over the strongest member and "demonstrability," the organizational-psychology construct of whether a team can recognize correct reasoning once it appears.

"These results suggest that organization itself can become an agent capability," the authors write, arguing that agent teams can "produce solutions their members could not reach independently." The paper was accepted as a poster at the Context Beyond the Window workshop at COLM 2026 and the DocInsights workshop at EMNLP 2026.