Stanford and Together AI researchers argue that collective intelligence depends not only on what team members individually know, but on how they organize their work. Their paper, *Self-Organizing Agent Teams Learn to Reason Together* (arXiv:2609.22682, submitted September 19, accepted as a poster at COLM 2026 and EMNLP 2026 workshops), proposes SAT (Self-Organizing Agent Teams): a fixed team of agents is given reusable strategies learned from prior collaboration to organize roles, conversation phases, participation, and information flow.
The authors call this mechanism collaborative computation—agents exchange, challenge, repair, and synthesize each other's partial reasoning, producing solutions no member could reach alone.
Experimental Setup and Headline Results
The team trained on just 40 problems—15 math and 25 graduate-level knowledge questions—then applied the learned strategies unchanged to unseen benchmarks.
Average accuracy across five math and physics benchmarks:
| Baseline | Average Accuracy | |---|---| | Self-organizing team (SAT) | 66.7% | | Strongest member alone | 48.8% | | Same member with matched compute budget | 58.7% | | Perfect routing over independent member answers | 59.0% |
Two comparisons matter here:
- Compute-matched baseline (58.7%): giving the strongest member extra compute explains part of the gap, but SAT still leads by 8 points.
- Perfect routing (59.0%): this oracle picks a correct answer whenever any member produced one independently—a ceiling for independent answering. SAT exceeds it by 7.7 points overall and by 13.4 points on AIME, evidence that discussion produces answers no member reached alone.
- The authors note the results await independent replication.
- Gains vary widely across benchmarks; 66.7% is a five-benchmark average, not a uniform improvement.
- The 40-problem practice set is small (a selling point), but learned strategies may be sensitive to its composition.
- The perfect-routing baseline is defined over independent answers; it does not cover cases where members answer correctly only after being prompted during discussion—the very case collaborative computation targets—so 59.0% is a ceiling only for independent answering.
- Aneesh Pappu, Mirac Suzgun, Yongchan Kwon, Federico Bianchi, Batu El, Mykel J. Kochenderfer, Hancheng Cao, James Zou, *Self-Organizing Agent Teams Learn to Reason Together*, arXiv:2609.22682, submitted 2026-09-19
- Benchmark figures: 66.7% vs 48.8%, compute-matched 58.7%, perfect routing 59.0%, +13.4 points on AIME
- Demonstrability correlation: Spearman 0.90, p = 0.005 across eight benchmarks
When Does Teaming Up Pay Off?
Gains vary substantially across benchmarks, which the authors turn into a second research question. The answer lies in an organizational psychology construct called demonstrability—whether a team can distinguish correct from incorrect reasoning. Across eight benchmarks, demonstrability correlates strongly with improvement over the strongest member (Spearman 0.90, p = 0.005).
Practically, this turns the decision to use multi-agent systems from an article of faith into a measurable quantity: test whether members can recognize correct reasoning before paying the compute cost of collaboration.
The Broader Claim
The paper's closing argument: organization itself can be a learnable agent capability. Where recent progress came from training larger single models or hand-crafting orchestration protocols, learned organization offers a cheaper, more general path—keep the members, change how they talk to each other.