Self-Organizing Agent Teams Learn to Reason Together

Self-Organizing Agent Teams Learn to Reason Together. By Aneesh Pappu and seven others, from Stanford University, Together AI, and one other institution. Published on the nineteenth of September. No code repository is linked to it.

Collective intelligence depends not only on what the members of a team know, but on how they organize their work. When the structure of a solution is unknown ahead of time, nobody can assign useful roles or divide the labor in advance. A team has to learn from experience how to organize its reasoning while that reasoning is unfolding. Human teams adapt this way routinely. Existing teams of AI agents do not. They rely on fixed protocols, on explicit task decomposition, or on routing a question to one member, and the organization is settled before the problem is ever seen. The difficulty is not a lack of skill in any member. However capable each one is, a plan fixed in advance cannot fit a solution whose shape nobody knows yet.

Self-Organizing Agent Teams, abbreviated SAT, keep the team itself fixed and let its way of working be learned. The agents learn reusable strategies from prior collaborations, and those strategies govern four things: the roles members take, the conversational phases a discussion moves through, who participates, and how information flows between them. What the strategies enable, the authors call collaborative computation. Agents exchange partial reasoning, challenge it, repair it, and synthesize it into solutions that no single member produced on its own.

The learning needs very little material. In two independent settings, the teamwork strategies are learned from only fifteen mathematics problems and twenty-five graduate-level knowledge problems. The strategies then transfer unchanged to benchmarks the team has never seen. Nothing is retuned per benchmark. What was learned from that small set is exactly what gets carried over. The small numbers make sense once you notice what is being learned. It is not knowledge of mathematics or physics, which the members already bring. It is a way of working together, and a way of working can be reused on problems the team has never met.

Across five mathematics and physics benchmarks, the self-organizing teams average sixty-six point seven percent accuracy. Three comparisons frame that figure. The team's strongest single member reaches forty-eight point eight percent. That same agent, given compute-matched inference, reaches fifty-eight point seven percent. A perfect router, one that always picks the right member's independent answer, reaches fifty-nine percent. On AIME, the teams beat even that perfect router by thirteen point four points.

The gains are not uniform, so the authors ask when self-organizing collaboration actually helps. Their answer borrows a construct from organizational psychology called demonstrability: whether a team can distinguish correct reasoning from incorrect reasoning. Across eight benchmarks, demonstrability strongly tracks the improvement over the strongest member, with a Spearman correlation of zero point nine and a p value of zero point zero zero five. Teams benefit most when correct reasoning can be recognized once it appears.

The broader suggestion is that organization itself can become an agent capability. Picking the best individual answer has a ceiling, and the perfect router sits on it. A team that has learned how to reason together goes past that ceiling, reaching solutions its members could not reach independently. Nothing about the members themselves changes along the way. They are not replaced or retrained. What changes is how they are organized, and that organization is the part that was learned.