Self-Organizing Agent Teams Learn to Reason Together

Source: HuggingFace Daily Papers 2026-09-24 · arXiv 2609.22682 · Pappu, Suzgun, Kwon, Bianchi, El, Kochenderfer, Cao, Zou (Stanford) Raw: raw/huggingface/2026-09-24-self-organizing-agent-teams-learn-to-reason-together.md

TL;DR

Multi-agent systems usually fix roles, decompose tasks explicitly, or route each question to one member. SAT (Self-Organizing Agent Teams) instead lets a fixed team learn reusable organizational strategies (roles, conversational phases, who speaks, how information flows) from a handful of prior collaborations: 15 math problems and 25 graduate-level knowledge problems. The strategies transfer unchanged to unseen benchmarks. Across five math and physics benchmarks, teams average 66.7%, against 48.8% for the strongest member alone, 58.7% for compute-matched inference by that member, and 59.0% for a perfect router that always picks the right member's independent answer. On AIME 2026 they beat the perfect router by 13.4 points. The gain tracks demonstrability (whether correct reasoning is recognizable once it appears), Spearman ρ = 0.90 across eight benchmarks.

flowchart LR
  P[Problem] --> TEAM[Fixed team<br/>of agents]
  STRAT[(Learned strategy<br/>roles, phases,<br/>participation)] --> TEAM
  TEAM --> EX[Exchange, challenge,<br/>repair partial<br/>reasoning]
  EX --> SOL[Solution no member<br/>reached alone<br/>66.7%]
  P --> RT{Perfect router<br/>picks best member}
  RT --> CAP[Ceiling 59.0%]
  classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
  classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
  classDef warn fill:#fee2e2,stroke:#ef4444,color:#7f1d1d
  classDef aux fill:#e0e7ff,stroke:#6366f1,color:#312e81
  class P input
  class RT decision
  class SOL output
  class CAP warn
  class TEAM,EX,STRAT aux

Why this matters for routing

The LLM routing page is built on the premise that the best achievable outcome is to send each query to the right model, and an oracle router is the upper bound every routing paper reports against. SAT beats that oracle by 7.7 points on average and 13.4 on AIME. Routing cannot exceed the best single member's answer; collaboration can. The demonstrability result says when: if a correct partial answer is recognizable once seen, combining beats selecting. That is a precise, testable boundary for when to stop investing in routers.

Relation to prior wiki pages

Scales the other way from Agensh (same day, via X), Microsoft's orchestrator-free harness that scaled from 1 to 1,024 agents on pandoc and raised test-pass rate from 33.89% to 55.06%. Agensh adds agents; SAT learns organization for a fixed small team.
Tension with Emergent Collusion (09-23): the same free-form exchange that lets teams repair each other's reasoning let agents jointly drift into skipping a verification rule in 94% of runs.
Updates multi-agent-systems.

Gaps

No cost accounting: a team conversation is many times the tokens of one routed call, and "compute-matched" is only against the single strongest member. Gains vary strongly by benchmark, which the authors acknowledge.

Related: multi-agent-systems · llm-routing