Self-Organizing Agent Teams Learn to Reason Together
Aneesh Pappu, Mirac Suzgun, Yongchan Kwon, Federico Bianchi, Batu El, Mykel J. Kochenderfer, Hancheng Cao, James Zou
cs.AI, cs.MA
2026-09-19
Self-organizing teams average 66.7% on five math/physics benchmarks, beating a 59.0% perfect router; knowledge coverage is 87.9% but selected accuracy is 72.8%.
Multi-agent systems usually take one of two routes. Debate, voting, and Mixture of Agents start from complete individual answers and combine them. Workflow search and learned routers split a problem into subtasks first. Both assume either comparable full answers or a decomposition you can write down in advance. Many hard problems have neither. A useful division of labor appears only in conversation: one member names an invariant and miscounts, another fixes the count, a third audits. Human teams learn those roles by working together. Most agent teams still treat organization as a fixed protocol, not as something the roster can learn and reuse.
Stanford, Together AI, and Emory treat organization as a capability. A fixed roster learns reusable teamwork strategies from old collaborations, freezes them, and runs them unchanged on unseen benchmarks.
A strategy is a small language. It lists conversational phases, a shared teamwork prompt, and persistent role prompts. Each phase sets who speaks, how many rounds, whether information stays local or is summarized afterward, and the phase instruction. It does not decompose a new problem. It only says how the conversation is staged.
Learning has three stages. The member with the best training accuracy does teamwork reflection: o3-mini on math, Gemini-2.5-Flash on knowledge. It reads prior strategies, transcripts, outcomes, and validation probes, then mutates roles, phases, and synthesis rules. Six rounds per source problem, up to three candidates per round. A mutation that solves its source problem is tested on five other training items. A semantic audit rejects strategies that encode answers, problem facts, or solution recipes. Coverage-greedy selection freezes at most ten strategies. At test time every strategy runs and emits a certificate with step-by-step checkable reasoning. The same member judges the pool by auditing local defects, without solving the problem independently.
Two independent rosters. Math and physics: o3-mini, Claude Sonnet 4, DeepSeek-V3, trained on 15 AIME 2024 problems. Knowledge and logic: Gemini-2.5-Flash, Llama-4-Maverick, GPT-4.1, trained on 25 GPQA Diamond problems. Controls include the best member, self-consistency at K=10, five-pass self-reflection, member vote, three-round debate, Mixture of Agents, a homogeneous team running the same strategies, and linearization (the strongest member plays every role serially at about the team's budget). The strictest bar is the routing oracle: a perfect per-problem selector over members' independent answers. Beating it means interaction created a solution nobody supplied alone.
On five math and physics benchmarks SAT averages 66.7%, versus 48.8% for the best member, 58.7% for linearization, 57.3% for Mixture of Agents, 59.0% for the routing oracle, and 75.2% team coverage. Held-out AIME 2024 is 84.7% against o3-mini at 64.4%. The same frozen bank transfers to AIME 2025 (60.8% vs 40.0% best member, 51.1% oracle), AIME 2026 (71.2% vs 42.2% and 57.8%, +13.4 points over the oracle), HMMT 2026 (39.4%, tied with linearization, best member 26.3%), and TheoremQA-physics (77.2% vs 71.1%). On HMMT problem 6 all three initial answers are wrong: o3-mini supplies the invariant and miscounts, DeepSeek-V3 repairs the count, Claude audits, and the team returns 3840.
| Method | Math/physics avg | AIME 2026 |
| Best member | 48.8% | 42.2% |
| Linearization (o3-mini) | 58.7% | 68.5% |
| Routing oracle | 59.0% | 57.8% |
| SAT | 66.7% | 71.2% |
| SAT coverage | 75.2% | 73.3% |
On three knowledge and logic benchmarks SAT averages 72.8%, a thin edge over linearization at 72.1% and Mixture of Agents at 71.2%, still below the oracle at 79.6%. Coverage is 87.9%. SAT leads on GPQA at 80.0%, trails Mixture of Agents on MMLU-Pro (82.4% vs 84.3%), and sits at 56.0% on BBEH behind a homogeneous team at 58.7%. Generating a correct certificate and selecting it are different jobs.
Across eight benchmarks, demonstrability (how often a ten-model panel picks the correct certificate over a wrong one) tracks improvement over the best member with Spearman ρ=0.90 (p=0.005). Collaboration pays when correct reasoning, once produced, can still be recognized.
Organization can be learned from old dialogues and reused across years, contests, and domains. Beating the routing oracle means the gain is not only "listen to whoever is already right." The practical split is sharp: on math, where steps can be checked, interaction can mint new solutions; on knowledge items the team often has the right certificate and then drops it. The rosters were chosen with headroom. Stronger models that saturate these benchmarks would hide the same teamwork effect.
Knowledge coverage is 87.9% while accuracy is 72.8%; a correct writeup does not guarantee a correct pick. Demonstrability is a post-hoc correlation and does not steer strategy search. Evolution uses 15 and 25 training problems; the bank is frozen and never adapted per test item. Linearization uses the training-best member, so a stronger serial model might close part of the gap. Compute is matched by candidate count (K=10), with no dollar or token report. The opening anecdote about agents building a channel in an OpenAI cybersecurity eval is motivation, not an experiment.