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.AISep 19, 2026

Self-Organizing Agent Teams Learn to Reason Together

Authors: Aneesh Pappu, Mirac Suzgun, Yongchan Kwon, Federico Bianchi, Batu El, Mykel J. Kochenderfer, Hancheng Cao, James Zou

Organizations: Stanford University · Together AI · Goizueta Business School, Emory University

Abstract

Collective intelligence depends not only on what team members know, but also on how they organize their work. When the structure of a solution is unknown, useful roles and divisions of labor cannot be specified in advance; teams must learn from experience how to organize reasoning as it unfolds. Human teams routinely adapt this way, while existing AI agent teams rely on fixed protocols, explicit task decomposition, or routing. We introduce Self-Organizing Agent Teams (SAT), fixed teams of AI agents that learn reusable strategies from prior collaborations to organize roles, conversational phases, participation, and information flow. These strategies enable what we call collaborative computation: agents exchange, challenge, repair, and synthesize partial reasoning into solutions no member produced independently. In two independent settings, we learn teamwork strategies that transfer unchanged to unseen benchmarks, using only 15 mathematics and 25 graduate-level knowledge problems. Across five mathematics and physics benchmarks, self-organizing teams average 66.7% accuracy, versus 48.8% for their strongest member, 58.7% for compute-matched inference by that agent, and 59.0% for a perfect router over members' independent answers; on AIME 2026, they exceed this router by 13.4 points. Because gains vary across benchmarks, we ask when self-organizing collaboration helps. Across eight benchmarks, demonstrability (the organizational-psychology construct of whether a team can distinguish correct from incorrect reasoning) strongly tracks improvement over the strongest member (Spearman $ρ=0.90$ , $p=0.005$ ): teams benefit most when correct reasoning can be recognized once it appears. More broadly, these results suggest that organization itself can become an agent capability: agent teams can learn how to reason together and produce solutions their members could not reach independently.

Figures & tables

Panel of Figure 1 — Figure 1: Self-Organizing Agent Teams ( SAT ) outperform their strongest member, compute-matched inference by the strongest individual agent, and perfect routing over independent answers. Across five mathematics and physics benchmarks, SAT averages $66.7\%$ accuracy, $17.9$ points above the strongest member ( $48.8\%$ ) and $7.7$ percentage points above routing-oracle coverage ( $59.0\%$ ). SAT also surpasses compute-matched inference by the strongest individual agent ( $58.7\%$ ; Table 1 ). Exceeding the routing oracle shows that some correct answers are produced through interaction rather than recovered by selecting among members’ independent outputs. Bars report final-answer accuracy; black rules mark routing-oracle coverage. The narrow blue bars, labeled with “+” values, show SAT ’s improvement over the strongest team member in absolute percentage points.

Figure 2: Self-organizing agent teams learn how to reason together from prior collaboration. A designated member reflects on earlier exchanges and outcomes, then revises the teamwork strategy governing subsequent collaboration. Candidate strategies are tested on training problems before a complementary bank is frozen for evaluation. Dialogue and role assignments are illustrative.

Panel of Figure 3 — Figure 3: Team organization is learned offline and frozen before evaluation. (a) Starting from a fixed three-model roster and $15$ AIME-2024 training problems, teamwork reflection proposes and tests organizational strategies; validation probes measure transfer, and a coverage-greedy step retains a bank of ten. At deployment, the frozen bank produces candidate certificates for a new problem, and one team member serves as judge, selecting the final answer. (b) In a GPQA mutation, the designated member diagnoses collaboration-induced loss and converts an observed member strength into an auditor role. Model identities were blinded during reflection; Agent 2 is Gemini-2.5-Flash. The resulting final_auditor_claim_recovery strategy (Appendix A.2 ) solved its source problem and all five sampled validation probes and was selected into the final bank. Reflection excerpts are verbatim and lightly trimmed.

	In-dist.	Out-of-distribution transfer
Method	AIME24	AIME25	AIME26	HMMT26	TQA-phys	Avg
Best member	64.4	40.0	42.2	26.3	71.1	48.8
Self-consistency ( $K{=}10$ ) ‡	75.8	44.4	53.4	31.9	71.5	55.4
Self-reflection ‡	71.1	41.1	52.2	29.3	72.2	53.2
Linearization (o3-mini) ‡	74.0	43.3	68.5	39.4	68.4	58.7
Member-vote	60.7	38.5	40.9	25.3	71.3	47.3

Table 1: SAT achieves the highest average accuracy across the mathematics-and-physics suite and exceeds routing-oracle coverage on average. Strategies learned on

15

AIME-2024 training problems are frozen and evaluated on held-out AIME-2024 problems and four transfer benchmarks. All values are percentages. Unshaded rows report final-answer accuracy; boldface marks the highest accuracy in each column, including ties. Shaded rows report perfect-selection coverage over independent member answers (routing oracle) or team certificates. Single-agent controls use the strongest team member on each benchmark unless otherwise stated; see Appendix D for evaluation details. Results report means over three seeds. ‡ denotes compute controls.

	In-dist.	Out-of-distribution transfer
Method	GPQA	MMLU-Pro	BBEH	Avg
Best member	72.7	79.7	45.3	65.9
Self-consistency ( $K{=}10$ ) ‡	71.6	81.2	49.8	67.5
Self-reflection ‡	77.7	81.7	54.7	71.4
Linearization (Gemini-2.5-Flash) ‡	79.0	84.0	53.3	72.1
Member-vote	67.0	79.9	43.6	63.5

Table 2: SAT achieves the highest average accuracy across the knowledge-and-logic suite, but selection leaves a substantial gap to team coverage. Across GPQA, MMLU-Pro, and BBEH, SAT averages

72.8\%

final-answer accuracy, while perfect selection from its certificate pool would reach

87.9\%

. All values are percentages. Unshaded rows report final-answer accuracy; boldface marks the highest accuracy in each column, including ties. Shaded rows report perfect-selection coverage over independent member answers (routing oracle) or team certificates. Single-agent controls use the strongest team member on each benchmark unless otherwise stated; see Appendix D for evaluation details. Results report means over three seeds. ‡ denotes compute controls.

Panel of Figure 4 — Figure 4: Teamwork reflection discovers qualitatively different ways to organize reasoning. ( 4(a) ) Problem-adaptive method diversification. Members propose distinct approaches to the current problem, divide them among the team, and compare and repair the resulting derivations before synthesis. The strategy transfers unchanged, while the choice of methods adapts to the problem. ( 4(b) ) Divergence reconciliation. DeepSeek-V3 can surface useful alternative derivations but also introduce errors. The strategy retains DeepSeek as a source of useful diversity while assigning o3-mini and Claude Sonnet 4 to independently verify disputed reasoning, with DeepSeek absent from that phase. Corroborated contributions are integrated; unresolved ones are downweighted.

Panel of Figure 5 — Figure 5: Cross-member repair produces an answer that no individual member initially had. On HMMT February 2026 problem 6, all three independent answers are wrong. o3-mini supplies the invariant and factorization, DeepSeek-V3 repairs the decisive counting error, and Claude Sonnet 4 audits the correction; synthesis returns $3840$ , absent from all three initial responses.

Figure 6: Demonstrability tracks when self-organizing teams improve over their strongest team member. Across eight benchmarks, demonstrability strongly tracks the rank order of team improvement over the strongest team member (Spearman

\rho=0.90

, exact permutation

p=0.005

). Demonstrability is the balanced rate at which a ten-model panel selects correct over incorrect team reasoning. The axes rank benchmarks by demonstrability and absolute improvement over the strongest team member; the dashed diagonal denotes perfect rank agreement. Raw scores and leave-one-benchmark-out sensitivity results appear in Appendix Tables 5 and 6 .

Panel of Figure 7 — Figure 7: Learned teamwork supports a richer interaction structure than common debate and feed-forward aggregation multi-agent systems. Debate repeats symmetric exchange before voting; Mixture of Agents aggregates independent proposals in one direction. A learned strategy can instead vary participation across phases, maintain persistent roles, and revisit earlier reasoning for targeted repair. For a transcript-grounded comparison on a common GPQA problem, see Appendix Figure 10 .

Approach	Collaborative computation [1pt]joint reasoning, not aggregation or routing	Problem-agnostic collaboration scaffold [1pt]reused across tasks and benchmarks	Learned computation [1pt]not hand-designed
Debate, voting, Mixture of Agents	$\times$	$\checkmark$	$\times$
Topology optimization (GPTSwarm, AgentNet)	$\times$	$\times$	$\checkmark$
Interaction-graph optimization (OPTAGENT)	$\checkmark$	$\times$	$\checkmark$
Problem-conditioned orchestration (Conductor)	$\times$	$\times$	$\checkmark$
PI-led scientific collaboration (Virtual Lab)	$\checkmark$	$\times$	$\times$
Open agent platform (EinsteinArena)	$\checkmark$	$\checkmark$	$\times$

Table 3: A landscape of multi-agent collaboration. Multi-agent approaches differ along three dimensions: whether agents jointly develop reasoning, whether their collaboration scaffold is agnostic to problem content and reused across tasks and benchmarks, and whether that scaffold is learned rather than hand-designed.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 8 — Figure 8: Teamwork reflection explores a broader strategy space than the final banks reveal. ( 8(a) ) Forced role reversal. Members first vote blindly, then each argues for an answer other than its own, cross-examines the alternatives, and votes again only after the temporary advocacy roles are released. This separates the strength of an argument from its proponent’s original answer and forces minority alternatives to be examined. ( 8(b) ) Overlapping pairwise verification. Each member states an answer and its critical assumption, after which the strategy verifies all three member pairs in an overlapping sequence. A detected flaw triggers full-team re-solving; if no flaw is found, unanimity still triggers a challenge to one key interpretation before synthesis. These strategies were discovered during earlier searches but were not selected for either final bank and were not used in reported deployment.

Method	AIME24	AIME25	AIME26	HMMT26	TQA-phys	Avg
$P_{\mathrm{init}}$	73.3	50.0	50.0	24.2	72.8	54.1
SAT (Ours)	84.7	60.8	71.2	39.4	77.2	66.7

Appendix

Table 4: SAT achieves higher average accuracy than the initialization in both task suites. Panels (a) and (b) compare SAT with

P_{\mathrm{init}}

, the initial teamwork strategy used to seed the evolutionary search that implements teamwork reflection. The initialization has each team member propose an independent solution, followed by two rounds of debate-like exchange and a final majority vote. All values are percentages; boldface marks the higher accuracy in each column, including ties.

Panel of Figure 9 — Figure 9: Cross-member composition constructs an answer absent from all three initial outputs. On HMMT 2026 problem 6 (gold $3840$ ; all three initial answers were incorrect), o3-mini supplies the gcd–lcm exponent invariant and factorization but miscounts the unit-exponent primes. DeepSeek repairs the count, Sonnet audits it, and synthesis preserves the composed answer. The transcript is condensed for legibility.

Figure 10: A learned peer-audit strategy repairs errors that debate and aggregation preserve. GPQA Diamond 132 asks the product count for 1-bromobenzene-2-d reacting with NaNH 2 in ammonia. All methods use Llama 4 Maverick, GPT-4.1, and Gemini 2.5 Flash; the correct answer is B (three products). (a) Mixture of Agents starts with proposals D/D/C. The aggregator merges the H- and D-removal intermediates and incorrectly retains deuterium, counting 2- and 3-deuterioaniline but omitting aniline. It returns D (two). (b) Debate starts with a wrong D majority (Llama and Gemini); GPT initially chooses the correct answer B. Gemini’s first-round analysis falsely equates the D-retaining pathway’s two products. Citing the other agents’ analyses, GPT adopts this equivalence and switches to D; Llama adopts the same product accounting. The wrong majority becomes unanimous, omitting 3-deuterioaniline. (c) SAT begins with all three agents independently recommending wrong answers: D/A/D (two/four/two). Its learned strategy asks each member to audit a specific peer claim, propose a correction, then acknowledge corrections, explain initial errors, and revise their answer before synthesis. Gemini challenges Llama’s merged pathways, then corrects its own assumption that deuterium survives removal. The team preserves the two-pathway mechanism and repairs atom tracking: removing D gives aniline; removing H gives 2- and 3-deuterioaniline. Green marks correct B and light blue incorrect D; teal marks the audit and proposed repair, orange the acknowledged correction. Transcripts and instructions are condensed; quotations are exact.

Panel of Figure 11 — Figure 11: A failed check becomes the missing construction. On AIME 2026 II-08 (gold $245$ ; initial outputs $028$ / $049$ /unparsed), a mandated integer check exposes the gap in the leading derivation. Sonnet turns that failed check into the divisibility constraint needed to construct $245$ , and DeepSeek verifies minimality.

Panel of Figure 12 — Figure 12: Verification protects a correct minority against a wrong majority. On AIME 2026 II-01 (gold $178$ ; initial answers $032$ / $178$ / $032$ ), two members share the same locally checkable error. A step audit names that error, both wrong members revise, and the team preserves the lone correct answer that voting would discard.

Benchmark	Demonstrability	Improvement over best member (pp)
AIME 2024	0.550	20.3
AIME 2025	0.559	20.8
AIME 2026	0.687	29.0
HMMT 2026	0.519	13.1
TheoremQA-physics	0.423	6.1
GPQA	0.534	7.3

Appendix

Table 5: Demonstrability and SAT improvement across eight benchmarks. Demonstrability is operationalized as the balanced rate at which a ten-model panel selects correct over incorrect reasoning. Improvement is SAT accuracy minus the best member’s accuracy, in percentage points.

Multi-Agent Collaboration

Multi-agent collaboration studies how autonomous agents coordinate, share information, and combine capabilities to solve tasks together.

See all 320 papers

May 8, 2026

TeamBench: Evaluating Agent Coordination under Enforced Role Separation

Agent systems often decompose a task across multiple roles, but these roles are typically specified by prompts rather than enforced by access controls. Without enforcement, a team pass rate can mask whether agents actually coordinated or whether one role effectively did another role's work. We present TeamBench, a benchmark with 851 task templates and 931 seeded instances for evaluating agent coordination under operating system-enforced role separation. TeamBench separates specification access, workspace editing, and final certification across Planner, Executor, and Verifier roles, so that no role can read the full requirements, modify the workspace, and certify the final answer. Prompt-only and sandbox-enforced teams reach statistically indistinguishable pass rates, but prompt-only runs produce 3.6 times more cases where the verifier attempts to edit the executor's code. Verifiers approve 49% of submissions that fail the deterministic grader, and removing the verifier improves mean partial score in the ablation. Team value is also conditional. Teams benefit when single agents struggle, but hurt when single agents already perform well. A 40-session human study under the same role separation shows that our benchmark exposes interaction patterns that pass rate misses. Solo participants work through the task directly, human participants paired with agents often collapse into quick approval, and human teams spend more effort coordinating missing information across roles.

Yubin Kim, Chanwoo Park, Taehan Kim +9

MIT · Google Research · Independent Researcher +1

Multi-Agent Coordination LLM Agent Evaluation AI Agent Security AI Agent Benchmarks Task Decomposition Tool Access Control for LLM Agents+1

Ask PDF

Sep 17, 2026

Scaling Discovery through Test-Time Communication

Science advances not in isolation but through collaboration, yet existing agentic systems capture little of this. Whether communicating agents help remains an open question with mixed prior results. We show that test-time communication can substantially outperform independent parallel attempts on challenging tasks, where sharing a breakthrough can push the whole group forward. We first study the effect of scaling multi-agent test-time communication, where agents have no predefined roles and communicate via a shared directory, on ARC-AGI-3, a benchmark requiring novel problem solving. We find that a team of

k

communicating agents, team@

k

, matches the success rate of

4k

independent agents, and this advantage grows with

k

, suggesting gains compound with scale. The effect is not merely efficiency: a task that no single agent can solve, a team of agents can solve reliably. Furthermore, these gains transfer to research-oriented tasks, given sufficient compute. On polyomino packing, communicating agents outperform best@

k

and exceed the prior best-known score. On MNIST classifier compression, communication surpasses the best-known human solution. A team of four agents produced a 1,957-byte classifier submission achieving 99.4% test accuracy, smaller than both the best-known human solution and the best single-agent result. These gains are not unconditional. Independent agents may outperform communication when compute is limited or when a clear measure of progress is absent. However, under sufficient compute and clear feedback, multi-agent communication consistently yields stronger results.

Jongho Park, Vasilis Kontonis, Shivam Garg +2

UC Berkeley · Microsoft Research

Multi-Agent Collaboration Multi-Agent Communication Test-Time Scaling AI Agents for Scientific Discovery Shared-Memory Multi-Agent Systems ARC-AGI

Ask PDF

May 27, 2026

AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation

Scientific research proceeds through iterative cycles of hypothesis generation, experiment design, execution, and revision. AI agents can automate parts of this process, but existing approaches typically follow a single research trajectory or coordinate through a central planner with fixed objectives. As a result, they struggle to sustain parallel exploration, adapt as experimental evidence changes, or preserve knowledge of failed directions over long-running experiments. We introduce AutoScientists, a decentralized team of AI agents for long-running computational scientific experimentation. Agents interpret a shared experimental state, self-organize into teams around promising hypotheses, critique proposals before using experimental compute, and share successes and failures to reduce redundant exploration. Under matched experimental budgets, AutoScientists improves over prior AI agents across biomedical machine learning, language-model training optimization, and protein fitness prediction. On BioML-Bench, spanning biomedical imaging, protein engineering, single-cell omics, and drug discovery, AutoScientists achieves a mean leaderboard percentile of 74.4% across 24 tasks, improving over the strongest AI agent by +8.33%. On GPT training optimization, AutoScientists reaches a target validation bits-per-byte 1.9x faster than Autoresearch and continues discovering improvements from a starting champion where the single-agent approach finds none (7 vs. 0 accepted improvements). On ProteinGym fitness prediction, AutoScientists discovers a method for ACE2-Spike binding that improves over the current state-of-the-art model by +12.5% in Spearman correlation. Applied without modification across all 217 ProteinGym assays, the same method improves over the prior state of the art by +6.5% (Spearman correlation).

Shanghua Gao, Ada Fang, Marinka Zitnik

Harvard University

Protein Fitness Prediction Multi-Agent Collaboration Multi-Agent Systems Autonomous Scientific Discovery Autonomous Experimentation

Ask PDF

Abstract

Multi-Agent Collaboration

Explore similar work

TeamBench: Evaluating Agent Coordination under Enforced Role Separation

Scaling Discovery through Test-Time Communication

AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation