acceptodds
Under review as a conference paper at ICLR 2027

Self-Organizing Agent Teams Learn to Reason Together

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 always 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 typically 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 reusable 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 accuracy, versus for their strongest member, for compute-matched inference by the strongest individual agent, and for a perfect router over members' independent answers; on AIME 2026, they exceed this router by percentage points. Because these gains vary across benchmarks, we ask when self-organizing collaboration improves over individual models. Across eight benchmarks, demonstrability—an organizational-psychology construct capturing whether correct reasoning can be distinguished from incorrect reasoning—strongly tracks how much the team improves over its strongest member (Spearman , ), suggesting that self-organizing agent 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.

Primary area:
foundation or frontier models, including LLMs
Keywords:
multi-agent, organization, emergence, teamwork, reflection, evolution, learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zPr4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces SAT, which is a fixed three LLMs team whose way of collaborating is learned by reflective evolutionary search on a small training set. On five math and physics benchmarks, SAT averages 66.2%, which beats the best member, a compute-matched single-agent control, and a routing oracle that perfectly selects among the member's independent answers. The paper also argues that demonstrability (whether correct reasoning can be told apart from incorrect ones) explains when collaboration helps.

Critical strengths

This paper is well-written. The question is important and motivated. The routing oracle is a valuable contribution. Details are pretty transparent. The full strategy banks are released, and reporting coverage alongside accuracy cleanly separates generating correct solutions from selecting them.

Critical weaknesses
  1. The oracle picks from a single answer per member, while SAT picks from 10 multi-agent conversations, so the comparison mixes up interaction with sample count. I think a better reference should be oracle coverage over many independent samples per member at matched token cost.

  2. Although the authors implemented a wide range of baselines already, I find the effect of learning is still confounded with pure ensembling and selection. Particularly, SAT runs 10 strategies and use an o3-mini auditor to choose among them. The initial strategy is run once with majority vote, and no baseline gets a comparable judge. The gains thus cannot be credited to learned organization until the same judge is applied to equally sized baseline pools. Could you also report each strategy's accuracy on its own?

  3. Gains over the strongest control seem small and statistical uncertainty is not reported. SAT beats linearization by 3.7 points on AIME26 and ties it on HMMT26. On the knowledge suite the margin is 0.7. In your experiments, the in-distribution AIME24 test set has only 15 problems and there are no CIs or paired tests.

  4. I think the framing somehow overstates. There is one member designing the strategies offline, and the bank is a frozen one, and a central judge picks the answer. This to me is a learned protocol ensemble rather than true self-organization as described in the text.

Questions for authors
  1. Figure 7 mentions earlier searches: how many search runs and team rosters were tried, and how was the final setup chosen?
  2. What are the per-problem token costs for each method, and what does each strategy in the bank achieve alone?
  3. Do the results hold with current frontier models? Could Claude Sonnet 4 have seen AIME 2025 problems during training?
Reviewer_zCxYOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks how a team can learn to organize reasoning when the useful decomposition of a test problem is unknown, rather than merely aggregating independent answers. SAT learns teamwork strategies covering roles, conversational phases, information flow, and synthesis, then freezes those strategies and transfers them to new tasks. Two team settings are evaluated across five and three benchmarks, with average accuracies of 66.2% and 72.8%. The relationship between demonstrability and performance gain is still a post hoc correlation.

Critical strengths
  1. Team organization is the learned and transferable object. The strategy specifies roles, phases, information flow, and verification order without hard-coding a decomposition for each test problem. It is frozen after training and transfers across years, competitions, and physics.
  2. The controls and diagnostic measures are strong. Linearization tests whether one model can reproduce the structure, the homogeneous team tests model heterogeneity, and the coverage/accuracy split exposes a selection bottleneck. The HMMT and AIME repair cases make the claim about collaborative solution construction inspectable.
Critical weaknesses
  1. The routing oracle may not be a same-compute upper bound. If the oracle selects one independent answer per member while SAT uses ten strategies and multiple candidate-selection passes, the gain over the oracle mixes interaction benefits with extra candidate sampling. This weakens the strongest claim about generating answers absent from independent inference, although it does not remove the paper’s more basic problem–solution contribution.
  2. The demonstrability result has limited statistical and causal force, and the source code cannot be checked. The ρ=0.88 association is based on eight aggregate benchmark points. Leave-one-out analysis helps, but the association may still reflect benchmark difficulty, model composition, and task type. The paper provides no prospective intervention or independent holdout test, and no anonymous source link for checking strategy search, bank freezing, candidate selection, and oracle implementation.
Questions for authors
  1. Can you recompute the independent-sampling routing oracle with the same number of members, samples, candidates, and total reasoning budget as SAT, and report per-benchmark and per-seed results for SAT, linearization, and the oracle?
  2. Can you fix demonstrability before evaluating a new benchmark that was not used for strategy learning or correlation analysis, and test whether it predicts SAT’s gain over the strongest member?
  3. Could the final certificate selector benefit by implicitly solving the problem rather than only auditing the certificate? Please report its independent problem-solving ability or compare stronger/different selectors, and provide anonymous source code and a review-access link to the complete strategy bank.
Reviewer_chPZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Self-Organizing Agent Teams (SAT), a framework that uses reflection-based optimization to learn reusable multi-agent collaboration strategies, including agent roles, communication patterns, and reasoning phases. Experiments on mathematical and knowledge-intensive benchmarks show improvements over several single-agent and multi-agent baselines. The authors further investigate how the ability to recognize correct reasoning relates to collaboration effectiveness.

Critical strengths
  1. Learning effective collaboration strategies for heterogeneous agents is an interesting and potentially useful research direction.

  2. The paper evaluates multiple benchmarks and includes several meaningful baselines, such as compute-matched single-agent inference and homogeneous agent teams.

Critical weaknesses
  1. The proposed method closely resembles existing reflection-based prompt optimization and automated multi-agent workflow optimization approaches. The main distinction appears to be a more flexible search space for collaboration protocols, rather than a fundamentally new methodology.

  2. The reflection process heavily depends on the quality of previously generated reasoning trajectories, which can be highly stochastic. Given the small training sets, it is unclear whether the learned strategies are stable across independent optimization runs or reflect incidental patterns in sampled trajectories.

  3. The paper suggests that different models develop complementary roles, but the supporting examples are largely anecdotal. It remains unclear whether these role assignments capture reliable model-specific strengths or whether similar gains could be achieved through arbitrary role assignments.

Questions for authors

How does SAT compare with existing prompt optimization methods, such as GEPA, under the same search space and optimization budget?

How consistent are the learned strategies and their performance across independent runs of the entire optimization process?

Have the authors considered randomly permuting agent roles to determine whether model-specific specialization actually contributes to the improvements?

Comments on the paper

1 comment

Hi everyone, please acknowledge the authors’ rebuttal with a brief reply, even if your assessment is unchanged (if you already did or are about to do so, please ignore this message).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.