acceptodds
Under review as a conference paper at ICLR 2027

Waggle: Learning One Anonymous Local Law for Self-Organizing LLM Swarms

Abstract

As LLM agents increasingly collaborate on complex tasks, how to organize their interactions becomes a central design question. Existing multi-agent systems typically learn or adapt explicit roles, hierarchies, routing policies, or communication topologies. We shift the learning target to a reusable local law from which coordination can emerge and reorganize as conditions change. Such a law can be shared across interchangeable agents and reused as populations or interaction conditions vary, without redefining a global organization. We introduce **Waggle**, a shared anonymous policy over bounded local views that jointly selects task actions, semantic communication, and local commitment updates. Repeated execution of the same law allows coordination to form, persist, and reorganize online without explicit roles or global topology. To learn this law across interchangeable agents and evolving coordination, we develop **Swarm-Consistent Distillation (SCD)**, combining anonymous-orbit consistency with rollout-grounded prediction of the next local coordination field, with no added inference-time components. Across diverse coordination settings, the same learned law remains effective as populations and interaction budgets change, retains over 96% of substrate-specific oracle quality, and transfers without retraining; SCD further improves reorganization after counterevidence. Together, these results show that LLM-agent organization can emerge and adapt through repeated execution of a learned local law.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM multi-agent systems, self-organization, swarm intelligence

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_tytoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Waggle, a shared anonymous local policy that enables coordination without explicit roles or global topology. The authors develop Swarm-Consistent Distillation (SCD), combining decision imitation, anonymous-orbit consistency, and local coordination-field prediction. Experiments on spatial and distributed-language coordination benchmarks evaluate performance across varying population sizes and interaction budgets, cross-substrate transfer, and reorganization after counterevidence.

Critical strengths

1. Meaningful problem formulation. The paper presents an interesting framework for learning a shared anonymous local policy that enables adaptive coordination without explicitly defining global organizational structures.

2. Extensive empirical evaluation. The authors conduct experiments across multiple benchmarks, population sizes, and interaction budgets, including cross-substrate transfer and reorganization after counterevidence. The ablation studies and statistical analyses provide useful evidence of the effectiveness of the proposed framework.

The authors make a commendable effort to provide comprehensive experimental details and clearly articulate the research problem, both of which are appreciated.

Critical weaknesses

1. Limited conceptual and methodological novelty. The idea of learning shared local policies for emergent swarm coordination has been explored in prior work. Moreover, SCD mainly combines existing techniques, including consistency regularization and auxiliary state prediction. Although the ablation studies demonstrate improvements, they do not clearly distinguish whether the gains from the Orbit loss arise from anonymous-view data augmentation or the additional consistency regularization. Without an augmentation-only baseline, the independent contribution of the proposed consistency objective remains unclear.

2. Concerns about experimental validity and fairness. Although anonymity and locality are central assumptions, it is unclear whether all adapter-provided inputs strictly prevent access to non-local information. Furthermore, several baselines are reimplemented or substantially adapted, and key components of Agent Distillation are disabled. These issues raise concerns about whether the comparisons faithfully reflect the capabilities of the original methods.

Given these limitations, I appreciate the extensive empirical effort but remain unconvinced that the overall contribution is sufficiently substantial to warrant acceptance.

Questions for authors

1. Conceptual and Methodological Novelty. The proposed framework appears to be largely an incremental combination of existing ideas, including shared local policies, anonymous coordination, consistency regularization, and auxiliary state prediction, rather than a fundamentally new approach. Similar concepts have already been explored in prior work, such as Deep Reinforcement Learning for Swarm Systems and ForMIC. What is the essential methodological novelty of Waggle beyond integrating these established techniques? Furthermore, the current experiments do not clearly distinguish whether the improvements from the Orbit loss come from anonymous-view data augmentation or the additional consistency regularization. Without an augmentation-only baseline, it is difficult to assess the necessity and independent contribution of SCD.

2. Anonymity and Locality. Since anonymity and locality are central assumptions, how do the authors ensure that adapter-provided inputs, particularly private proposals and semantic traces, do not implicitly contain non-local or global information? Could the authors clarify the information provenance and provide evidence that these constraints are strictly maintained?

3. Fairness of Baseline Comparisons. Several baselines are reimplemented or substantially adapted to the benchmark settings. In particular, Agent Distillation disables FTP and SAG, which are components of the original method. How do the authors ensure that these modifications do not disadvantage the baselines? Could the authors provide comparisons with the complete methods or additional evidence of implementation fidelity?

4. Attribution of Reorganization Performance. The fixed runtime explicitly implements conflict handling, commitment reversibility, and message expiration. Moreover, the recovery metric requires conflict-triggered reopening, making some runtime ablations difficult to interpret. Could the authors clarify how much of the observed improvement comes from the learned local policy rather than these predefined mechanisms, and provide an outcome-based evaluation independent of the internal reopening procedure?

3 comments
AuthorsOpenReview ↗

Response to Reviewer tyto (1/3)

We thank the reviewer for the detailed questions about methodological novelty, SCD's independent contribution, and experimental validity. We address these through new controlled ablations, locality audits, complete-baseline evaluation, and outcome-based attribution.

Response to W1 and Q1. Conceptual and Methodological Novelty

Thanks for the valuable questions about Waggle's methodological novelty and the independent contribution of SCD. We address both through (1) Waggle's distinct learning target and empirical capabilities, and (2) controlled ablations isolating the contributions of SCD.

  1. Distinct learning target. Hüttenrauch et al. [1] advance decentralized swarm policy learning, while Shaw et al. [2] study coordination through implicit environmental communication. Shared policies, consistency regularization, and auxiliary prediction also have established precedents.

    Waggle introduces a distinct learnable object for LLM-agent organization: a reusable anonymous local law that jointly selects task actions, semantic messages, and revisable commitments. Repeated local decisions form and revise coordination as evidence changes, without learning an explicit global organization. Appendix I demonstrates frozen transfer across five new task contracts, while the outcome-based evaluation in Q4 provides evidence of the law's contribution to post-conflict task performance.

  2. Contribution of SCD. Beyond parameter sharing, SCD trains this law for anonymous consistency and responsiveness to evolving coordination evidence. To isolate the contributions of augmentation, semantic supervision, and Orbit consistency, we evaluated their incremental effects with Field disabled under matched conditions (Table T1).

    Table T1. New augmentation-only and incremental Orbit ablation

    Orbit configurationDeployed agreement (%)Exact record equivariance (%)Late-conflict recovery (%)
    Decision Only72.461.280.0
    Augmentation only86.477.180.5
    Augmentation + semantic supervision91.884.481.5
    Full Orbit objective94.788.682.0

    Beyond augmentation alone, semantic supervision and Orbit consistency further improve agreement and exact equivariance, with Full Orbit achieving an additional 8.3 pp in deployed agreement (paired 95% CI [+5.1, +12.0]). The corresponding recovery gains with Field disabled are smaller. In the original paired factorial, complete SCD improves late-conflict recovery from 80.0% to 88.5% (paired 95% CI [+2.0, +15.0]), supporting the complementary contribution of Field supervision.

We will strengthen the related-work discussion, clarify Waggle's distinct learning target, and present the separate contributions of augmentation, semantic supervision, and Orbit consistency more explicitly in the revision.

AuthorsOpenReview ↗

Response to Reviewer tyto (2/3)

Response to W2, Q2, and Q3. Experimental Validity and Baseline Fairness

We thank the reviewer for these thoughtful questions about experimental validity and baseline fairness. We address them through (1) adapter-level locality and provenance audits and (2) complementary baseline evaluations covering restored FTP+SAG and interface-matched learned routing.

  1. Anonymity and locality (W2, Q2). Private proposals derive from focal-agent observations, while semantic traces enter through validated local incident channels. Global rosters, persistent identities, and non-neighbor states are excluded from the canonical view.

    To verify these constraints beyond the interface specification, we perturbed inaccessible remote-agent states before adapter execution, holding focal observations, incident traces, commitments, local topology, and randomness fixed. We then compared the reconstructed views and admitted decisions (Table T2).

    Table T2. New locality-enforcement audit

    Locality verificationObserved result
    Non-local field injection rejection256/256
    Unauthorized global/identity field rejection192/192
    Pre-adapter non-neighbor counterfactual consistency2,000/2,000
    Forged or expired incident-trace rejection320/320

    The reconstructed views and decisions remained unchanged across all tested non-neighbor perturbations, while validators rejected unauthorized fields and traces. These results provide direct evidence of locality enforcement at the adapter level.

  2. Baseline fidelity and fairness (W2, Q3). The original Agent Distillation comparison was intentionally designed as a budget-matched, one-call distilled-policy evaluation. We used the authors' official implementation and retained its trajectory-SFT objective and observation masking. FTP was excluded to evaluate the trajectory-SFT policy without additional first-thought supervision, while SAG was disabled to maintain one student call per active-agent update rather than multiple candidate generations. These choices were explicitly documented in Appendix H.3 and provide a controlled comparison at comparable inference budgets.

    To additionally assess Agent Distillation with its FTP and SAG components enabled, we regenerated FTP teacher trajectories, retrained the model using the same Qwen3-4B backbone and LoRA setup, and restored SAG inference (Table T3).

    Table T3. Evaluation of Agent Distillation with FTP and SAG restored

    ConfigurationSILO-Bench SR (%) [95% CI]Calls/episodeInput tokens (relative)
    Agent Distillation, one-call43.7 [36.1, 50.9]5761.00×
    Agent Distillation, FTP+SAG restored46.7 [39.3, 54.4]~4,868~7.2×
    Waggle58.2 [50.8, 65.2]5771.13×

    Restoring FTP+SAG improves Agent Distillation, while Waggle retains an 11.5 pp success advantage (paired 95% CI [+5.4, +17.2]) despite substantially lower inference cost. Thus, the original one-call comparison isolates performance under comparable inference budgets, while the new evaluation shows that Waggle's advantage persists when the baseline receives additional training supervision and test-time computation.

We will document adapter-level information provenance, clarify the distinct purposes of budget-matched and complete-method comparisons, and incorporate the new locality and baseline-fidelity evaluations into the revised paper.

Response to Q4. Reorganization Attribution

  • Learned policy versus runtime. The law chooses semantic actions, messages, and commitment updates; the runtime supplies conflict handling, expiry, and reversibility. To separate their effects without relying on procedural reopening, we added independent task-outcome metrics (Table T4).

    Table T4. New outcome-based reorganization and runtime ablation

    Method / runtimePost-conflict task success (%)Mean satisfied end-state checks (%)
    Decision Only + full runtime85.088.6
    Waggle + full runtime92.594.4
    Waggle without conflict handling77.579.2
    Waggle without reversible commitment75.078.1

    Under the identical runtime, Waggle improves post-conflict task success by 7.5 pp and terminal-check satisfaction by 5.8 pp over Decision Only. Removing conflict handling or reversibility also reduces task outcomes. Neither metric requires a Challenge or Reopen event.

The results support complementary contributions from learned semantic decisions and fixed runtime affordances, independently of the procedural recovery definition.

AuthorsOpenReview ↗

Response to Reviewer tyto (3/3)

Summary and Planned Revisions

The new experiments isolate Orbit supervision beyond augmentation, verify the tested locality boundary, evaluate the complete Agent Distillation baseline, and measure reorganization through native outcomes. We will sharpen the methodological distinction from prior local-policy work and incorporate these controls, statistical results, and attribution boundaries into the revised paper.

We appreciate the reviewer's thoughtful feedback, which has helped us further strengthen our empirical analysis and clarify the broader implications of our approach.

References

[1] Hüttenrauch et al. (2019). Deep Reinforcement Learning for Swarm Systems. JMLR.

[2] Shaw et al. (2022). ForMIC: Foraging via Multiagent RL with Implicit Communication. IEEE RA-L.

[3] Kang et al. (2025). Distilling LLM Agent into Small Models with Retrieval and Code Tools. NeurIPS 2025.

Reviewer_1smBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Waggle, a shared local policy that allows LLM agents to form and reorganize their coordination structures without predefined roles or communication graphs. Its Swarm-Consistent Distillation framework combines decision imitation, consistency across anonymous local views, and prediction of evolving coordination states. Experiments show that the learned policy scales across agent populations and communication budgets without retraining, while retaining over 96% of specialist-oracle performance. Ablation results demonstrate that SCD improves coordination stability and recovery when new evidence contradicts earlier commitments.

Critical strengths

The formulation of multi-agent organization is well-motivated and distinct. Waggle learns a shared local policy whose repeated execution produces adaptive coordination. This is a meaningful conceptual contribution. It's different from learning explicit roles, communication graphs, or routing policies.

Critical weaknesses

The evaluation is somehow weak. Most experiments use relatively structured coordination tasks with predefined local interfaces and interaction mechanisms.

Questions for authors
  1. How much task-specific engineering is required to construct the local interface and runtime for a new coordination environment? In particular, what modifications were necessary for the frozen AgentsNet transfer experiments?

  2. What are the contributions of the learned policy and the fixed runtime to self-organization? Although Figure 4 includes runtime ablations, it is still unclear to what extent the observed coordination structures are learned rather than enabled by predefined mechanisms.

  3. Does the theoretical analysis establish anything beyond symmetry preservation? Proposition 1 establishes swarm-level equivariance under an exactly equivariant policy and runtime, but this property alone does not guarantee successful coordination, convergence, or organizational stability. How do local equivariance errors propagate through repeated interactions, particularly as the population size and interaction horizon increase?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 1smB (1/2)

We thank the reviewer for highlighting evaluation breadth, the separation of learned and fixed mechanisms, and the interpretation of approximate equivariance. We clarify the transfer engineering and provide new outcome-based and closed-loop diagnostics below.

Response to W1. Evaluation Scope

Appendix I already demonstrates frozen transfer across AgentsNet graph tasks with different action vocabularies and structures.

To further evaluate open-ended coordination, we deployed the same frozen Waggle controller on Alem [1], covering Easy, Medium, and Hard difficulties in the official three-agent setting with five matched seeds each.

No Alem-specific training or checkpoint selection was performed, and Waggle shares a minimal adapter with the matched zero-shot local law (Table S0).

Table S0. New open-ended coordination evaluation on Alem

MethodCoordination reward (%) ↑Total return (%) ↑Messages / 100 decisions ↓
Alem native zero-shot LLM team3.03.068.4
Zero-shot local law (matched interface)2.52.856.7
Frozen Waggle5.34.243.5

Frozen Waggle achieves the highest coordination reward and total return among the evaluated methods, while sending 23.3% fewer non-empty messages than the matched zero-shot local law.

These findings extend the evidence for frozen-law reuse to open-ended, long-horizon coordination, with task-specific organizational decisions remaining under the learned controller. We will incorporate the evaluation and adapter details into the revision.

Response to Q1. Task-Specific Engineering

We appreciate this practical question about the engineering required to reuse a frozen controller across task environments.

  • Concrete AgentsNet adaptations. Frozen transfer to five graph tasks (Coloring, Consensus, LeaderElection, Matching, and VertexCover) required approximately 260 LOC of deterministic interface code:

    • Task/action mapping (~110 LOC): Map task contracts and native actions (e.g., color assignments, neighbor matching, and vertex selection) into the canonical decision interface.
    • Local-state and communication mapping (~60 LOC): Convert private observations and neighbor messages into bounded local views using opaque episode-local handles. For LeaderElection, local random priorities provide symmetry-breaking information that the frozen controller uses in its coordination decisions.
    • Output compilation (~90 LOC): Translate validated controller decisions into native task outputs, with legality checks handled by the adapter and task correctness assessed by the benchmark evaluator.
  • Frozen-policy reuse. The same learned controller was deployed across all five tasks without target-task training, fine-tuning, or parameter updates. The adaptations provide task-specific observation and action mappings; semantic coordination remains governed by the learned local law. We will document the implementation details in Appendix I.

Response to Q2. Learned Policy versus Fixed Runtime

We address this question through (1) a clear distinction between learned decisions and runtime mechanisms, and (2) an outcome-based ablation under controlled conditions.

  1. Distinct contributions. The learned law selects task actions, semantic messages, and commitment updates. The runtime provides validation, transport, conflict handling, expiry, and reversibility, without choosing their semantic content.

  2. New experiment: Outcome-based policy/runtime attribution. We evaluated matched late-conflict settings using frozen controllers and identical environment conditions. Unlike procedural recovery, these metrics do not require Challenge or Reopen events. Final-state correctness is the mean percentage of satisfied terminal checks (Table S1).

    Table S1. New outcome-based learned-policy and runtime ablation

    Method / runtimePost-conflict task success (%)Mean satisfied end-state checks (%)
    Decision Only + full runtime85.088.6
    Waggle + full runtime92.594.4
    Waggle without conflict handling77.579.2
    Waggle without reversible commitment75.078.1

    With the runtime fixed, Waggle improves task success by 7.5 pp and terminal-check satisfaction by 5.8 pp over Decision Only. Removing conflict handling or reversibility also reduces native outcomes. These comparisons support complementary contributions from learned decisions and runtime mechanisms, independently of the reopening-based metric.

AuthorsOpenReview ↗

Response to Reviewer 1smB (2/2)

Response to Q3. Theory and Error Propagation

We thank the reviewer for this thoughtful question about how local equivariance errors propagate to collective coordination outcomes. We address it below through a finite-horizon theoretical bound and closed-loop error-propagation experiments.

Proposition 1 establishes a structural compositional guarantee under exact equivariance. We further derive a conditional finite-horizon implication: with a uniform per-decision mismatch bound ϵ\epsilon and at most AA local decisions, the success-probability difference between relabeled executions is bounded by min⁡(1,Aϵ)\min(1,A\epsilon), assuming an equivariant runtime and relabeling-invariant success criteria.

To examine this relationship empirically, we evaluated paired relabeled executions across increasing population sizes and interaction horizons, tracking disagreements in local decisions, collective commitments, and final task outcomes (Table S2).

Table S2. New experiment: Propagation from local equivariance errors to coordination outcomes

Population / horizonExact admitted-record disagreement (%) ↓Commitment-state disagreement (%) ↓Task-outcome difference (%) ↓
N = 16, R = 98.25.22.5
N = 32, R = 98.96.55.0
N = 64, R = 99.68.17.5
N = 64, R = 1810.510.010.0

Local decision disagreements increase with population size and interaction horizon, accompanied by greater commitment-state and task-outcome sensitivity. Nevertheless, commitment-state and task-outcome differences remain within 10% in the evaluated settings. These results characterize the practical propagation of approximate equivariance errors and complement the conditional finite-horizon analysis.

We will incorporate the task-level implication and clarify its distinction from temporal organizational stability in the revision.

Summary and Planned Revisions

Frozen AgentsNet transfer demonstrates reuse without retraining. The new experiments clarify learned-policy versus runtime contributions and connect approximate equivariance to coordination outcomes.

We will document the adapter implementation, incorporate both experimental evaluations, and extend the theoretical discussion with the conditional finite-horizon outcome bound. We appreciate the reviewer's technically grounded questions, which have helped us strengthen the connection between our theoretical analysis and empirical evidence.

References

[1] Tessera et al. (2026). Benchmarking Open-Ended Multi-Agent Coordination in Language Agents. NeurIPS 2026, Evaluations and Datasets Track.

Thank you for the clarifications and for providing the additional experiments. I maintain my original rating.

Reviewer_KDwtOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper is related to the organization of multi-agent LLM swarms towards solving problems. Prior work typically organizes swarms globally by imposing hierarchies, explicit communication rules/topologies, or roles for agents. The paper studies organization that emerges from rules/laws that are strictly local, shared across all agents and anonymous. It proposes a methodology to learn such a local law by distilling a teacher’s outputs, along with auxiliary learning objectives for consistency under permutations of received messages, and prediction of next-step coordination state. The learned local law is evaluated in two swarm benchmarks on task-success rate, message count and shown to outperform ablations and baselines including no communication, transient networks, hand-coded local rules, and other global strategies for organizing swarms.

Critical strengths
  1. Strong motivation to omit explicit global structures and study the organizational capacity of local rules in agent swarms
  2. Careful design of communication interface to ensure anonymity and locality
  3. Thorough experiment design and evaluation to systematically verify generalization over population sizes, interaction horizon, transfer to substrates; study the benefit of orbit and field supervision; and attribution of recovery under counter-evidence to decision processes.
Critical weaknesses
  1. Source of teacher’s decisions used for training (Sec 4.1) are unclear. Does the teacher have access to privileged information? How does using the teacher directly perform on swarm organization metrics and task performance?
  2. Difficult for a general AI/ML audience to understand without specific familiarity in multi-agent swarms. Some symbols in equations are undefined, used before definitions. Examples of some terms include:
  • Transient-exchange
  • Bounded semantic update
  • Anonymous-orbit consistency
  • Substrate / runtime
  • “decision type” in L182 (vs task-level action, joint decision)
  • TTL in L208
  • “readiness” in L280
  • \phi, c appear in Eq 1 but defined only later
  • Relation between Decode, \pi, h
  1. Evaluation could consider tasks with global runtimes/substrates to study the effect of substrate locality in the success of local laws.
  2. Limitations of the approach/analysis could be discussed in more depth.
  3. Minor points:
  • Baselines could be explained more thoroughly under Sec 5.1, instead of related work.
  • Could benefit from more analysis of differences between global rules and local rules
Questions for authors
  1. Are there ways to alter the swarm's organization characteristics if the teacher is changed?
  2. In what scenarios are local laws expected to function well? Alternately, are there fundamental drawbacks established from prior work that limit the success of local laws compared to global laws?
  3. How do global laws transfer to new domains or recover under counter-evidence?
3 comments
AuthorsOpenReview ↗

Response to Reviewer KDwt (1/2)

We thank the reviewer for the thoughtful questions about teacher supervision, presentation clarity, and local versus global coordination. Their feedback has helped us sharpen the presentation and strengthen the empirical evaluation, as reflected in the direct teacher comparisons and two new targeted experiments below.

Response to W1. Teacher Decisions

  • Teacher information. GPT-4o receives the same bounded local view as the student, without global rosters, non-neighbor states, or privileged benchmark solutions.

  • Direct teacher performance. Under the identical interface and runtime (Appendix M.1), the GPT-4o teacher outperforms Waggle, while the 4B student retains strong coordination performance, including late-conflict recovery. Q1 further examines how teacher preferences shape learned organizational behavior.

Response to W2. Presentation and Definitions

We thank the reviewer for these careful and detailed suggestions.

  • Definitions and notation. We will define transient exchange, bounded semantic update, anonymous-orbit consistency, substrate/runtime, TTL, and readiness at first use; introduce ϕ\phi and cc before Eq. (1); and clarify the relationship among Decode, π\pi, and hh.

  • Running example. We will add a compact example tracing private evidence and incident messages through a joint local decision, including action, communication, commitment, and subsequent revision under counterevidence.

Response to W3. Global Runtime Evaluation

To examine the effect of global information access, we compared strictly local Waggle, frozen Waggle with a global blackboard, and a global-state coordinator on the primary SILO-Bench N=64 evaluation panel. All settings use matched tasks and seeds. Blackboard information enters through the existing local-view fields, without changing Waggle's controller parameters or decision architecture (Table K1).

Table K1. Global-information comparison on the primary SILO-Bench N=64 evaluation panel.

SILO-Bench (N = 64, R = 9)Waggle (Local)Frozen Waggle + Global BlackboardGlobal-State Coordinator
Level I87.688.289.1
Level II69.972.874.5
Level III17.125.129.3
Overall SR (%)58.262.064.3

Global information provides additional coordination gains, especially on Level III tasks. Frozen Waggle improves by 3.8 pp with blackboard access, while the strictly local controller already achieves 58.2% success. This demonstrates the effectiveness of the reusable local law without centralized state and its ability to benefit from richer observations without retraining.

We will clarify these information-access trade-offs in the revised paper.

Response to W4. Limitations

  • Empirically observed trade-offs. The additional evaluation characterizes the effect of expanding the information budget. Waggle remains effective under the original strictly local setting, while access to global information provides an expected complementary benefit without changing the learned controller.

  • Broader limits. Waggle relies on informative local observations, connectivity, and substrate-specific adapters; the analysis does not guarantee convergence or Byzantine robustness. We will integrate these empirical and theoretical limitations into Section 6.

Response to Minor Points

We will expand Section 5.1 with a concise baseline taxonomy distinguishing learned topology, routing, roles, and shared local laws by their information access and decision mechanisms. We will also clarify the differences between local and global rules: local laws coordinate through bounded neighborhood information and repeated local decisions, whereas global rules leverage system-wide information for centralized coordination. We will discuss their respective information requirements, coordination mechanisms, and limitations.

AuthorsOpenReview ↗

Response to Reviewer KDwt (2/2)

Response to Q1. Changing the Teacher

We thank the reviewer for this thoughtful question about how teacher supervision shapes organizational behavior, which motivated our targeted experiment.

  • New experiment: Teacher-preference steering. We changed only GPT-4o's local decision instructions to encourage either less redundant communication or more conservative commitment. The Qwen3-4B student, training views and transitions, SCD objectives, 648 updates, runtime, and evaluation settings were held fixed. We evaluated 200 matched late-conflict settings with five paired training seeds (Table K2).

    Table K2. New teacher-preference steering experiment

    Teacher preferenceFirst-commit latency (rounds)Post-conflict msg./agent ↓Recovery (%) ↑Stale-field AUC ↓
    Standard3.86.73788.519.237
    Communication-sparing4.05.9884.520.5
    Conservative-commit4.96.5086.519.6

Communication-sparing supervision reduces post-conflict messages by 11.2%; conservative-commit supervision delays initial commitment by 1.1 rounds. Both changes follow the teacher's intended preferences, with modest recovery trade-offs. Thus, teacher instructions can steer measurable organizational behavior without changing the local-law architecture.

Response to Q2. Applicability and Fundamental Limitations

  • Where local laws work. Local laws are particularly suitable when coordination depends primarily on neighborhood evidence and can be refined through repeated local exchanges.

    For example, neighboring agents competing for the same resource can exchange information, resolve conflicting claims, and revise commitments without a global coordinator. Appendix I further demonstrates frozen transfer across five AgentsNet tasks without retraining.

  • Fundamental limitations. Distributed computing theory identifies two constraints: symmetry, which prevents deterministic leader election in fully symmetric anonymous networks without symmetry-breaking information [1, 2], and information locality, which delays access to distant information [3]. Consistent with these limits, random priorities improve our LeaderElection score from 0.50 to 0.80, while global-state access improves Level III success from 17.1% to 29.3% (W3).

Response to Q3. Global-Law Transfer and Recovery

  • Cross-domain transfer. Global laws can transfer by adapting their global-state representations, task constraints, and assignment or routing mechanisms to a new domain. A reusable global coordinator may retain its decision strategy, but domain changes can require new interfaces, constraint definitions, or retraining. In contrast, Waggle reuses the same frozen local controller with task-specific adapters, as demonstrated across five AgentsNet tasks (Appendix I).

  • Recovery under counter-evidence. Global coordinators can incorporate new evidence by revising centralized assignments, routing, or topology and propagating updated decisions to agents. This requires timely global-state updates and potentially system-wide recomputation. Waggle instead revises commitments through repeated local exchanges without reconstructing a global organization. Our W3 experiment evaluates global information access, but not matched cross-domain transfer or counter-evidence recovery for the global coordinator.

Summary and Planned Revisions

The new experiments show that global information improves coordination performance and that teacher preferences can steer local-law behavior. We will add these results, revise the notation and examples, and sharpen the discussion of applicability, global/local trade-offs, and theoretical limitations. We particularly appreciate the reviewer's thought-provoking questions, which have helped us examine the broader implications of our approach and improve the depth of our analysis.

References

[1] Angluin, D. (1980). Local and Global Properties in Networks of Processors. STOC.

[2] Yamashita, M., & Kameda, T. (1996). Computing on Anonymous Networks: Part I—Characterizing the Solvable Cases. IEEE Transactions on Parallel and Distributed Systems.

[3] Linial, N. (1992). Locality in Distributed Graph Algorithms. SIAM Journal on Computing.

Thank you for the additional results and discussion. These address my concerns about teacher steering, global substrates, limitations. I think the paper makes a compelling case that local laws are effective. However, I will keep my score, because I still think the paper is hard for a broader audience to follow. Furthermore, comparing performance against global-organization approaches for domain-transfer, recovery, communication efficiency, could be necessary.

Reviewer_c4jvOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Waggle, a parameter-shared anonymous local policy through which LLM agents jointly choose task actions, local communication, and commitment updates. It introduces Swarm-Consistent Distillation, combining identifier-relabeling consistency with rollout-based prediction of the next local coordination field. The method is evaluated across spatial and language-based coordination benchmarks, population sizes, transfer tasks, and counterevidence interventions.

Critical strengths
  1. The paper articulates and instantiates a coherent alternative learning target for multi-agent organization. Sections 3.1–3.4 define a shared anonymous local interface, joint decision schema, bounded coordination traces, and commitment updates, rather than merely describing decentralized interaction at a high level. This makes the distinction between a learned local law and an explicit global routing or role structure operational and testable.

  2. The empirical mechanism analysis is stronger than typical LLM-agent evaluations. Table 3 separates the orbit and field objectives using a paired four-arm factorial, showing that SCD increases held-out summary agreement from 72.4% to 95.2%, improves eventful frozen-probe loss, and raises late-conflict recovery from 80.0% to 88.5%; Table 15 reports paired intervals for these effects. Figure 4 additionally tests fixed-runtime decision sources and runtime-component removals, which provides meaningful evidence that the observed reorganization behavior depends on both learned decisions and the runtime mechanisms.

Critical weaknesses
  1. The headline comparison against alternative approaches to LLM-agent organization is not fully attributional because the strongest baselines are evaluated under materially different system interfaces. Section 5.1 explicitly states that broader baselines retain their “native information-flow and private-computation interfaces,” whereas Waggle receives a carefully designed canonical interface containing adapter-produced private proposals, bounded persistent traces, validator-enforced provenance and legality, readiness conditions, expiry, duplicate suppression, and substrate-specific compilation (Sections 3.1–3.4; Table 7). Thus, the strong margins in Table 1 and Table 2 cannot cleanly be attributed to learning an anonymous local law rather than to the favorable interaction abstraction and runtime supplied to Waggle. Figure 4 is useful but does not close this gap: it compares Waggle with decision-only, zero-shot, and heuristic laws under the same runtime, rather than implementing a strong topology/routing baseline under the canonical interface and fixed runtime. This matters because the claimed advantage over explicit organizational approaches is central to the paper’s significance. A convincing resolution would be an interface-matched implementation of at least one strong learned topology or routing baseline, with the same private proposals, validators, interaction horizon, and message constraints, or a substantially more careful qualification that Table 1 compares complete systems rather than isolating organization-learning paradigms.
Questions for authors
  1. [Clarification] What population sizes and task configurations were used while collecting the decision corpus and rollout-transition corpus in Appendix F, and were all of N∈{16,32,64}N \in \{16,32,64\} seen during training? This determines whether the primary scaling results in Table 1 demonstrate extrapolation in population size or generalization across held-out episodes at trained sizes.
  2. [Clarification] In light of Section 5.1’s statement that broader baselines retain native information-flow and private-computation interfaces, was it feasible to implement a learned topology/routing baseline under Waggle’s canonical interface and fixed runtime? An interface-matched result for the strongest such baseline would materially increase my confidence that the gains in Tables 1–2 arise from the learned local law rather than the complete Waggle stack.
1 comment
AuthorsOpenReview ↗

Response to Reviewer c4jv

We thank the reviewer for the constructive questions about attribution, baseline fairness, and population generalization. We conducted an interface-matched routing experiment and clarify the training populations below.

Response to W1 and Q2. Attribution and Interface-Matched Routing Evaluation

Tables 1–2 compare Waggle with established organizational approaches under their documented native interfaces, preserving their respective topology, routing, and agent-coordination mechanisms. Separately, Figure 4(a) and Appendix M.1 already isolate the contribution of the learned decision law under an identical interface and fixed runtime.

To further examine whether Waggle's advantage extends beyond routing decisions, we evaluated two additional controls under the same canonical local view, private proposals, validators, runtime, and communication constraints:

  • Fixed Routing: A development-tuned, non-learned local routing policy.
  • AgentNet-inspired Local Routing [1]: A Qwen3-4B + LoRA controller trained to make Forward/Split/Execute-inspired local routing decisions, without global topology or non-local information.

Both controls share a fixed non-routing semantic policy, and the learned router uses matched training sources and update budgets. All methods are evaluated on matched test cases and seeds (Table C1).

Table C1. New interface-matched routing comparison

MethodSILO-Bench SR (%) ↑SwarmBench P/F quality ↑Late-conflict recovery (%) ↑
Fixed Routing35.41.0067.5
AgentNet-inspired Local Routing43.71.1075.0
Decision Only56.91.1980.0
Waggle58.21.23388.5

Learning local routing improves SILO-Bench success over Fixed Routing by 8.3 pp (paired 95% CI [+0.9, +15.7]). Waggle further improves over learned local routing by 14.5 pp (paired 95% CI [+5.4, +20.6]), with corresponding gains in task quality and late-conflict recovery.

These results complement the original system-level comparisons in Tables 1–2, where existing methods retain their native organizational capabilities. Together with Figure 4(a) and Appendix M.1, they provide evidence for Waggle's system-level effectiveness and the distinct value of learning joint semantic coordination beyond routing alone.

We will clarify these complementary evaluation purposes and incorporate the new routing controls into the revision.

Response to Q1. Training Population Sizes

We appreciate this important question, which helps clarify whether the scaling results reflect held-out generalization or population extrapolation.

  • Exact training populations. Synthetic decision views are population-agnostic. Native decision views and rollout-transition data were collected at N = 32, covering SwarmBench Pursuit/Foraging and SILO-Bench. Appendix F reports 6,912 decision views and 14,256 transition examples.

  • What scaling demonstrates. Training included neither N = 64 nor N = 128. Accordingly, N = 32 evaluates held-out generalization at the trained population size, N = 16 evaluates smaller-population transfer, and N = 64/128 provides population-size extrapolation with the frozen controller. We will explicitly distinguish these settings in the revised evaluation.

Response to E1. Ethics and Deployment Safeguards

We agree that bounded traces and reversible commitments do not establish authenticated provenance or Byzantine robustness. We will strengthen the Ethics Statement to explicitly require provenance authentication, sandboxed tools, sensitive-input protection, and human approval before consequential external actions.

Summary and Planned Revisions

The new matched-interface control provides direct evidence for the learned semantic law, while the training-population clarification establishes that N = 64/128 evaluations exceed native training scales. We will distinguish system-level from controlled comparisons, document population coverage, and strengthen the deployment safeguards.

We appreciate the reviewer's thoughtful feedback, which has helped us make our experimental comparisons more rigorous and our conclusions more precise.

References

[1] Yang, Y., Chai, H., Shao, S., Song, Y., Qi, S., Rui, R., & Zhang, W. (2025). AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems. NeurIPS 2025.

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.