acceptodds
Under review as a conference paper at ICLR 2027

Theory of Scene: Breaking the Symmetry Trap in Multi-Agent LLM Coordination

Abstract

Multi-agent systems built on large language models (LLMs) are largely homogeneous, as their agents behave alike even across distinct LLMs. We show that when such agents act concurrently without communication, they collide on targets they must split and diverge on targets they must take together, a double failure we term the symmetry trap. Theory of Mind (ToM), widely used for coordination without communication, cannot escape this trap, since homogeneous agents form the same prediction of one another and respond to it in the same way. We propose Theory of Scene (ToS), a training-free reasoning schema in which each agent reads its public role, the only difference between the agents, and the task context they all observe. Homogeneous agents thereby derive one division of labor, each taking the part its role fixes, which turns homogeneity from the cause of the trap into the cure. ToS reads the role together with the scene through role gating, which determines whether ownership overlaps or is already divided, and the task context through task coupling, which infers whether the team must converge on each target, divide it, or take its stages in turn. We evaluate on DivvyBench, a controlled environment we introduce, whose target types make an episode Competitive, Cooperative, or Mixed across Tabletop, Airspace, and Household scenarios, and on two established agentic benchmarks, GovSim and Overcooked. ToS outperforms all six baselines on every benchmark, and each baseline falls far behind it in at least one setting. Against ToM given the same inputs, ToS raises the DivvyBench success rate from 71.1% to 99.6%, the GovSim total gain from 207 to 400, and the Overcooked level-normalized throughput from 1.41 to 1.67.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent System, LLM Coordination

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_9EnoOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the symmetry trap in multi-agent LLM coordination, where homogeneous agents acting concurrently without communication may collide on targets requiring division or diverge on targets requiring joint action. The authors propose Theory of Scene (ToS), a training-free reasoning schema that derives a division of labor from shared observations and public roles rather than predicting teammates' intentions. ToS uses role gating to distinguish overlapping from already divided ownership, and task coupling to identify joint, exclusive, or sequential dependencies. Agents then select actions through a shared ordering and role-based allocation. Experiments on the newly introduced DivvyBench and adapted versions of GovSim and Collab-Overcooked show substantial improvements over six baselines. The paper also provides probability bounds characterizing coordination under shared and agent-specific action distributions.

Critical strengths
  1. Clear problem formulation and controlled evaluation. The symmetry trap captures two opposing coordination requirements: divergence on exclusive targets and convergence on joint targets. DivvyBench provides a controlled setting for evaluating both requirements within the same episode. ToS achieves nearly perfect completion and substantially outperforms the strongest baseline.

  2. Substantial empirical improvements across different coordination settings. The same ToS reasoning schema improves performance on both shared-resource allocation and sequential cooking tasks. Experiments with additional model backbones, heterogeneous agents, observation perturbations, and component ablations provide useful evidence of its effectiveness and robustness.

Critical weaknesses
  1. The contribution of the structured reasoning schema is not sufficiently isolated from the explicit coordination rules. The ToS prompt directly instructs agents to use a canonical ordering and role-based allocation for different task couplings. Although Table 2 examines the role and task readings, it does not isolate the benefit of these intermediate reasoning fields from that of the supplied coordination rules. Table 3 compares direct action derivation with prediction-based action selection, but does not include a simpler, rule-matched direct-action baseline. Consequently, it remains unclear how much of ToS's advantage comes from its structured reasoning schema rather than the explicit coordination convention.
  2. The theoretical analysis does not establish a general limitation of Theory of Mind. The probability bounds characterize coordination under shared or agent-specific action distributions, but do not demonstrate that ToM inherently cannot escape the symmetry trap. In particular, a ToM policy could incorporate public roles and a shared allocation convention when predicting teammates' actions. Although the evaluated ToM baselines perform poorly, these results establish limitations of the tested reasoning schemas rather than a general impossibility for ToM. The abstract's categorical claim that ToM "cannot escape this trap" therefore appears stronger than the theoretical and empirical evidence warrants.
Questions for authors
  1. How does ToS compare with a compact baseline that receives the same shared observations, public roles, and coordination rules, but proceeds directly to planning and action selection without explicitly generating separate role and task readings? This comparison would help isolate whether the intermediate reasoning structure provides benefits beyond the supplied coordination convention.
  2. Is the claimed limitation of ToM intended to apply only to the evaluated prediction-based reasoning schemas? Given that the theoretical bounds do not exclude role-dependent action distributions, what assumption would prevent a role-aware ToM policy from predicting complementary or matching assignments under a shared convention?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 9Eno

We thank the reviewer for the careful and constructive review.

W1 & Q1. The Schema and the Stated Rules

We have run three variants, (i) ReAct + Coupling Rules, the rule-matched direct-action baseline in W1, (ii) ToS (Reasoning without Rules), the intermediate reasoning fields without the coordination rules in W1, and (iii) ToS (Compact Baseline), the compact baseline in Q1 that proceeds directly to plan and action without the role and task readings. None of the three variants reaches ToS across the three benchmarks, so the advantage of ToS needs all three parts below.

  • (i) Explicit Coordination Rules. The rules state how the team responds to each task coupling, so an agent can apply them only once it knows the task coupling of each target.
  • (ii) Intermediate Reasoning Fields. The fields infer the ownership and the task coupling of each target from the observation, and without the rules the agents know the scene but not what to do with it.
  • (iii) Explicit Reasoning Structure. Writing the two readings in the output makes the agent infer them at every step, and without the fields this inference is not guaranteed.
MethodDivvyBench Competitive Success (%)DivvyBench Cooperative Success (%)DivvyBench Mixed Success (%)GovSim Survival (%)GovSim Total GainGovSim Survival (%), thinking onGovSim Survival (%), 20 agentsOvercooked Normalized Throughput
ReAct40.796.756.726.515548.68.41.00
ReAct + Coupling Rules (Reviewers hBtU, 9Eno)96.795.395.327.616765.88.61.17
ReAct + Stated Convention (Reviewer 9JF5)98.7100.098.732.318048.98.81.07
ReAct + Explicit Convention (Reviewer jbte)92.0100.088.733.319457.89.51.04
ToS (Reasoning without Rules)96.770.753.341.322381.716.31.34
ToS (Compact Baseline)98.097.392.773.833099.246.41.40
ToS99.3100.099.385.240099.276.11.67

In the revised paper, we have added ReAct + Coupling Rules to Appendix A.3 (Figure 4) and ToS (Compact Baseline) to Table 2 as w/o fields.

W2. The Propositions and the Abstract's Claim about ToM

We want to clarify that the propositions and the abstract's sentence about ToM are two separate claims. The propositions bound the per-step probabilities of any policy, ToM included, and Proposition 2 shows the importance of the public role, so a role-aware ToM is a potential solution under them. The abstract's sentence is an empirical claim about prediction-based schemas, in which a homogeneous agent's prediction of the others returns only its own reasoning (Section 3).

We have also run the policy the reviewer describes, Role-Aware ToM under a Shared Convention. The public role and the convention raise it above Classic ToM on DivvyBench, and it remains below ToS on all three benchmarks, because its action still depends on a prediction of the other agents, whereas ToS derives its action from the scene alone (Q2).

MethodDivvyBench Competitive Success (%)DivvyBench Cooperative Success (%)DivvyBench Mixed Success (%)GovSim Survival (%)GovSim Total GainOvercooked Normalized Throughput
ReAct + Coupling Rules96.795.395.327.61671.17
Classic ToM72.786.754.032.22071.41
Classic ToM + role + task88.768.778.739.12471.66
Role-Aware ToM under a Shared Convention95.392.080.731.02041.36
ToS99.3100.099.385.24001.67

In the revised paper, we have revised the abstract's statement and the discussion of Table 3 in Section 5.2.

Q2. A Role-Aware ToM under a Shared Convention

We want to clarify that our claim is empirical and restricted to schemas that derive their action from a prediction of homogeneous agents, the more challenging and practical setting, as LLM agents behave alike even across distinct LLMs. In W2, we have shown that a role-aware ToM can predict matching or complementary assignments under a shared convention and remains below ToS on all three benchmarks. We therefore report the prediction accuracy and the stuck rate of the ToM schemas, where the role-aware ToM reaches the highest prediction accuracy and the lowest stuck rate of the ToM methods, yet its stuck rate remains above that of ToS, because its action still depends on the prediction.

MethodPrediction Acc., Competitive (%)Prediction Acc., Cooperative (%)Prediction Acc., Mixed (%)Stuck, Competitive (%)Stuck, Cooperative (%)Stuck, Mixed (%)
ReAct---48.72.526.9
Classic ToM31.091.555.035.68.227.1
ProAgent30.179.848.737.220.135.0
HypMinds31.374.544.939.526.630.9
Classic ToM + role + task40.780.670.224.020.117.4
Classic ToM + Coupling Rules51.995.261.024.74.422.0
Role-Aware ToM under a Shared Convention65.196.070.216.33.915.8
ToS---15.40.09.1

Dear Reviewer 9Eno,

We thank you again for your time and effort in reviewing our paper. In our response, we have conducted ablation studies on the reasoning schema and the coordination rules, including the rule-matched direct-action baseline (W1) and the compact baseline (Q1), benchmarked a role-aware ToM under a shared convention with its prediction accuracy and stuck rate (W2 & Q2), and revised the abstract's statement about ToM (W2).

With three days left in the discussion period (November 18, AoE), we would greatly appreciate hearing whether our response has addressed your concerns, and we would be glad to address any that remain.

Best regards,
The Authors

Thank you once again for your valuable comments on our submission. We have posted a summary of our rebuttal at the top of this page. With the discussion period ending on November 18 (AoE), we would greatly appreciate hearing whether our response has addressed your concerns, and if you are satisfied, we would be grateful for your consideration in updating the score. We sincerely look forward to your feedback.

Reviewer_jbteOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies homogeneous LLM agents (copies of one model with one prompt, differing only by a public role label) that act simultaneously without communication. It names a "symmetry trap": such agents collide on targets they must split (Competitive) and, once they are randomized or predicted apart, they also split on targets they must take together (Cooperative). The paper argues that Theory of Mind (ToM) prompting cannot escape this trap, because a homogeneous agent predicting its teammate just reproduces its own reasoning. The proposed fix is "Theory of Scene" (ToS), a training-free prompt schema with four fields (state → role → task → action). The role gating field decides whether ownership of the work is overlapping or already divided. The task coupling field labels each target Joint (converge on it), Exclusive (split it by role over a "canonical order"), or Sequential (each role works its own stage). Two propositions bound the per-step probabilities of divergence and convergence: under one shared distribution, and in terms of the total-variation distance between agents' distributions.

Critical strengths

The idea is simple and easy to understand. It is easy to reprosuce since full prompts was given

Critical weaknesses

ToS is effectively a prompted convention, which puts doubts on the novelty

The authors removed GovSim's communication phase and universalization hint, opened up the Overcooked kitchen, and removed the request action. Each is motivated, but the paper should say whether these choices were fixed before comparative results were seen. It should also report at least one unmodified-protocol result per benchmark (e.g., GovSim with the universalization hint given to every method), so readers can relate the numbers to the original papers.

Questions for authors

How was the ToS schema (Fig. 15) developed? Which benchmarks, scenarios, and episodes were used during prompt iteration, and how many versions were tried? Was any part of the evaluation held out from that process?

Can you report DivvyBench results with layouts randomized per episode (target names, counts, positions of Cooperative targets, odd numbers of Exclusive targets) and with target types left unlabeled?

How does a "ReAct + explicit convention" baseline (two sentences stating the converge-first and role-indexed split rule) compare with ToS on all three benchmarks?

Were the GovSim and Overcooked protocol modifications fixed before comparative results were seen? Can you report results under the unmodified protocols?

5 comments
AuthorsOpenReview ↗

Response to Reviewer jbte (1/2)

We thank the reviewer for the constructive and encouraging feedback.

W1. ToS and the Convention Methods

We frame a convention as the multi-user layer of a harness. A system prompt works when one agent follows it, whereas a convention works only when every agent, with no channel to the others, reaches the same division of the work. We want to clarify the novelty of ToS in three parts.

  • Convention Is a Typical Solution. Focal (EMNLP 2026 Findings) and CFD (NeurIPS 2026) are convention methods and two of our baselines in Table 1, and GovSim universalization (NeurIPS 2024) is one as well.
  • Ours Differs from Every Convention Below. Our convention addresses a challenge that is common in coordination and absent from the conventions below, targets of different kinds in one scene, competitive and cooperative targets in the same scenario. One response for the whole scene cannot serve both kinds, which is where Focal and CFD fail (Table 1).
  • ToS Is Reasoning + Convention. (i) The reasoning reads the key information from the scene. (ii) The convention turns that information into the decision. With the reasoning alone, the agents know the scene but not what to do with it, and they still fail to coordinate (Reviewer 9Eno, Q1). With the convention alone, the agents know the rule but not who they are or which target it applies to, so given to ReAct it brings ReAct close to ToS on DivvyBench, where the labels supply the information, and improves ReAct slightly on GovSim and Overcooked (Q3).
WorkConvention
Social Convention (Science Advances 2025)A population settles on one name through repeated interaction
GovSim (NeurIPS 2024)Universalization provides the share at which the resource still regrows, with the share given
Focal (EMNLP 2026 Findings)Select the option that stands out
CFD (NeurIPS 2026)Coordination-friendly definitions and a self-check that rejects any convention that could have been flipped
ToSAgents that differ only in their public role derive one division of labor over targets with conflicting requirements (Competitive / Cooperative)

In the revised paper, we have stated the convention of ToS explicitly in Section 4.2 and placed it against these convention methods in Related Work.

W2 & Q4. Protocol Changes

We fixed every protocol change before any comparative result was seen. We report one unmodified-protocol result per benchmark, GovSim with universalization given to every method, and Collab-Overcooked on its original layout. With universalization, every method survives longer, since it states the share that ToS derives on its own, and ToS remains the method that survives most often. We note that CFD collects a higher total gain (614) than ToS (587) in this setting, because CFD follows an answer directly once it is stated. Without universalization no answer is stated, and ToS leads CFD in both survival (85.2% vs. 59.2%) and total gain (400 vs. 342).

ProtocolMethodGovSim Survival (%)GovSim Total GainOvercooked Raw ThroughputOvercooked Normalized Throughput
Table 1ReAct26.51552.411.00
Focal20.21432.431.11
CFD59.23422.311.16
Classic ToM32.22073.091.41
ProAgent16.41173.081.36
HypMinds46.52623.131.49
ToS85.24003.671.67
OriginalReAct82.05161.201.00
Focal80.75121.111.06
CFD96.86141.261.24
Classic ToM70.44511.171.04
ProAgent52.22731.190.99
HypMinds79.24641.211.48
ToS99.25871.371.52

In the revised paper, we have added GovSim with universalization given to every method to Appendix A.3 (Figure 5).

AuthorsOpenReview ↗

Response to Reviewer jbte (2/2)

Q1. How the Schema Was Developed

We developed the schema in three stages but did not count the versions. Airspace, Household, GovSim, and Overcooked were first run after its text was frozen, and every experiment in this response runs it unchanged.

  • (i) Probe. We tested the coordination idea on the pirate game (five pirates dividing 100 gold coins) as a probe of the coordination.
  • (ii) Iteration on Tabletop. We built DivvyBench Tabletop, which covers the Competitive, Cooperative, and Mixed settings, and iterated there from the intuition that every scene admits at least one rational division of labor to the role and task readings.
  • (iii) Generalization. The Tabletop text contained words such as light, heavy, and ball, so we rewrote it in generic terms and polished and extended the wording, the version in the paper, which names no scenario or target (Fig. 17 in the revised paper).
MethodTabletop, DevelopmentAirspace, Held OutHousehold, Held OutGovSim Survival (%), Held OutOvercooked Normalized Throughput, Held Out
ReAct52.777.364.026.51.00
Focal32.762.044.720.21.11
CFD9.364.786.059.21.16
Classic ToM64.778.070.732.21.41
ProAgent39.366.070.016.41.36
HypMinds36.772.066.746.51.49
ToS100.098.7100.085.21.67

Q2. Randomized Layouts and Unlabeled Types

Randomized Layouts. We have run 1800 random DivvyBench layouts, random in four ways: (i) random split between Competitive and Cooperative targets, (ii) random positions of the Competitive and Cooperative targets, (iii) random names of the targets, and (iv) random order of the targets.

Success rate (%) within 10 steps for layout of Competitive / Cooperative (1800 distinct instances):

Method0 / 81 / 72 / 63 / 54 / 45 / 36 / 27 / 18 / 0Avg.
ReAct88.089.078.067.055.549.045.030.033.059.4
Focal0.035.023.521.023.029.041.052.540.029.4
CFD8.021.018.011.515.515.019.520.561.021.1
Classic ToM53.556.062.059.553.052.054.566.566.058.1
ProAgent50.554.555.041.532.534.542.051.056.546.4
HypMinds19.037.541.539.545.540.048.547.069.543.1
ToS99.590.598.093.093.592.087.591.598.593.8

Unlabeled Types. We have also run DivvyBench Tabletop Mixed with the target types unlabeled. The state gives each target's weight in kg instead of its kind, and the rules state the 5 kg threshold once, so the kind of every target has to be inferred.

MethodLabeled Mixed Success (%)Unlabeled Mixed Success (%)
ReAct48.054.0
Focal48.04.0
CFD12.012.0
Classic ToM64.060.0
ProAgent18.016.0
HypMinds40.042.0
ToS100.088.0

In the revised paper, we have added the randomized-layout results to Appendix A.8 (Table 12).

Q3. ReAct + Explicit Convention

We have run the baseline the reviewer describes, ReAct + Explicit Convention, on all three benchmarks. On DivvyBench it brings ReAct close to ToS. GovSim provides no convention, and ToS outperforms all three convention baselines there in every setting, including thinking on and 20 agents. The explicit convention is hand-written for one scene, as GovSim's universalization is, whereas ToS reads from the scene which rule applies, with one text for all three benchmarks, so on GovSim, ReAct needs universalization (82.0%) to reach the level of ToS w/o universalization (85.2%), as in W2 & Q4. A hand-written convention has two limits that ToS does not:

  • Knowledge. Someone has to read the scene and write the rule, the explicit convention for DivvyBench, the sustainable share for GovSim.
  • Generality. The rule holds only in the scene it was written for. The explicit convention names nothing the agents can act on in GovSim, and universalization names nothing in DivvyBench.
MethodDivvyBench Competitive Success (%)DivvyBench Cooperative Success (%)DivvyBench Mixed Success (%)GovSim Survival (%)GovSim Total GainGovSim Survival (%), thinking onGovSim Survival (%), 20 agentsOvercooked Normalized Throughput
ReAct40.796.756.726.515548.68.41.00
ReAct + Coupling Rules (Reviewers hBtU, 9Eno)96.795.395.327.616765.88.61.17
ReAct + Stated Convention (Reviewer 9JF5)98.7100.098.732.318048.98.81.07
ReAct + Explicit Convention (Reviewer jbte)92.0100.088.733.319457.89.51.04
ToS99.3100.099.385.240099.276.11.67

In the revised paper, we have added the comparison with a stated convention to Appendix A.3 (Figure 4), using ReAct + Coupling Rules.

Dear Reviewer jbte,

We thank you again for your time and effort in reviewing our paper. In our response, we have conducted ablation studies on the reasoning and the convention, including ReAct + explicit convention on all three benchmarks (W1 & Q3), benchmarked every method under an unmodified protocol on GovSim and Overcooked (W2 & Q4), reported the held-out evaluation (Q1), and tested randomized layouts and unlabeled target types (Q2).

With three days left in the discussion period (November 18, AoE), we would greatly appreciate hearing whether our response has addressed your concerns, and we would be glad to address any that remain.

Best regards,
The Authors

Reviewer_9JF5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Theory of Scene (ToS) addresses what the authors call the symmetry trap in multi-agent LLM systems. It is based on the pattern that when agents built on the same or similar LLMs act at the same time without communicating, they tend to make the same choice, so they collide on targets they should split between them. Fixes that push them apart, such as higher sampling temperature or Theory of Mind (where each agent predicts the others and responds), push them apart on targets they must take together, because when several options look equally good each agent breaks the tie on its own. ToS is a training free reasoning schema in which each agent reads the shared scene together with its public role. From these it decides whether the work is already divided, and whether each target needs all agents at once, one agent each, or agents taking stages in turn. It then derives a division of labor over a canonical ordering of the targets and uses its role to find its own part. This turns the agents’ similarity into an advantage, since agents reading the same scene arrive at the same division. The authors also introduce DivvyBench, a controlled environment with Competitive, Cooperative, and Mixed settings. They report ToS outperforms six baselines on DivvyBench and two established benchmarks.

Critical strengths
  • Training free, elegant solution to the identified symmetry trap
  • ToS needs even less information than the ToM methods (doesn’t rely on other agents past actions). It also scales better than ToM as agents are added, since it doesn’t predict each agent individually.
  • Quantitatively surpasses six baselines on three benchmarks with very notable improvements on DivvyBench and GovSim
Critical weaknesses
  • No closed frontier LLM comparison. On GPT-OSS-20B, for example, ReAct already averages 84.4% success on DivvyBench, compared with 64.7% on Qwen3.5-35B-A3B (Table 6). Stronger models may therefore narrow the gap between ToS and the baselines, and it would be valuable to see whether ToS’s advantage holds with them.
  • Main quantitative results are only measured with Qwen3.5-35B-A3B. Appendix results with other backbones are only on the authors’ own DivvyBench benchmark, not on GovSim or Overcooked. Since those two benchmarks provide the more independent evidence, extending the backbone experiments to them would considerably strengthen the claim that ToS generalizes across models.
Questions for authors

The authors disable thinking mode so the generation budget goes to the reasoning schema (Appendix C.1). Does ToS’s advantage over the baselines hold with thinking mode enabled? Longer reasoning might help the baselines derive a division on their own, or it might lead agents to different divisions under ToS.

How does ToS compare with a simple stated convention? A baseline prompt such as “Leader takes odd-numbered light targets, Follower takes even ones, and both take the first heavy target” might perform close to ToS on DivvyBench. The paper discusses conventions in related work but doesn’t test one directly.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 9JF5

We thank the reviewer for the positive assessment and the concrete suggestions.

W1. Closed Frontier Models

We have run Gemini 4 Argon, and as the reviewer anticipated, the gap closes on the standard settings, where Argon reasons out a convention by itself, e.g., for Competitive targets, the Leader takes from one end of the list and the Follower from the other. Argon is smart enough to reason out a convention, but it lacks one that all agents share, so the agents stop agreeing on the division of labor once the setting gets harder. We therefore add three harder settings, where the advantage of ToS returns, (i) DivvyBench with reversed observation, (ii) GovSim with 100 agents, and (iii) Overcooked, where throughput does not saturate, since it counts the dishes delivered within 50 steps.

SettingReActClassic ToMToS
DivvyBench Mixed Success (%), budget 6100.0100.0100.0
DivvyBench Mixed Success (%), Reversed, budget 636.048.7100.0
DivvyBench Mixed Success (%), Reversed, budget 898.0100.0100.0
DivvyBench Mixed Success (%), Reversed, 5 agents, budget 530.026.798.0
GovSim Survival (%)100.0100.0100.0
GovSim Survival (%), 100 agents8.38.3100.0
Overcooked Throughput5.635.275.93
Overcooked Normalized Throughput1.000.911.12

In the revised paper, we have added the Gemini 4 Argon results to Appendix A.4 (Table 7).

W2. Backbones on GovSim and Overcooked

We have extended Table 6 and run its three further backbones on GovSim and Overcooked, where ToS remains the best method on every backbone.

BackboneMethodGovSim Survival (%)GovSim Total GainOvercooked Normalized Throughput
Gemma-4-26BReAct55.33311.00
Focal14.11241.11
CFD83.65190.86
Classic ToM57.73550.97
ProAgent34.41660.82
HypMinds78.34700.99
ToS83.95211.22
GPT-OSS-20BReAct25.81781.00
Focal11.91131.05
CFD27.01830.95
Classic ToM20.31481.13
ProAgent22.11511.06
HypMinds23.71591.09
ToS32.92011.20
Qwen3.5-9BReAct15.21281.00
Focal9.81050.96
CFD25.21750.87
Classic ToM19.61430.82
ProAgent20.81400.69
HypMinds28.31730.87
ToS46.02711.17

In the revised paper, we have added every backbone on GovSim and Overcooked to Table 6.

Q1. Thinking Mode

We have run Qwen3.5-35B-A3B with thinking mode enabled, the generation budget set to 16384 tokens, and the max model length to 24576 tokens. As the reviewer expected, longer reasoning helps the baselines derive a division on their own on the standard DivvyBench settings. This costs more than 10x the output tokens per reply on the same DivvyBench prompts (e.g., ReAct 220 → 5,070, Classic ToM 318 → 5,139, ToS 253 → 5,439), whereas ToS reaches the same success in Table 1 with thinking disabled.

SettingReActClassic ToMToS
DivvyBench Competitive Success (%)100.0100.0100.0
DivvyBench Cooperative Success (%)100.098.0100.0
DivvyBench Mixed Success (%)94.086.0100.0
DivvyBench Mixed Success (%), 5 agents, budget 564.060.092.0
GovSim Survival (%)48.661.799.2
GovSim Total Gain302404621
Overcooked Throughput3.533.934.17
Overcooked Normalized Throughput1.001.281.31

In the revised paper, we have added the thinking-mode results to Appendix A.4 (Table 8).

Q2. A Stated Convention

We want to clarify that we tested two convention baselines, Focal (EMNLP 2026 Findings) and CFD (NeurIPS 2026), in Table 1. We have also run the baseline the reviewer describes, ReAct + Stated Convention, and as the reviewer expected, it performs close to ToS on DivvyBench. DivvyBench labels every target, so dividing the work is the whole task and a convention alone solves it, whereas GovSim and Overcooked first require working out what the work is, which is the reasoning of ToS, so there the stated convention stays close to ReAct. GovSim provides no convention, and ToS outperforms all three convention baselines there in every setting, including thinking on and 20 agents.

MethodDivvyBench Competitive Success (%)DivvyBench Cooperative Success (%)DivvyBench Mixed Success (%)GovSim Survival (%)GovSim Total GainGovSim Survival (%), thinking onGovSim Survival (%), 20 agentsOvercooked Normalized Throughput
ReAct40.796.756.726.515548.68.41.00
ReAct + Coupling Rules (Reviewers hBtU, 9Eno)96.795.395.327.616765.88.61.17
ReAct + Stated Convention (Reviewer 9JF5)98.7100.098.732.318048.98.81.07
ReAct + Explicit Convention (Reviewer jbte)92.0100.088.733.319457.89.51.04
ToS99.3100.099.385.240099.276.11.67

In the revised paper, we have added this comparison to Appendix A.3 (Figure 4), using ReAct + Coupling Rules.

Dear Reviewer 9JF5,

We thank you again for your time and effort in reviewing our paper. In our response, we have benchmarked ToS on Gemini 4 Argon (W1), on more model backbones across GovSim and Overcooked (W2), with thinking mode enabled (Q1), and against a stated convention baseline (Q2).

With three days left in the discussion period (November 18, AoE), we would greatly appreciate hearing whether our response has addressed your concerns, and we would be glad to address any that remain.

Best regards,
The Authors

I am happy with this response and appreciate you showing these additional results. It quells my concerns for the paper, I will retain my score.

Thank you again for the reply and for the suggestions, which have significantly improved the paper, and we are glad that the additional results have addressed your concerns.

Reviewer_hBtUOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies coordination among homogeneous LLM agents—using the same model and prompt and differing only in publicly visible role labels—when they cannot communicate and must act simultaneously. The authors argue that such teams fall into a “symmetry trap”: agents collide when objectives require them to split up, while symmetry-breaking mechanisms such as temperature sampling and Theory-of-Mind-style prediction cause them to disperse when objectives require joint action. The proposed Theory of Scene (ToS) is a prompting framework in which each agent first determines whether ownership is overlapping or divided, then classifies each objective’s coupling type as joint, exclusive, or sequential, and finally applies the corresponding rule over a canonical ordering of objectives. The authors introduce DivvyBench, featuring two agents, eight objectives, and a ten-step budget across competitive, cooperative, and mixed settings. They also evaluate ToS on modified versions of GovSim and Collab-Overcooked.

Critical strengths

DivvyBench cleanly isolates objective selection as the decision being evaluated. The theoretical minimum number of steps has a closed-form solution, and the paper provides curves showing success rates under different step budgets, making the results easy to interpret.

Critical weaknesses
  1. My primary concern is the practical value of this approach. Harnesses such as Claude Code and Codex are already standard in most agent deployments and also support multi-agent capabilities. Can the proposed method improve the performance of these mainstream harnesses? More broadly, do the paper’s conclusions remain generalizable under today’s mature and highly optimized harnesses?

  2. The prompting framework effectively gives ToS the answer. The ToS prompt in Fig. 15 does more than structure reasoning: it directly specifies the coordination convention. For joint objectives, all agents converge on “the first such target in that order,” while for exclusive objectives, agents “take your share by role over that order.” Moreover, DivvyBench prompts explicitly label objective types—light or heavy—and state their associated rules, so the task-coupling step largely amounts to reading these labels. Removing the role field has almost no effect on DivvyBench performance (97.3/100/99.3 versus 99.3/100/99.3; Table 2), which further supports this interpretation. None of the baselines in Table 1 receive the same coordination rule. Table 3 adds the rule to Classic ToM, but it remains embedded within a prediction–best-response framework. I therefore cannot determine whether the gains come from the reasoning framework itself or simply from explicitly providing the coordination convention.

  3. DivvyBench is very small in scale. The layout of each scenario is fixed (Appendix C.2), meaning that the 450 episodes in the main experiment correspond to only nine distinct instances, each sampled 50 times.

Questions for authors

N/A

5 comments
AuthorsOpenReview ↗

Response to Reviewer hBtU (1/2)

We thank the reviewer for the detailed and constructive review.

W1. Practical Value in Mainstream Harnesses

We want to show that our setting differs from the multi-agent modes of Claude Code and Codex, which assume a central orchestrator that assigns each agent its task and a channel over which the agents can talk at any time. Many deployments have neither, because the agents belong to different endpoints, as in the table below, where each harness completes the task alone, and the team gains parallelism but loses part of it to conflicts. Lowering those conflicts without an orchestrator or a channel is the motivation of the paper, and ToS enters any harness as prompt text with no training.

SettingAgentsTaskOrchestratorChannel
Claude Code / Codex multi-agent modeSubagents of one sessionOne task, split by the orchestrator✓✓
Software EngineeringDifferent engineers' Claude Code / CodexOne codebase, many files✗✗
Paper WritingDifferent authors' agentsOne paper, many sections✗✗
Personal AssistantDifferent friends' agents (Meta Muse, OpenAI Dots)One social circle, many users✗✗
DivvyBenchDifferent robot arms, UAVs, household robotsOne scene, many targets✗✗
OvercookedDifferent kitchen robotsOne dish, many stations✗✗

Codex CLI Experiments. We have also run ToS inside the Codex CLI itself. Engineer A and Engineer B are two runs of the Codex CLI on the same Qwen3.5-35B-A3B, and aim to address one repository of 8 random HumanEval questions. Each round both name one open task at the same time, and the only text that differs between the methods is the reasoning schema in the task prompt, the same text each method receives in DivvyBench. An Exclusive task has its acceptance cases in the repository, one engineer's pull request merges, and a second one on the same task is discarded. A Joint task has its acceptance cases from two stakeholders, and each engineer holds one stakeholder's cases, so a pull request by one engineer alone passes only half of them.

Success rate (%) within 10 steps for layout of Competitive / Cooperative (1800 distinct instances):

Method0 / 81 / 72 / 63 / 54 / 45 / 36 / 27 / 18 / 0Avg.
ReAct36.526.022.525.036.048.559.591.5100.049.5
Focal21.527.031.039.047.557.578.093.0100.054.9
CFD5.59.513.520.029.532.560.591.0100.040.2
Classic ToM1.53.06.017.538.562.087.599.5100.046.2
ProAgent1.50.55.523.537.068.093.599.5100.047.7
HypMinds1.00.56.513.025.555.579.098.5100.042.2
ToS31.543.052.067.071.583.593.599.0100.071.2

In the revised paper, we have added this motivation to the Introduction and the Codex CLI experiment to Appendix A.7 (Table 11).

W2a. The Coordination Convention of Each Method

We frame the convention of ToS as the multi-user layer of a harness. A harness gives one user's agent the answer to how its work is done, through its system prompt and skills, and it gives no answer to how the agents of different users divide that work, which is the answer ToS states.

We also want to clarify that the SOTA methods have different coordination mechanisms, as analyzed in the Introduction, Related Work, and Section 5.1. Focal (EMNLP 2026 Findings) and CFD (NeurIPS 2026) are convention methods, Classic ToM, ProAgent, and HypMinds are prediction methods, and ToS has its convention together with its reasoning. GovSim (NeurIPS 2024) reports universalization, which states the sustainable share itself, as a main finding, and we show that ReAct needs universalization (82.0%) to reach the level of ToS w/o universalization (85.2%), please see full comparison in our reply to Reviewer jbte, W2 & Q4. Although Focal and CFD are convention methods as well, they fail on DivvyBench (Table 1), which is the problem ToS addresses, and the advantage of ToS extends to two established benchmarks, GovSim and Overcooked.

MethodDivvyBench Competitive Success (%)DivvyBench Cooperative Success (%)DivvyBench Mixed Success (%)GovSim Survival (%)GovSim Total GainOvercooked Normalized Throughput
Focal (Most Obvious Target)47.342.050.020.21431.11
CFD (Non-Flippable Convention)57.366.736.059.23421.16
ToS (Joint, Exclusive, Sequential)99.3100.099.385.24001.67

In the revised paper, we have stated the convention of ToS explicitly in Section 4.2.

AuthorsOpenReview ↗

Response to Reviewer hBtU (2/2)

W2b. Gains from the Convention vs. Reasoning

To separate the two, we have run ReAct + Coupling Rules, in which ReAct is given the three convention rules of ToS verbatim. We show that the convention and the reasoning serve different parts of the task.

  • Convention. It divides the work. We built DivvyBench to isolate the symmetry trap (Section 3), labeling every target and removing everything else a team would have to work out, so the division of labor is the whole task, and there the convention alone closes most of the gap.
  • Reasoning. It works out what the work is before the convention divides it. GovSim and Overcooked ask for both. GovSim provides no convention, and ToS leads there in every setting, including thinking on and 20 agents. Removing the role field lowers ToS much there and little on DivvyBench (Table 2).
MethodDivvyBench Competitive Success (%)DivvyBench Cooperative Success (%)DivvyBench Mixed Success (%)GovSim Survival (%)GovSim Total GainGovSim Survival (%), thinking onGovSim Survival (%), 20 agentsOvercooked Normalized Throughput
ReAct40.796.756.726.515548.68.41.00
ReAct + Coupling Rules (Reviewers hBtU, 9Eno)96.795.395.327.616765.88.61.17
ReAct + Stated Convention (Reviewer 9JF5)98.7100.098.732.318048.98.81.07
ReAct + Explicit Convention (Reviewer jbte)92.0100.088.733.319457.89.51.04
ToS99.3100.099.385.240099.276.11.67

In the revised paper, we have added ReAct + Coupling Rules to Appendix A.3 (Figure 4).

W2c. Target Types without Labels

We have run DivvyBench Tabletop Mixed with the type labels removed. The state gives each target's weight in kg instead of its kind, and the rules state the 5 kg threshold once, so the kind of every target has to be inferred.

MethodLabeled Success (%)Unlabeled Success (%)
ReAct48.054.0
Focal48.04.0
CFD12.012.0
Classic ToM64.060.0
ProAgent18.016.0
HypMinds40.042.0
ToS100.088.0

W3. Scale of DivvyBench

We follow the protocol of GovSim (NeurIPS 2024) in DivvyBench, fixed scenes sampled repeatedly, with 9 scenes against 3 in GovSim. Each is sampled 50 times rather than once because the randomness of an agentic run compounds over its steps, so under one budget we spent the episodes on repeats of each instance.

Experiments with More Distinct Instances. We have also run 1800 random DivvyBench layouts, random in four ways: (i) random split between Competitive and Cooperative targets, (ii) random positions of the Competitive and Cooperative targets, (iii) random names of the targets, and (iv) random order of the targets. Together with the Codex CLI experiments in W1, these layouts provide large-scale evidence of the effectiveness of ToS.

Success rate (%) within 10 steps for layout of Competitive / Cooperative (1800 distinct instances):

Method0 / 81 / 72 / 63 / 54 / 45 / 36 / 27 / 18 / 0Avg.
ReAct88.089.078.067.055.549.045.030.033.059.4
Focal0.035.023.521.023.029.041.052.540.029.4
CFD8.021.018.011.515.515.019.520.561.021.1
Classic ToM53.556.062.059.553.052.054.566.566.058.1
ProAgent50.554.555.041.532.534.542.051.056.546.4
HypMinds19.037.541.539.545.540.048.547.069.543.1
ToS99.590.598.093.093.592.087.591.598.593.8

In the revised paper, we have added the randomized-layout results to Appendix A.8 (Table 12).

Dear Reviewer hBtU,

We thank you again for your time and effort in reviewing our paper. In our response, we have tested ToS and all baselines inside the Codex CLI (W1), conducted ablation studies on the reasoning framework and the coordination convention (W2), and evaluated every method on 1800 distinct DivvyBench instances (W3).

With three days left in the discussion period (November 18, AoE), we would greatly appreciate hearing whether our response has addressed your concerns, and we would be glad to address any that remain.

Best regards,
The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

Dear Area Chair and Reviewers,

We thank you for your time and effort in reviewing our paper. We summarize below the new experiments in our rebuttal, grouped by the concerns raised. The revisions are marked in blue in the revised paper.

Reasoning vs. Convention (hBtU W2, 9JF5 Q2, jbte W1/Q3, 9Eno W1/W2/Q1/Q2)

ToS is reasoning + convention, and we have conducted ablation studies on all three benchmarks that separate the two. None of the variants below reaches ToS, which shows that the contribution of ToS is the combination of the two.

  • Convention without Reasoning. ReAct is given the convention in three forms, (i) the coupling rules of ToS verbatim (hBtU, 9Eno), (ii) a stated convention (9JF5), and (iii) an explicit convention (jbte), and ToS is run as a compact baseline (9Eno Q1). On GovSim, ReAct under all three forms stays low (at most 33.3%), and ReAct needs the universalization convention (82.0%) to reach the level of ToS w/o universalization (85.2%).
  • Reasoning without Convention (9Eno W1). On all three benchmarks, ToS outperforms its variant without the coordination rules (on GovSim, 85.2% vs. 41.3%).
  • Role-Aware ToM under a Shared Convention (9Eno W2/Q2). This policy reaches the highest prediction accuracy of the ToM methods and remains below ToS on all three benchmarks (on GovSim, 85.2% vs. 31.0%).

ToS Improves Codex Harnesses (hBtU W1)

We frame the convention of ToS as the multi-user layer of a harness. A harness gives one user's agent the answer to how its work is done, but it gives no answer to how the agents of different users divide that work, which is the answer ToS states. The multi-agent modes of Claude Code and Codex have a central orchestrator and a channel, whereas many deployments have neither, because the agents belong to different users. ToS enables multiple Claude Code and Codex agents to work together (e.g., Engineer A's and Engineer B's Codex).

  • Codex CLI Experiments. Engineer A and Engineer B are two runs of the Codex CLI that address one repository of 8 random HumanEval questions. We have tested ToS and all baselines on 1800 distinct instances, where ToS has the highest average success rate (71.2% vs. 54.9% for the best baseline).

Robustness across Protocols, Scale, and Backbones (hBtU W2/W3, 9JF5 W1/W2/Q1, jbte W2/Q2/Q4)

We have extended our evaluation and shown that ToS has the highest performance on all three benchmarks. Parentheses give ToS vs. the best baseline in that experiment.

  • Protocols (jbte W2/Q4). GovSim with universalization given to every method (99.2% vs. 96.8%), and Collab-Overcooked on its original layout (1.52 vs. 1.48).
  • Scale (hBtU W3, jbte Q2). 1800 randomized DivvyBench layouts (93.8% vs. 59.4%).
  • Unlabeled Target Types (hBtU W2, jbte Q2). DivvyBench with the target types unlabeled (88.0% vs. 60.0%).
  • Backbones (9JF5 W1/W2). Gemini 4 Argon and three more backbones across GovSim and Overcooked (on Argon, GovSim with 100 agents, 100.0% vs. 8.3%).
  • Thinking Mode (9JF5 Q1). All three benchmarks with thinking mode enabled (on GovSim, 99.2% vs. 61.7%).

We hope this summary is helpful for the discussion, and we would be glad to address any remaining concerns.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.