acceptodds
Under review as a conference paper at ICLR 2027

Degrees of Freedom in LLM Multi-Agent Coordination: Between Centralized and Decentralized Regimes

Abstract

Large language model (LLM)-based multi-agent systems (MAS) are commonly organized around two coordination regimes. Centralized coordination provides explicit planning and control but can create a bottleneck at a single coordination locus, whereas decentralized coordination offers local autonomy but must achieve effective collaboration through distributed decisions and experience. These complementary trade-offs motivate a systematic study of the space between the two regimes. We introduce **Degrees of Freedom** (DoF), a continuous configuration framework that quantifies the size of the decision space delegated to downstream agents. DoF jointly controls the granularity of delegated work and, within each granularity regime, the target probability that each task unit is assigned to the decentralized branch. Its endpoints recover the two pure coordination regimes. Across multiple task domains, DoF reveals a hierarchy of organizational effects. Delegated granularity and within-regime routing produce structured but bounded profile reallocations, whereas task domain and retained experience are associated with stronger differences between the centralized- and decentralized-branch profiles. The resulting performance landscape is task-dependent: different domains favor different DoF configurations, while atomic-target conditions are generally more consistent across runs. In-domain retained experience also brings the two branch profiles closer together. The same controlled sweep reveals a task-dependent coordination landscape: some domains exhibit clearly higher-performing configurations, whereas others show flatter responses, and no single configuration dominates across domains. In four of the six evaluated state–dataset settings, an intermediate DoF configuration achieves higher mean accuracy than both pure endpoints, demonstrating the value of exploring the coordination space between them. Together, these findings provide empirical guidance for studying and designing intermediate coordination regimes in LLM-based MAS.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent Systems, Degree of Centralization, LLM Agents

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

The paper studies centralized and decentralized LLM multi-agent coordination through a configuration index controlling decomposition granularity and per-unit routing probability. Using MetaAgent and AgentNet with a shared model and five-agent roster, it evaluates 15 configurations across five benchmark collections. It reports task-dependent accuracy peaks, changes in agent-allocation profiles, and generally lower run-to-run variance under atomic delegation.

Critical strengths

The systematic sweep across several task domains provides useful descriptive evidence about coordination choices. Shared model and agent configurations, repeated runs, and matched-routing comparisons improve interpretability. The paper also appropriately acknowledges that intermediate accuracy peaks do not establish benefits from mixing branches.

Critical weaknesses

The comparisons leave routing effects difficult to distinguish from decomposition and execution differences. Intermediate configurations introduce Coordinator-level decomposition and aggregation that the endpoints bypass. For example, BBH accuracy increases from 0.217 at d=0 to 0.750 at d=0.0001, where routing remains almost entirely centralized. Appendix B.3 specifies different branch temperatures and execution limits, without reporting total token or model-call costs. Section 5.3 acknowledges these confounds, but they limit the design guidance supported by the descriptive findings.

The evidence for reliable configuration preferences and stability differences is limited. Evaluation subsets contain 90–168 items, with only two or three runs per condition. Table 1 reports observed ranges, without uncertainty estimates for paired accuracy differences or variance ratios. The highest observed mean among 15 configurations does not establish a reproducible preference. Appendix C.2 explains condition inclusion, but the variance aggregation remains unclear. One model and one pair of branch implementations further limit generality. These limitations motivate my weak-reject recommendation.

Questions for authors
  1. Which routing conclusions hold within a fixed granularity regime? A focused comparison with identical decompositions and aggregation, matched decoding, and reported computation would help clarify the performance differences.
  2. Can paired uncertainty analyses of existing task-level outcomes establish how reliably the preferred configurations outperform alternatives?
  3. How are per-condition variances aggregated and weighted, and how sensitive are the ratios to individual high-variance conditions?

I would consider raising my rating if additional analysis, focused controls, or appropriately narrowed claims make the performance and stability findings sufficiently reliable and informative.

4 comments
AuthorsOpenReview ↗

Response to Reviewer boPo

Thank you for the specific requests about routing controls, uncertainty, and variance sensitivity. We address the scope of each analysis directly below.

[W1, Q1] Which routing conclusions hold within fixed granularity, and what remains uncontrolled?

The supported conclusion is that workflow performance varies with routing within a granularity regime. It is not that mixed routing consistently outperforms decomposed single-branch execution.

We separate this conclusion from endpoint-to-intermediate changes. Those comparisons introduce Coordinator decomposition and aggregation and therefore cannot isolate branch mixing. The BBH example highlighted in the review is one such comparison. The following selected endpoint contrasts illustrate the associated workflow changes.

SettingDecomposed configurationUndecomposed endpointBootstrap 95% CINominal pp-value
BBHd=0.0001d=0.0001: 0.750d=0d=0: 0.217[0.436, 0.628]1.37×10−271.37\times10^{-27}
AIMEd=0.0001d=0.0001: 0.600d=0d=0: 0.619[-0.089, 0.054]0.608
MMLU-Pro, BBH-trainedd=0.0001d=0.0001: 0.824d=0d=0: 0.662[0.103, 0.222]9.33×10−89.33\times10^{-8}
MMLU-Pro-trainedd=0.0001d=0.0001: 0.839d=0d=0: 0.705[0.080, 0.188]6.14×10−76.14\times10^{-7}
R-Bench-Td=0.0001d=0.0001: 0.633d=0d=0: 0.558[0.014, 0.137]0.0178
SMFRd=0.0001d=0.0001: 0.411d=0d=0: 0.243[0.090, 0.264]8.19×10−58.19\times10^{-5}

Within a granularity regime, the requested unit scope, decomposition prompt contract, and aggregation strategy remain fixed while routing varies. The selected paired comparisons below identify routing-dependent differences in several settings; other comparisons remain unresolved within their uncertainty intervals.

SettingGranularitydAd_A vs. dBd_BAccuracy A / BBootstrap 95% CINominal pp-value
BBHAtomic0.5 vs. 0.00010.883 / 0.750[0.074, 0.195]1.70×10−51.70\times10^{-5}
AIMESubtask0.6 vs. 0.99990.667 / 0.581[0.030, 0.148]0.00524
MMLU-Pro, BBH-trainedAtomic0.5 vs. 0.00010.856 / 0.824[-0.001, 0.064]0.058
MMLU-Pro-trainedSubtask0.5001 vs. 0.99990.848 / 0.586[0.205, 0.318]3.27×10−193.27\times10^{-19}
R-Bench-TAtomic0.2 vs. 0.00010.704 / 0.633[0.022, 0.123]0.00625
SMFRSubtask0.7 vs. 0.99990.396 / 0.295[0.038, 0.163]0.00213

We did not cache and replay identical decompositions. The recovered R-Bench-T example at d=0.2d=0.2 and d=0.4d=0.4 contains 15 and 14 units. Both follow five broad stages, but differ locally. It supports comparable high-level decomposition structure rather than exact identity. The aggregate decomposition-scale summaries are:

SettingAtomic units, mean ± SDSubtask units, mean ± SD
BBH16.98 ± 0.044.49 ± 0.02
AIME19.87 ± 0.185.11 ± 0.02
MMLU-Pro, BBH-trained12.54 ± 0.083.70 ± 0.02
MMLU-Pro-trained12.46 ± 0.063.51 ± 0.46
R-Bench-T20.28 ± 0.095.30 ± 0.02
SMFR39.13 ± 0.217.67 ± 0.96

The decoding and execution settings also differ between branches. Coordinator and MetaAgent calls use temperature 0.3; AgentNet uses temperature 0 and top-p=1.0p=1.0. MetaAgent permits twelve planning rounds and one redo per step; AgentNet permits three forward hops and five executions per task. These settings are held fixed within each branch, not matched across branches.

Likewise, the shared 2,048-token output limit is per call, not a matched total-task budget. We have not established a cost-controlled routing advantage. The results concern assignment between these implemented workflows, including their execution behavior. Appendix B.3 gives the settings. We have incorporated the decomposition evidence into Appendix G.1 and the selected paired comparisons into Appendix G.2. Section 5.3 now distinguishes these findings from gains associated with the decomposed workflow.

[W2, Q2] What do the paired uncertainty analyses establish?

They quantify task-level uncertainty for the selected configuration contrasts. They do not establish that the highest observed configuration is a reproducible global optimum.

For each task ii, correctness is averaged across the observed runs at configuration dd, giving yi(d)y_i(d). We then calculate the paired difference

Δi=yi(dA)−yi(dB).\Delta_i=y_i(d_A)-y_i(d_B).

The task-level bootstrap resamples task identities while preserving the paired outcomes for each selected comparison. It therefore estimates uncertainty over tasks conditional on the observed runs. It does not replace additional independent runs or quantify all variability expected from future reruns.

The following table reports the selected intermediate max–min comparisons.

Settingdmax⁡d_{\max}, accuracydmin⁡d_{\min}, accuracyBootstrap 95% CINominal pp-value
BBH0.9999, 0.8870.5001, 0.747[0.078, 0.204]1.23×10−51.23\times10^{-5}
AIME0.6, 0.6670.9, 0.563[0.048, 0.163]4.65×10−44.65\times10^{-4}
MMLU-Pro, BBH-trained0.5, 0.8560.6, 0.818[0.012, 0.065]0.00570
MMLU-Pro-trained0.25, 0.8480.9999, 0.586[0.211, 0.312]2.78×10−232.78\times10^{-23}
R-Bench-T0.5001, 0.7300.0001, 0.633[0.048, 0.149]1.28×10−41.28\times10^{-4}
SMFR0.0001, 0.4110.9999, 0.295[0.045, 0.191]0.00168

We also distinguish the comparisons’ purposes. The intermediate max–min contrasts describe the observed sweep range. The endpoint contrasts examine decomposition-associated workflow changes. The within-granularity contrasts address routing sensitivity. Some intervals exclude zero; others do not. An interval containing zero leaves that contrast unresolved and is not evidence of equivalence.

The reported pp-values are nominal and unadjusted for the number of comparisons examined. The max–min configurations were selected after inspecting the sweep, so their pointwise intervals and pp-values should not be interpreted as selection-adjusted evidence for the best configuration. They answer the narrower question of how the selected observed outcomes differ.

Appendix G.2 reports these comparisons separately. The revised accuracy claim concerns promising observed intermediate configurations, while the proposed validation procedure separates configuration selection from held-out evaluation.

[W2, Q3] How are condition variances aggregated and weighted?

We first calculate the sample variance of accuracy across runs separately for each complete DoF condition:

\left(\mathrm{Acc}_{d,r}-\overline{\mathrm{Acc}}_d\right)^2.$$ We then average these condition variances with equal weight within each granularity regime: $$\mathrm{Var}_g=\frac{1}{\lvert\mathcal{D}_g\rvert}\sum_{d\in\mathcal{D}_g}s_d^2, \qquad \rho=\frac{\mathrm{Var}_{\mathrm{subtask}}}{\mathrm{Var}_{\mathrm{atomic}}}.$$ Thus, the statistic averages run-to-run variation within conditions; it does not pool accuracy values across different DoF settings. Each included condition has equal weight, and each regime is normalized by its own number of included conditions. The endpoints $d=0,1$ and conditions without their complete planned run cohort are excluded. The planned counts are three runs for each BBH-trained setting and two for the MMLU-Pro-trained setting. **How sensitive is the ratio to individual high-variance conditions?** We added symmetric top-1 trimming: remove the highest-variance condition from each regime, recompute both arithmetic means over the remaining conditions, and calculate their ratio. | Setting | Untrimmed ratio | Symmetric top-1 trimmed ratio | |---|---:|---:| | BBH | 1.108 | 5.074 | | AIME | 1.854 | 1.958 | | MMLU-Pro, BBH-trained | 3.183 | 2.620 | | MMLU-Pro-trained | 1.174 | 0.726 | | R-Bench-T | 2.432 | 1.937 | | SMFR | 4.012 | 1.497 | All six untrimmed ratios exceed one. Five remain above one after symmetric trimming; MMLU-Pro-trained falls below one. The lower atomic variance is therefore a recurring pattern in the evaluated settings, but its robustness differs by setting. This trimming analysis measures dependence on extreme conditions; it is not a confidence interval for the variance ratio. The paired accuracy intervals also do not provide uncertainty intervals for variance ratios. We therefore report the stability finding through the observed run variances and the specified sensitivity check. We have incorporated the trimmed results into Table 1, the full calculation and cohort rules into Appendix C.2, and the resulting conclusion into Section 5.3.
Reviewer_boPoOpenReview ↗

Thanks for the detailed reply

Thanks for the detailed reply! I've updated my rating upwards.

AuthorsOpenReview ↗

Thank You for Your Feedback and Reassessment

Thank you for your feedback and for updating your rating! We greatly appreciate your recognition that the additional analyses strengthen the evidence for task-dependent coordination choices and the reported stability patterns, and address the main concerns.

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

This paper studies intermediate configurations between centralized and decentralized coordination in LLM-based MAS. The authors propose a Degrees of Freedom framework that uses a single variable d to characterize the coordination configuration: d=0 and d=1 correspond to purely centralized and purely decentralized coordination, respectively. For intermediate configurations, the Coordinator first decomposes the task at one of two granularities determined by d, then randomly assigns each task unit to the centralized or the decentralized branch with a probability also determined by d, and finally aggregates the unit outputs. Experimental results show that decomposition granularity and routing probability have a limited effect on the distribution of agent invocations, whereas the task domain and retained experience have a stronger effect. The optimal configuration varies across tasks: in most of the settings, the highest mean accuracy is achieved by an intermediate configuration. Based on these findings, the authors argue that coordination should be treated as a task- and state-dependent design choice rather than a fixed preference for either endpoint.

Critical strengths
  • This paper explores intermediate configurations that mix the two dominant coordination regimes, centralized and decentralized, and uses a continuous degree-of-freedom variable to jointly control the task decomposition granularity and the probability of assigning task units to each coordination branch. This expands the design space of MAS and may offer useful insights for future work.
Critical weaknesses
  • The comparison used to support the claim that intermediate configurations are superior is confounded, and under a fairer comparison, mixing the two coordination regimes shows no consistent advantage. The endpoint configurations (d=0 and d=1) pass the original task directly to a single branch, whereas all intermediate configurations additionally go through task decomposition and result aggregation by the Coordinator. Therefore, the abstract's claim that "in five of the six evaluated state–dataset settings, an intermediate DoF configuration achieves higher mean accuracy than both pure endpoints" does not show that the coordination space between the centralized and decentralized regimes is itself valuable. In fact, the configurations with d=0.0001, 0.5, 0.5001, and 0.9999 in Table 1 approximately correspond to "same decomposition + single branch" configurations, and the mixed configurations show no consistent advantage over single-branch configurations with the same decomposition. This suggests that the advantage of intermediate configurations comes mainly from task decomposition rather than from mixing the two coordination regimes. The paper acknowledges this in Section 5.3, but the abstract and introduction still present these gains as demonstrating the value of exploring the coordination space between the two regimes.

  • The paper's core conclusion is that the optimal configuration varies across tasks, but it does not further analyze which task characteristics determine these configuration preferences. As a result, when facing a new task, practitioners can still only choose a configuration by sweeping over DoF values, and the findings are insufficient to support the abstract's claim of providing "empirical guidance for studying and designing intermediate coordination regimes."

Questions for authors
  • Based on the motivation in the abstract and introduction, the paper focuses on the coordination space between the centralized and decentralized regimes. The paper treats decomposition granularity as one dimension of the decision space delegated to downstream agents and binds it with the branch-routing probability into a single variable d. What is the rationale for this design? How do the authors view the relationship between these two variables, and how can their respective effects on performance be disentangled?

  • For designers building MAS for unseen tasks, what insights can be drawn from this work? Which task characteristics can help predict the performance gains of different coordination configurations? Does the choice of task decomposition matter?

6 comments
AuthorsOpenReview ↗

Response to Reviewer oDkE

Thank you for distinguishing gains over undecomposed endpoints from benefits attributable to branch mixing. We have used that distinction to revise the claims and organize the supplementary analyses.

[W1] Do intermediate accuracy peaks establish an advantage from mixing coordination regimes?

They do not establish a consistent mixing advantage. The intermediate space includes decomposed single-branch configurations as well as mixed-routing configurations. A useful configuration can therefore lie between the original endpoints while assigning nearly all units to one branch.

Our interpretation has two parts.

First, explicit Coordinator decomposition contributes to a different workflow from either undecomposed endpoint. The large endpoint-to-intermediate changes highlighted in the review occur even when routing remains almost entirely within one branch. These comparisons support the importance of the decomposed workflow, but they jointly change decomposition, aggregation, and execution structure.

The following comparisons retain predominantly centralized assignment while adding the Coordinator-level workflow.

SettingDecomposed configurationUndecomposed endpointBootstrap 95% CINominal p-value
BBHd=0.0001d=0.0001: 0.750d=0d=0: 0.217[0.436, 0.628]1.37×10−271.37\times10^{-27}
AIMEd=0.0001d=0.0001: 0.600d=0d=0: 0.619[-0.089, 0.054]0.608
MMLU-Pro, BBH-trainedd=0.0001d=0.0001: 0.824d=0d=0: 0.662[0.103, 0.222]9.33×10−89.33\times10^{-8}
MMLU-Pro-trainedd=0.0001d=0.0001: 0.839d=0d=0: 0.705[0.080, 0.188]6.14×10−76.14\times10^{-7}
R-Bench-Td=0.0001d=0.0001: 0.633d=0d=0: 0.558[0.014, 0.137]0.0178
SMFRd=0.0001d=0.0001: 0.411d=0d=0: 0.243[0.090, 0.264]8.19×10−58.19\times10^{-5}

Second, routing remains consequential within a fixed granularity regime. The paired comparisons show that performance changes across routing probabilities even when the requested decomposition granularity is unchanged. Thus, the observed variation is not explained solely by whether Coordinator decomposition is enabled. This supports studying both delegation scope and routing locus, without implying that their optimum must use both branches.

SettingGranularitydAd_A vs. dBd_BAccuracy A / BBootstrap 95% CINominal p-value
BBHAtomic0.5 vs. 0.00010.883 / 0.750[0.074, 0.195]1.70×10−51.70\times10^{-5}
AIMESubtask0.6 vs. 0.99990.667 / 0.581[0.030, 0.148]0.00524
MMLU-Pro, BBH-trainedAtomic0.5 vs. 0.00010.856 / 0.824[-0.001, 0.064]0.058
MMLU-Pro-trainedSubtask0.5001 vs. 0.99990.848 / 0.586[0.205, 0.318]3.27×10−193.27\times10^{-19}
R-Bench-TAtomic0.2 vs. 0.00010.704 / 0.633[0.022, 0.123]0.00625
SMFRSubtask0.7 vs. 0.99990.396 / 0.295[0.038, 0.163]0.00213

We agree that a mixed configuration should be compared with the stronger decomposed single-branch alternative, not only the weaker one. For example, the best observed AIME subtask configuration is close in accuracy to its predominantly centralized alternative.

AIME subtask configurationddTarget q(d)q(d)Accuracy
Predominantly centralized0.50010.00020.659
Mixed routing0.60.20.667
Predominantly decentralized0.99990.99980.581

Accordingly, we do not interpret selected improvements over predominantly decentralized execution as a general benefit from mixing.

The revised Abstract states that intermediate configurations attain the highest observed mean accuracy in several settings and that this makes the unified space worth exploring. The Introduction, Related Work, and Section 5.3 follow the same claim. We have incorporated the endpoint and within-granularity paired comparisons into Appendix G.2.

[Q1] Why bind granularity and routing, and how can their effects be disentangled?

Our starting point was to connect centralized and decentralized coordination through a common configuration variable. We assign task units probabilistically between the branches, rather than selecting one branch for the entire original request. This creates two design choices: what constitutes a delegated unit, and where that unit is processed.

Granularity controls the semantic scope left for downstream decisions. In MetaAgent, a broader request leaves more internal planning and worker-selection decisions to the branch. In AgentNet, it leaves more scope for local execution, further splitting, or forwarding. Routing controls the fraction of units assigned to decentralized execution.

The single index provides a compact schedule over these choices. Crossing d=0.5d=0.5 changes the granularity regime and restarts the routing sweep. This does not make the variables inseparable: matched-routing comparisons examine granularity, while within-regime comparisons examine routing under the same requested granularity.

These are protocol-level controls, not cached-decomposition replay experiments. The following unit-count summaries and the recovered example support comparable decomposition scale and structure, but do not establish identical task-unit realizations.

SettingAtomic units, mean ± SDSubtask units, mean ± SD
BBH16.98 ± 0.044.49 ± 0.02
AIME19.87 ± 0.185.11 ± 0.02
MMLU-Pro, BBH-trained12.54 ± 0.083.70 ± 0.02
MMLU-Pro-trained12.46 ± 0.063.51 ± 0.46
R-Bench-T20.28 ± 0.095.30 ± 0.02
SMFR39.13 ± 0.217.67 ± 0.96

The recovered R-Bench-T example compares d=0.2d=0.2 and d=0.4d=0.4. Both decompositions follow five broad stages, with 15 and 14 units. Sections 3.1–3.2 explain the parameterization. We have incorporated the decomposition-scale summaries and example into Appendix G.1, alongside the paired analyses in Appendix G.2.

[W2, Q2] What can designers infer for unseen tasks, and does the choice of decomposition matter?

The choice of decomposition matters, and finer decomposition is not uniformly preferable. AIME’s results suggest evaluating coarser units for tightly coupled reasoning. They do not establish that context fragmentation is the measured cause of the atomic configurations’ lower accuracy.

AIME configuration groupObserved accuracy
Atomic configurations0.581–0.611
Undecomposed centralized endpoint0.619
Highest observed subtask configuration0.667

The study identifies three properties to consider when defining a validation search:

  1. Task structure: Whether useful units are relatively independent or depend on a tightly coupled reasoning sequence.
  2. Accuracy versus repeatability: Atomic delegation generally has lower observed across-run variance, while some tasks reach their highest observed accuracy under subtask delegation.
  3. Retained-state alignment: Branch profiles change with the retained state, so a routing preference should be evaluated using the state intended for deployment.

The stability results supporting the second consideration are summarized below.

SettingUntrimmed variance ratioSymmetric top-1 trimmed ratio
BBH1.1085.074
AIME1.8541.958
MMLU-Pro, BBH-trained3.1832.620
MMLU-Pro-trained1.1740.726
R-Bench-T2.4321.937
SMFR4.0121.497

For retained-state alignment, the same MMLU-Pro evaluation collection has a cross-branch profile JSD of approximately 0.583 under the BBH-trained state and 0.015 under the MMLU-Pro-trained state. These are candidate diagnostics, not universal selection rules. In particular, cross-branch profile similarity is an observed organizational measurement, not a proven predictor of which branch will answer a new task correctly.

We recommend selecting dd on a disjoint validation set and freezing it before held-out evaluation. We have added this procedure to Appendix G.2, but have not conducted a separate held-out validation of the procedure itself. The practical contribution is to make the relevant configuration dimensions explicit and provide evidence for evaluating them, rather than claim that the current study removes the need for validation. The supporting accuracy, stability, and retained-state results appear in Table 1 and Section 5.2.

Reviewer_oDkEOpenReview ↗

Reponse to the Authors

Thank you for the detailed response and the additional analyses. I appreciate that you acknowledge the intermediate accuracy peaks do not establish a consistent advantage from mixing coordination regimes, and that you have revised the claims accordingly. However, my main concerns remain.

Regarding W1, the new endpoint comparisons confirm that the gains of intermediate configurations over the endpoints come mainly from Coordinator decomposition. The within-granularity comparisons do not change this conclusion. Three of the six pairs (BBH, MMLU-Pro (BBH-trained), and MMLU-Pro-trained) compare two single-branch configurations. They show that the choice of branch matters, but not that mixing the two branches helps. Table 1 points the same way. Intermediate configurations achieve the highest mean accuracy in five settings. In three of them, the peak is a decomposed (near-)single-branch configuration. In the other two (AIME and MMLU-Pro-trained), the mixed configuration is roughly on par with the best single-branch configuration.

Regarding Q1, the response clarifies that binding granularity and routing into a single index is mainly a parameterization choice, which is reasonable. However, d is then less a single continuous dimension than a combination of two granularity regimes and a routing probability. This somewhat weakens the conceptual contribution of a unified DoF variable.

Regarding W2, the three considerations are helpful. However, as you acknowledge, they are candidate diagnostics rather than validated selection rules, and the task-structure consideration rests mainly on AIME. Practitioners facing a new task would still need to search for a configuration on a validation set, so my concern about practical guidance remains.

Overall, the revisions make the claims more accurate. However, the additional analyses mainly re-examine existing results and do not resolve the core concerns. I will therefore keep my score unchanged.

Reviewer_Cs8KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies how to combine centralized and decentralized coordination in LLM multi-agent systems, where the proposed DoF framework controls task decomposition and assigns subtasks to the two branches. In the experiments, the authors compared different settings on reasoning tasks and examined how coordination and previous experience affect performance.

Critical strengths
  • Looking at the settings between fully centralized and fully decentralized systems is an interesting idea.

  • The paper is easy to follow, with detailed descriptions of the framework and experimental setup.

Critical weaknesses
  • My main concern is what DoF tells us beyond combining task decomposition with two existing systems. The parameter (d) controls both decomposition granularity and routing probability, but the probability drops from 1 to almost 0 when (d) passes 0.5. It is not clear to me why increasing (d) should mean giving agents more freedom.

  • I find the comparison with the two endpoints difficult to interpret. The intermediate settings include extra decomposition and aggregation steps that the endpoints do not have. Section 5.3 acknowledges this, but we still cannot tell whether the gains come from mixing coordination methods or simply breaking the task into smaller parts.

  • The experiments use only one underlying model and one pair of coordination systems. I wonder whether the findings would hold with other implementations, particularly since the two branches use different temperatures and execution limits.

Questions for authors
  • Could you keep the decomposed tasks fixed and compare sending everything to the centralized branch, everything to the decentralized branch, and a mixture of both? A similar token budget would make this comparison more convincing.

  • Why combine granularity and routing probability into a single parameter? Would varying them separately help explain where the improvements come from?

  • How would we choose (d) for a new task? Could you select it on a validation set and check whether it works well on held-out tasks?

5 comments
AuthorsOpenReview ↗

Response to Reviewer Cs8K

Thank you for the questions about delegated decision scope, decomposition controls, and configuration selection.

[W1, Q2] What does DoF capture, and why does the routing probability reset?

DoF connects two aspects of delegated decision space: the semantic scope of each delegated unit and the coordination branch that processes it. Its contribution is a unified framework for studying these choices under a shared model, agent roster, and pair of branch implementations.

Although the Coordinator dispatches units to a branch, their semantic scope affects the decisions made inside that branch:

  • MetaAgent: The planner prefers a single step unless the request contains separable reasoning stages. An atomic unit specifies a narrow request, leaving fewer reasoning stages to organize internally. A broader subtask leaves more of this organization to MetaAgent through its plan–select–execute–review loop, including worker selection, replanning, and redo_current revisions.
  • AgentNet: Each active agent chooses among execute, split, and forward. An atomic unit already specifies a relatively fine execution scope. A broader subtask leaves the receiving agent more responsibility for deciding whether to execute directly, split the request, or forward it to another node. Atomic delegation does not disable these operations; it narrows the request over which they operate.

The transition from d=0.5d=0.5 to d=0.5001d=0.5001 changes delegated granularity from atomic to subtask. The routing sweep then restarts within the broader granularity regime. Thus, the transition concerns the scope of the delegated request, while increasing q(d)q(d) within either regime concerns where those requests are processed. The configuration index is real-valued, but its mapping to execution is piecewise.

The two controls are connected through the delegation interface: granularity determines which task-organization decisions the Coordinator resolves before delegation, while routing selects the mechanism that handles the remaining decisions. Binding them into one index provides a compact schedule over these allocations. Their effects can still be examined separately through matched-routing granularity comparisons and within-regime routing comparisons. We have added this explanation to Sections 3.1 and 3.2.

[W2] Are gains over the endpoints attributable to decomposition or mixing?

We distinguish two findings.

First, endpoint-to-intermediate gains include changes associated with Coordinator-level decomposition, which is itself part of DoF. The original endpoints bypass this outer decomposition and aggregation, whereas intermediate configurations retain the Coordinator’s organization of delegated work. The BBH comparison highlighted by the reviewer is therefore consistent with an important contribution from the decomposed workflow, even when execution remains almost entirely centralized.

A decomposed single-branch configuration is not the corresponding original endpoint. At d=0.5d=0.5, for example, the Coordinator determines the atomic units, dependencies, and aggregation, while AgentNet controls execution within each unit. At d=1d=1, AgentNet receives the original task directly. Both MetaAgent and AgentNet already support internal task organization; DoF changes how requests enter those mechanisms rather than replacing them.

The following centralized-branch comparisons illustrate the performance consequences, including the setting where decomposition does not improve the observed mean.

SettingDecomposed configurationOriginal endpointBootstrap 95% CINominal p-value
BBHd=0.0001d=0.0001: 0.750d=0d=0: 0.217[0.436, 0.628]1.37×10−271.37\times10^{-27}
AIMEd=0.0001d=0.0001: 0.600d=0d=0: 0.619[−0.089, 0.054]0.608
MMLU-Pro, BBH-trainedd=0.0001d=0.0001: 0.824d=0d=0: 0.662[0.103, 0.222]9.33×10−89.33\times10^{-8}
MMLU-Pro-trainedd=0.0001d=0.0001: 0.839d=0d=0: 0.705[0.080, 0.188]6.14×10−76.14\times10^{-7}
R-Bench-Td=0.0001d=0.0001: 0.633d=0d=0: 0.558[0.014, 0.137]0.0178
SMFRd=0.0001d=0.0001: 0.411d=0d=0: 0.243[0.090, 0.264]8.19×10−58.19\times10^{-5}

Decomposed single-branch or near-single-branch configurations also attain the highest observed mean accuracy in four settings: MMLU-Pro under atomic decentralized execution, MMLU-Pro-trained and R-Bench-T under predominantly centralized subtask execution, and SMFR under predominantly centralized atomic execution. The MMLU-Pro-trained peak is tied with a mixed atomic configuration.

These results identify an additional practical finding: Coordinator-level task organization can improve an existing MAS without retraining its model or updating its retained state. For the BBH-trained state evaluated on other domains, these improvements require no additional target-domain training.

Second, performance still varies with routing within a fixed granularity regime. The sweep contains predominantly centralized, predominantly decentralized, and mixed-routing configurations under the same requested granularity. The paired comparisons below show routing-dependent differences in several settings. These results indicate that specifying decomposition granularity alone does not determine performance; branch assignment also matters.

SettingGranularitydAd_A vs. dBd_BAccuracy A / BPaired bootstrap 95% CINominal p-value
BBHAtomic0.5 vs. 0.00010.883 / 0.750[0.074, 0.195]1.70×10−51.70\times10^{-5}
AIMESubtask0.6 vs. 0.99990.667 / 0.581[0.030, 0.148]0.00524
MMLU-Pro, BBH-trainedAtomic0.5 vs. 0.00010.856 / 0.824[−0.001, 0.064]0.058
MMLU-Pro-trainedSubtask0.5001 vs. 0.99990.848 / 0.586[0.205, 0.318]3.27×10−193.27\times10^{-19}
R-Bench-TAtomic0.2 vs. 0.00010.704 / 0.633[0.022, 0.123]0.00625
SMFRSubtask0.7 vs. 0.99990.396 / 0.295[0.038, 0.163]0.00213

The direction depends on the setting: BBH favors decentralized assignment in the displayed atomic comparison, whereas MMLU-Pro-trained favors predominantly centralized assignment in the displayed subtask comparison. The BBH-trained MMLU-Pro atomic contrast remains unresolved by its interval.

Our revised claim is therefore that the unified intermediate space contains useful task-dependent configurations. These may be decomposed single-branch configurations or mixed-routing configurations. The benefit includes configuring how an existing MAS receives and organizes delegated work, even when execution remains within one branch. We do not treat gains over the original endpoints as evidence that mixing is consistently superior.

The Abstract, Introduction, Related Work, and Section 5.3 now preserve the intermediate-accuracy-peak finding and add the Coordinator-level decomposition finding. We have incorporated the paired comparisons into Appendix G.2.

[Q1] Does the existing sweep implement the requested fixed-decomposition, budget-matched comparison?

It includes the relevant decomposed routing controls, but it is not the exact experiment proposed. We did not cache one decomposition for each task and replay that identical decomposition under all three routing policies.

The sweep fixes the decomposition procedure, prompt contract, and requested granularity within each regime. Decomposition precedes routing, and its specified inputs do not include q(d)q(d). The following table provides aggregate unit-count summaries across routing values.

SettingAtomic units, mean ± SDSubtask units, mean ± SD
BBH16.98 ± 0.044.49 ± 0.02
AIME19.87 ± 0.185.11 ± 0.02
MMLU-Pro, BBH-trained12.54 ± 0.083.70 ± 0.02
MMLU-Pro-trained12.46 ± 0.063.51 ± 0.46
R-Bench-T20.28 ± 0.095.30 ± 0.02
SMFR39.13 ± 0.217.67 ± 0.96

The recovered R-Bench-T example in revised paper compares d=0.2d=0.2, q=0.4q=0.4, and d=0.4d=0.4, q=0.8q=0.8. The two executions follow the same five broad stages, with 15 and 14 units and local differences. This supports comparable high-level structure, not identical decomposition across all executions.

We also did not match total token budgets. The common maximum of 2,048 output tokens applies to individual model calls, not entire task executions. Similar unit counts do not imply equal total tokens or model-call counts. Consequently, the comparisons characterize workflow performance under the specified implementations, not cost-matched gains.

We have incorporated the unit-count summaries and recovered example into Appendix G.1. Appendix B.3 documents the execution settings.

[W3] How do model choice, branch implementations, temperatures, and execution limits affect the conclusions?

The exact configuration preferences have not been established across other backbones or branch pairs. Our design fixes those components to examine configuration changes within one consistent execution environment.

The organizational focus is related to prior studies of debate protocols [1] and communication connectivity [2]. These studies motivate examining coordination choices directly; they are not evidence that our particular results transfer unchanged. AgentNet [3] supplies the decentralized coordination mechanism, while the diagnostic perspectives of [4, 5] motivate examining organization and execution rather than assuming that greater system complexity improves performance.

The execution differences are specific: Coordinator and MetaAgent calls use temperature 0.3; AgentNet calls use temperature 0 and top-p=1.0p=1.0. MetaAgent permits up to twelve planning rounds and one redo per step. AgentNet permits up to three forward hops and five executions per task. Both clients have a maximum output length of 2,048 tokens per call.

These settings remain fixed within each implementation throughout the sweep, but they are not matched between branches. Changing routing therefore changes assignment between implemented branches with different execution procedures. It does not isolate coordination topology from decoding and execution limits. Our interpretation is a controlled characterization of this system pair, with the exact settings reported in Appendix B.3.

[Q3] Can validation-selected dd be evaluated on held-out tasks?

Yes, that is the appropriate procedure for selecting a deployment configuration. However, we have not performed a separate validation-selection and held-out evaluation experiment in this study.

The revised paper now gives a concrete starting procedure based on the available MAS:

Available MASInitial configurationsPurpose
Centralized-orientedd=0.0001d=0.0001, d=0.5001d=0.5001Compare atomic and subtask delegation under predominantly centralized assignment
Decentralized-orientedd=0.5d=0.5, d=0.9999d=0.9999Compare atomic and subtask delegation under predominantly decentralized assignment
Both branchesAll four configurationsCompare delegation granularity and branch assignment jointly

The original workflow remains a baseline. These probes first test whether external task organization helps and which granularity is promising. Further validation can explore additional routing probabilities within a promising regime or perform the full DoF sweep. The model and retained state remain fixed during selection; the chosen dd is frozen before held-out evaluation.

Our results also inform candidate selection: AIME favors coarser subtask delegation, atomic delegation generally shows lower observed run-to-run variance, and retained-state changes alter branch profiles.

Candidate diagnosticObserved result
Decomposition scopeAIME atomic accuracy: 0.581–0.611; highest observed subtask accuracy: 0.667.
RepeatabilityAll six untrimmed subtask-to-atomic variance ratios exceed one; five remain above one after symmetric top-1 trimming.
Retained-state alignmentOn MMLU-Pro, cross-branch profile JSD is approximately 0.583 with the BBH-trained state and 0.015 with the MMLU-Pro-trained state.

These observations guide what to validate; they are not a validated predictor of the best dd for an unseen collection. The initial probes likewise do not guarantee recovery of the full-sweep optimum. Section 5.3 now summarizes this procedure, and Appendix G.2 provides the configuration probes and validation protocol.

References cited in this response

[1] Should We Be Going MAD? A Look at Multi-Agent Debate Strategies for LLMs. ICML 2024.
[2] Improving Multi-Agent Debate with Sparse Communication Topology. Findings of EMNLP 2024.
[3] AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems. NeurIPS 2025.
[4] Why Do Multi-Agent LLM Systems Fail? NeurIPS 2025, Datasets and Benchmarks Track.
[5] The Illusion of Multi-Agent Advantage. arXiv:2606.13003, 2026.

Comments on the paper

3 comments

Dear Reviewers,

Could you please review the authors’ rebuttals and follow up with any further comments or questions?

Many thanks for your time and efforts.

Best regards, AC

AuthorsOpenReview ↗

General Response

General Response

We thank the reviewers for recognizing the value of exploring intermediate coordination configurations (Cs8K, oDkE), the clarity of the framework and experimental setup (Cs8K, boPo), and the usefulness of the systematic task-domain sweeps (boPo). We address their concerns through four main revisions:

  1. DoF interpretation: clarify how delegated granularity affects branch-internal decision scope and explain why the routing sweep restarts at the granularity boundary.
  2. Performance attribution: separate decomposition-associated endpoint contrasts from within-granularity routing comparisons and refine the claims about intermediate configurations.
  3. Statistical reliability: add task-aligned paired uncertainty analyses, specify variance aggregation, and assess sensitivity to high-variance conditions.
  4. Decomposition and configuration selection: add decomposition-scale summaries and a recovered same-task prompt comparison, and describe validation-based selection for new task collections.

The responses below address each reviewer's specific concerns and provide key results, with full details in the revised paper (changes are highlighted in blue).

Dear Reviewers,

As the AC–reviewer discussion phase is about to begin, please highlight any disagreements or inconsistencies across the reviews, as well as any concerns that remain unresolved after the author rebuttal.

Please share your perspectives on these points and engage with one another to help us reach a well-supported assessment of the paper.

Thank you for your time and contributions.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.