acceptodds
Under review as a conference paper at ICLR 2027

CEPO: A General Framework for Improving Prompt Optimization in Multi-Agent Systems

Abstract

Multi-agent systems (MAS) require both effective coordination and specialized task execution. Existing prompt optimization approaches typically couple these functions within agent-specific prompts. We introduce Coordinator-Executor Prompt Optimization (CEPO), a general framework for improving prompt optimization in MAS. CEPO separates system-level coordination from agent-level execution through a shared coordinator prompt and agent-specific executor prompts, and guides their updates according to their respective roles. This structure allows shared coordination rules to be refined consistently across agents while keeping execution instructions separately adjustable. CEPO provides the prompt structure and optimization guidance, while existing prompt optimizers perform the underlying search. Extensive experiments spanning 11 benchmarks across four domains demonstrate that applying CEPO to four existing prompt optimizers consistently improves their performance. Ablations and qualitative analyses support the proposed design, while controlled studies on a document-understanding MAS further demonstrate improved optimization budget efficiency and performance scaling with the number of agents.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-agent Systems, Prompt Optimization, Self-evolving Agents

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bX3nOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work focuses on prompt optimization for multi-agent systems (MAS), particularly addressing the challenge of credit assignment in complex, interconnected agent systems. To improve optimization, the authors propose CEPO, a framework based on two principles: (1) coordinator-executor prompt decomposition, which separates each agent’s prompt into a shared coordination prompt and an agent-specific execution prompt; and (2) role-aligned optimization, which provides distinct optimization guidance for coordination and execution prompts. The proposed framework aims to improve prompt optimization by explicitly separating system-level coordination from agent-level execution. Experimental results across multiple domains demonstrate performance improvements when applying CEPO to existing prompt optimizers, compared with unoptimized systems.

Critical strengths
  1. Relevant and important research problem. Prompt optimization is important but challenging in multi-agent systems, particularly due to the difficulty of credit assignment in complex and tightly coupled agent interactions. Investigating how to improve credit assignment to guide more effective prompt optimization is a meaningful research direction, and this paper addresses this problem through structured prompt decomposition and role-aligned optimization.
Critical weaknesses
  1. Insufficient justification of the proposed design principles. The two proposed principles are not sufficiently justified. Regarding the first principle, the assumption that all agents should share a common coordination prompt is not well motivated. In many real-world multi-agent systems, a shared coordination prompt is not necessary. For example, orchestration-based multi-agent systems are widely adopted in industry, where individual agents maintain their own prompts and coordination objectives are communicated through context rather than predefined shared instructions. Although the proposed decomposition may be applicable to certain multi-agent architectures, this strong architectural assumption limits the generalizability of the framework. Regarding the second principle, while distinguishing coordination-level and execution-level optimization is reasonable, the proposed role-aligned optimization primarily relies on LLM-based optimization guided by role-specific meta-instructions. The paper does not sufficiently justify how this mechanism achieves better credit assignment or provides generalizable insights for designing future optimization methods.

  2. Insufficient experimental validation of the proposed framework’s superiority. Although the experimental results demonstrate performance improvements, the comparisons in Tables 2 and 3 do not fully establish the superiority of the proposed framework. In particular, Table 2 primarily compares CEPO-enhanced systems against their unoptimized counterparts, making the observed improvements less convincing given the additional optimization process. While Table 3 includes comparisons with MASPO, a more comprehensive evaluation against other existing prompt optimization approaches would strengthen the claims. Considering the paper’s focus on improving prompt optimization, comparisons with methods such as TextGrad (Yuksekgonul et al., 2024), MiCRO (Opsahl-Ong et al., 2024), and more recent approaches would provide stronger evidence of its effectiveness.

Questions for authors
  1. How does the proposed framework compare with existing prompt optimization methods for multi-agent systems, particularly under comparable optimization budgets?
  2. How sensitive is the proposed framework to the capabilities of the optimizer model? Specifically, what happens when a weaker model is used as the optimizer, particularly when its capabilities are comparable to or weaker than those of the executor models?
  3. How stable are the optimization results across different runs? Since verbalized prompt optimization can be less stable than numerical parameter optimization, evaluating its robustness across multiple runs is important. Could the authors report the mean performance and standard error over 3 or 5 independent optimization runs to demonstrate the stability and reliability of the proposed method?
2 comments
AuthorsOpenReview ↗

Response - (1)

Response Summary: Thank you for the detailed questions. We address orchestration and role-aligned guidance (W1, Tables R8 and R9) and provide optimization-budget controls (W2/Q1, Table R10). Added optimizer-model comparisons show gains over PE2 (Q2, Table R11, Appendix B.6). For MDocAgent, we report three-run means and standard errors (Q3, Table R12, Appendix B.7). The manuscript clarifies guidance in Section 3.2 and summarizes robustness in Section 4.3.

W1, architecture. Applicability to orchestration.

CEPO improves a system that coordinates through context, matching the orchestration mechanism described in the review (Table R8, paper Table 7). In LongVideoAgent, the master dynamically chooses which agent to invoke, constructs vision queries, and accumulates agent responses in its conversation context. Grounding and vision agents execute requests using their role prompts. Thus, coordination is communicated through requests and observations in context. CEPO improves this system by optimizing rules for requests, evidence reporting, and evidence sufficiency while retaining the same interaction and routing mechanism.

Table R8. Orchestration in LongVideoAgent.

MethodAccuracyAdditional CEPO gain over PE2
Baseline60.3--
PE263.4--
PE2+CEPO64.9+1.5

W1, guidance. The design lesson.

Separate the optimized functions and provide feedback appropriate to each function. In our MDocAgent implementation, coordinator updates use the full interaction trace. Executor updates focus on the target agent's input and output with task and outcome feedback. The swap ablation shows that matching guidance to its role matters (Table R9). Our contribution here is a functional decomposition with corresponding optimization guidance, rather than a separate credit assignment algorithm. We clarified this feedback scope and design lesson in Section 3.2.

Table R9. Guidance alignment on MMLB.

Coordinator guidanceExecutor guidanceAccuracy
Coordination guidanceExecution guidance61.0
Execution guidanceCoordination guidance57.7

W2/Q1. Existing optimizers and matched budgets.

CEPO's gain is not explained by additional optimization calls. We evaluate it with PE2, ProTeGi, OPRO, and MASPO. Table R10 shows that both comparable and lower budgets favor CEPO. This addresses the concern that extra optimization alone explains the improvement. These budget controls are already in paper Table 5 and Appendix B.1. We made PE2 explicit in the method labels and added the common optimizer comparison in Appendix B.5.

Table R10. Controls on optimization calls on MMLB. Iteration pairs denote (TC,TE)(T_C, T_E). MAS calls count internal agent LLM API requests.

MethodOptimizer callsMAS callsAccuracy
PE2 (0,3)105,71457.9
PE2+CEPO (1,1), lower budget23,12060.1
PE2 (0,6)2211,94557.4
PE2+CEPO (3,3), comparable budget2011,45861.0

Q2. Weaker optimizers and the same model as executors.

CEPO remains effective with a weaker optimizer and when the optimizer uses the same model as the executors. We added both conditions on MDocAgent+MMLB (Table R11). All settings are identical to the default GPT-5.4 experiment, except for the optimizer model. The additional gains over PE2 are larger in these conditions than in the original GPT-5.4 setting. The observed benefit therefore does not require a stronger optimizer. We present this as the observed result on MMLB, without claiming that absolute performance is invariant to the optimizer model. We added these comparisons to Appendix B.6 and paper Table S12, with the main finding summarized in Section 4.3.

Table R11. Sensitivity to the optimizer model on MDocAgent+MMLB.

OptimizerBaselinePE2PE2+CEPOGain over baselineCEPO gain over PE2
GPT-5.454.157.961.0+6.9+3.1
GPT-5-mini54.154.057.4+3.3+3.4
Qwen3.5-4B (same model as executors)54.155.860.0+5.9+4.2

Q3. Stability across runs.

CEPO has higher mean accuracy on every benchmark across three independent optimization runs on MDocAgent (Table R12). Table R12 reports mean accuracy and standard error. We added the reporting protocol to Section 4.1 and the statistics to Appendix B.7 and paper Table S13, with a summary in Section 4.3.

Table R12. Stability across three independent optimization runs on MDocAgent.

BenchmarkPE2 mean±SE (%)PE2+CEPO mean±SE (%)
MMLB57.9±0.4661.0±0.40
LongDocURL74.4±0.2375.4±0.29
PaperTab48.1±0.5853.7±0.69
PaperText71.4±0.6976.4±0.58
FetaTab77.8±0.2980.8±0.29

Dear Reviewer bX3n,

With one day remaining in the discussion period, we would appreciate your feedback on our response above, including the optimization-budget controls, weaker-optimizer experiments, and three-run stability results (Tables R10–R12), as well as our clarification of the shared coordinator and role-aligned guidance (W1). Could you let us know which concerns, if any, remain unresolved? We would be happy to provide further clarification.

Thank you for your time and consideration.

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

This paper studies how to improve prompt optimization for LLM-based multi-agent systems (MAS). The authors propose Coordinator-Executor Prompt Optimization (CEPO), which decomposes agent prompts into a shared coordinator prompt for system-level collaboration and agent-specific executor prompts for individual task execution. The two components are optimized separately using role-aligned guidance and existing prompt optimizers, rather than introducing a new underlying optimization algorithm. Experiments across 11 benchmarks in four domains demonstrate that CEPO can improve the performance of several existing optimizers, including PE2, OPRO, ProTeGi, and MASPO. Compared with prior work that optimizes agent prompts without explicitly separating coordination and execution, CEPO provides a structured approach to optimizing these two functions.

Critical strengths
  1. Separating system-level coordination from agent-level execution is an interesting idea. This decomposition provides a natural way to optimize different components of agent prompts according to their roles, which could be particularly useful for MAS with heterogeneous agents.

  2. Comprehensive empirical evaluation. The authors evaluate CEPO across multiple domains, including document understanding, mathematical reasoning, code generation, and video understanding. The experiments also consider four different prompt optimizers, including both general-purpose and MAS-specific methods. The ablation studies and optimization budget analysis further support the effectiveness of the proposed design.

Critical weaknesses
  1. Confusing presentation of experimental results and unclear attribution of improvements. The results in several tables are not presented clearly enough to distinguish the improvements from the underlying optimizer and those from CEPO. For example, Table 2 reports results labeled "+ CEPO (Ours)," but these results are actually obtained using PE2 + CEPO, rather than CEPO alone. For MDocAgent on MMLongBench-Doc, the reported improvement is +6.9 points over the original system, while Table 8 shows that PE2 alone already improves performance from 54.1 to 57.9. Thus, the incremental gain from CEPO is only +3.1 points. Moreover, the corresponding DocAgent + PE2 results are not provided, making it difficult to isolate CEPO's contribution. I encourage the authors to present all results consistently and explicitly distinguish the underlying optimizer's contribution from the additional improvement introduced by CEPO.

  2. Although CEPO is claimed to be a general framework applicable to different prompt optimizers, the optimizers are not evaluated on a consistent set of tasks. For example, PE2, ProTeGi, and OPRO are evaluated on document understanding benchmarks, while MASPO is evaluated on mathematical reasoning and code generation tasks. This makes it difficult to determine whether CEPO consistently improves different optimizers under the same conditions. It also raises a concern about potential selection bias, as it is unclear whether the reported optimizer-task combinations were selected in advance or based on their performance. The reviewer encourage the authors to include at least a subset of common benchmarks on which all four optimizers, where applicable, are evaluated with and without CEPO. This would provide more convincing evidence for the claimed generality.

Questions for authors
  1. What motivated the selection of different benchmark sets for different prompt optimizers? Were there any compatibility or computational constraints preventing a comparison on shared tasks? A controlled evaluation across the same benchmarks would help establish whether CEPO provides consistent improvements.
4 comments
AuthorsOpenReview ↗

Response - (1)

Response Summary: Thank you for identifying the attribution and presentation issues. The revision separates baseline, optimizer, and optimizer+CEPO results and adds the missing DocAgent+PE2 comparison (W1, Tables R5 and R6, Section 4.2 and paper Tables 2 and 8). A common evaluation on the same sequential math/code MAS shows improvements for all four optimizers on both benchmarks (W2/Q1, Table R7, Appendix B.5).

W1. Attribution and the missing DocAgent baseline.

CEPO accounts for most of the average improvement across the document benchmarks (Table R5). Per benchmark, CEPO improves over the base optimizer in all 15 optimizer and benchmark combinations (paper Table 8). The base optimizers alone provide only small average gains, while CEPO delivers substantially larger gains across all three optimizers. We also added the requested DocAgent+PE2 results on MMLongBench-Doc (MMLB) and the other document benchmarks (Table R6). Thank you for identifying the missing comparison. We now present baseline, optimizer, and optimizer+CEPO consistently and make CEPO's contribution explicit. Paper Table 2 now includes the PE2 comparisons, and paper Tables 2, 3, 7, and 8 consistently report CEPO gains over the base optimizer. Section 4.2 clarifies these contributions.

Table R5. Average improvement across five document benchmarks on MDocAgent.

OptimizerBaselineOptimizer aloneOptimizer+CEPOOptimizer gainCEPO gain
PE265.265.969.5+0.7+3.5
ProTeGi65.265.368.5+0.1+3.1
OPRO65.265.367.1+0.1+1.8

Table R6. DocAgent with the added PE2 comparison.

MethodMMLBLongDocURLPaperTabPaperTextFetaTabAverage
Baseline59.572.960.075.682.070.0
+PE264.371.760.076.081.970.8
+PE2+CEPO67.473.262.278.382.672.7

W2/Q1. Common evaluation across optimizers.

CEPO improves all four optimizers on both common benchmarks (Table R7). PE2, ProTeGi, and OPRO already share the same document evaluation, where CEPO improves every evaluated combination (paper Table 8, with averages in Table R5). MASPO follows the evaluation setting of its original study. This assignment was fixed in advance, and no optimizer and benchmark combination was chosen after seeing results. Following your suggestion, we added PE2, OPRO, and ProTeGi to the same sequential math/code MAS used for MASPO, evaluated on MATH-500 and HumanEval-ET. All four optimizers use the same experimental setting as the MASPO experiments in Section 4.1. CEPO improves all eight optimizer and benchmark combinations, with average gains from 1.21 to 5.34 points. Table R7 includes the MASPO results from paper Table 3. We added the common comparison to Appendix B.5 and paper Table S11, and summarize its findings in Section 4.3.

Table R7. Four optimizers on the same sequential math/code MAS and benchmarks.

MethodMATH-500HumanEval-ETAverage (CEPO gain)
Baseline75.1068.9072.00
PE275.8070.7373.27
PE2 + CEPO77.4073.7875.59 (+2.33)
OPRO74.4070.1272.26
OPRO + CEPO76.8075.6176.21 (+3.95)
ProTeGi69.8068.9069.35
ProTeGi + CEPO75.6073.7874.69 (+5.34)
MASPO77.8073.7875.79
MASPO + CEPO79.6074.3977.00 (+1.21)

Dear Reviewer heqX,

With one day remaining in the discussion period, we would appreciate your feedback on our response above. We added the missing optimizer-only baselines and evaluated all four optimizers on the same two benchmarks (Tables R5–R7), directly addressing the attribution and controlled-comparison concerns you raised. Do these additions address your main concerns, or are there any remaining points we could clarify?

Again, thank you for your time and consideration.

The authors solved most of the concerns, and I would like to raise the score.

Dear reviewer heqX,

Thank you for your follow-up and for revising your rating to 3 (Weak Acceptance). We are glad to hear that our response addressed your concerns, and we appreciate your constructive feedback, which helped us improve the paper.

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

CEPO is a general framework designed to improve existing prompt optimization methods for multi-agent systems. It decomposes agent prompts into two components: a shared coordinator prompt that governs system-level collaboration and role-specific executor prompts that guide individual task execution. CEPO employs role-aligned optimization, using distinct meta-instructions to optimize coordination and execution separately. By default, the coordinator prompt is optimized first, followed by individual executor prompts. The framework is evaluated with four existing prompt optimizers across 11 benchmarks spanning four domains, demonstrating consistent performance improvements over the corresponding baseline optimizers.

Critical strengths
  1. Extensive Evaluation: The paper evaluates the proposed framework across 11 benchmarks across four domains. The consistent performance improvements highlight its effectiveness across diverse tasks.

  2. Comprehensive Ablation Studies: The authors have provided extensive ablation studies that help in understanding the contributions of individual components and design choices.

  3. Generalizability Across Prompt Optimizers: The proposed framework consistently improves the performance of four existing prompt optimizers, demonstrating its broad applicability and compatibility with different optimization approaches.

Critical weaknesses

1. Insufficient motivation for the problem: The authors argue that existing prompt optimization methods couple coordination and execution but fail to demonstrate the failure modes arising from this coupling. While it is well known that multi-agent systems suffer from coordination issues, the claim that separating coordination and execution prompts can help address these issues is not sufficiently justified.

2. Limited justification for a shared coordinator prompt: CEPO assumes that a single shared prompt can capture coordination across all agents. However, coordination requirements may differ across agent roles and interactions. The paper does not compare this design with role-specific or hierarchical coordination prompts, leaving the benefits of a globally shared coordinator insufficiently justified.

3. Limited evidence of improved coordination: Since improving coordination is a central objective, the paper should analyze how optimization affects intermediate agent outputs and interactions. The lack of such evidence makes it difficult to assess whether the observed gains actually stem from improved coordination.

4. Cross-benchmark generalizability: Although the authors evaluate cross-benchmark transfer, the observed gains are relatively low compared to benchmark-specific optimization.

Questions for authors
  1. Can the authors provide concrete failure cases arising from coupling coordination and execution, and show how CEPO addresses them?

  2. Why is a shared coordinator preferred over role-specific or hierarchical coordination prompts?

  3. Can the authors demonstrate how CEPO improves intermediate agent interactions and reduces coordination failures?

  4. As shown in Table S7 (Appendix B.2), why does optimizing all five executor prompts reduce accuracy from 61.0% to 56.9%, compared to optimizing only the General agent's executor?

4 comments
AuthorsOpenReview ↗

Response - (1)

Response Summary: Thank you for the constructive suggestions. We address prompt separation and the shared coordinator with a failure analysis and coordinator-design comparison (W1/Q1, W2/Q2, Table R1). Intermediate-answer analysis supports improved coordination (W3/Q3, Table R2, Section 4.3 and paper Table 9). We also report coordinator transfer and executor coverage results (W4, Q4, Tables R3 and R4, Appendix B.2 and B.4).

W1/Q1. Why separate coordination from execution?

Optimizing one executor can change its response without updating the other agents' evidence reporting rules. Consider unanswerable questions, where the document lacks enough information to determine an answer. These decisions require considering all agents' evidence, since one agent may miss information that another finds or propose an unsupported answer.

We directly optimized the SUM agent with PE2. The SUM agent combines the other agents' answers to produce the system's final response. The prompt learned by PE2 instructs the SUM agent to return "Not answerable" whenever there is ambiguity or conflict. The optimized SUM agent returns "Not answerable" on 22.0% of questions that are actually answerable, compared with 1.6% for the original SUM agent prompt with CEPO's learned coordinator. Optimizing only the SUM agent with PE2 changes the final response locally without updating the other agents' evidence reporting rules. CEPO instead optimizes shared rules for reporting and combining evidence, so the coordination update reaches all agents.

W2/Q2. Why a single shared coordinator?

The shared design achieves the highest accuracy with the fewest optimizer and MAS calls (Table R1). We compared three designs on MDocAgent with MMLongBench-Doc (MMLB). The "role-specific" design independently optimizes one coordinator for each agent and applies each coordinator only to its corresponding agent. The "hierarchical" design first optimizes a global coordinator, fixes it, and then separately optimizes coordinators for three groups: initial reasoning, evidence processing, and final synthesis. Each agent receives the global coordinator and its group coordinator.

Table R1. Coordinator design comparison on MDocAgent+MMLB. Optimizer calls count optimizer API requests. MAS calls count internal agent LLM API requests.

Coordinator designAccuracy (%)Optimizer callsMAS calls
PE2+CEPO (shared)61.02011,458
Role-specific57.46039,900
Hierarchical58.75033,279

With role-specific or hierarchical coordinators, common rules are optimized in multiple places, and updates to these prompts are not synchronized. In the role-specific design, the GENERAL agent, which provides the initial answer, learns to "give the strongest supported partial answer" when evidence is insufficient. The SUM agent instead learns: "If the trace does not support a unique answer, say so rather than guessing." These independently learned rules can create conflicting expectations about how partial answers should be passed on and accepted. CEPO keeps common rules in one optimization target, so related revisions reach both the reporting and synthesis agents together. The shared prompt can still contain instructions for individual roles. We clarified the shared-prompt rationale in Section 3.2.

W3/Q3. Evidence of improved coordination.

CEPO better preserves correct intermediate answers through final synthesis (Table R2). When at least one intermediate agent is correct, the final error rate drops from 18.7% in the baseline to 12.1% with PE2+CEPO, compared with 16.3% for PE2 alone. Some individual agents also improve, but correct intermediate answers are not more frequently available than under the baseline. This supports coordination as an important source of the gain. As supporting evidence, accuracy on unanswerable questions improves from 53.4% with PE2 to 71.3% with PE2+CEPO. Combining agents' evidence and uncertainty judgments is especially important for these questions. Table R2 uses regenerated traces. We added this analysis to Section 4.3 and paper Table 9, with details in Appendix B.3.

Table R2. Individual agent performance and preservation of intermediate answers on MMLB.

Metric (%)BaselinePE2PE2+CEPO
Per agent accuracy
↳ GENERAL agent55.655.557.0
↳ CRITIC agent49.649.654.8
↳ TEXT agent46.844.947.6
↳ IMAGE agent57.456.056.8
At least one intermediate agent correct68.466.667.3
Final wrong given at least one correct intermediate answer ↓18.716.312.1
AuthorsOpenReview ↗

Response - (2)

W4. Transfer across benchmarks.

Optimizing CEPO on the target dataset gives the best result in this comparison, so a transfer gap is expected. CEPO transfer outperforms the baseline on average and PE2 optimized separately on every target dataset (Table R3). These results support reusing coordination rules without target-specific optimization. We added the target-specific PE2 comparison and averages to paper Table S9 and revised Appendix B.2 accordingly.

Table R3. Coordinator transfer compared with target-specific optimization.

SettingLongDocURLPaperTabPaperTextFetaTabAverage
Baseline74.946.372.178.868.0
PE2 optimized on target data74.448.171.477.867.9
PE2+CEPO, transferred coordinator74.849.773.479.469.3
PE2+CEPO optimized on target data75.453.776.480.871.6

Q4. Optimizing all executors.

We conjecture that the degradation is tied to the combination of PE2 and MDocAgent, as it also occurs with vanilla PE2 (paper Table S7). CEPO partially mitigates the drop when all executors are optimized. In the sequential MAS with MASPO on AQuA, optimizing all executors outperforms optimizing only the Predictor, and CEPO further improves both settings (Table R4). We clarified the sequential update procedure in Section 3.3, Eq. 4, and Algorithm 1, revised Appendix B.2, and added the comparison to Appendix B.4 and paper Table S10.

Table R4. Optimizing one agent versus all agents in the sequential MAS on AQuA.

Optimization settingBaselineMASPOMASPO+CEPO
One agent (Predictor)83.4783.0785.04
All agents83.4785.5687.01

Thank you, authors, for the comprehensive response. I am satisfied with the response and will increase the score.

Dear reviewer eTyA,

Thank you for your follow-up and for revising your rating to 3 (Weak Acceptance). We are glad to hear that our response addressed your concerns, and we appreciate your constructive feedback, which helped us improve the paper.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Global response

We thank all reviewers for their thoughtful and constructive feedback. We appreciate the recognition of our extensive evaluation and ablation studies (eTyA, heqX), and the importance of improving prompt optimization for multi-agent systems (bX3n).

Across our detailed responses and revised manuscript, we address all reviewers’ concerns and questions. The main additions and clarifications are:

  • Expanded controlled comparisons. We added a common evaluation of all four optimizers on shared benchmarks and provided the missing optimizer-only baselines, making CEPO’s incremental contribution explicit.
  • Additional evidence on coordination and design choices. We provide concrete failure analyses, comparisons with alternative coordinator designs, and intermediate-answer analyses that support CEPO’s motivation and its ability to preserve correct intermediate answers through final synthesis.
  • New robustness experiments. We evaluate weaker optimizer models, report results across three independent optimization runs, and examine the effect of optimizing additional executors.
  • Clearer presentation. We clarify CEPO’s contribution as a framework that complements existing optimizers, explain the optimization procedure and experimental settings, and distinguish the underlying optimizers’ gains from the additional gains introduced by CEPO.

The reviews have helped us substantially improve the paper. We welcome further questions and would be happy to provide additional clarification.

Dear Reviewer bX3n,

Thank you for your follow-up, for acknowledging our additional experiments, and for finding the proposed designs reasonable. We briefly clarify the two remaining points.

Credit assignment. We do not claim CEPO as a credit assignment method. CEPO is a design principle that matches feedback scope to the coordinator and executor functions (Section 3.2), and it complements existing credit assignment approaches rather than replacing them. For example, MASPO addresses cross-agent credit assignment by evaluating each agent's prompt by its effect on downstream agents, and CEPO further improves MASPO on all five benchmarks (Table 3). The gains also hold across four optimizers and weaker optimizer models (Tables S11 and S12), and swapping the guidance drops accuracy from 61.0 to 57.7 (Table 6). Together, these results suggest a reusable principle rather than setting-specific prompt engineering.

Generalizability. Section 3.1 follows the standard MAS prompt optimization setting, which existing optimizers such as MASPO share. CEPO only adds a shared prompt to this setting (Eq. 2) without constraining MAS topology or routing. In our experiments, CEPO improves four different MAS structures:

  • parallel agents with final synthesis (MDocAgent)
  • an actor-reviewer-reflector loop (DocAgent)
  • a sequential predictor-reflector pipeline (math and code)
  • dynamic orchestration through context (LongVideoAgent) CEPO is therefore not restricted to a particular MAS structure, either conceptually or empirically.

Thank you again for your feedback, which has substantially improved the paper.

AuthorsOpenReview ↗

Summary of the discussion period

Dear Area Chair and Reviewers,

We thank all reviewers for their thoughtful reviews and engagement during the discussion.

Strengths recognized by reviewers

  • An important problem: prompt optimization and credit assignment in multi-agent systems (bX3n)
  • A natural and interesting decomposition of system-level coordination and agent-level execution (heqX)
  • Extensive evaluation on 11 benchmarks across four domains with four prompt optimizers (eTyA, heqX)
  • Comprehensive ablations and budget analysis (eTyA, heqX), with consistent gains across optimizers (eTyA)

During the rebuttal, we addressed each reviewer's concerns and updated the manuscript accordingly. Below we summarize the outcome for each reviewer.

Reviewer eTyA (2 → 3). Concerns about motivation, the shared coordinator design, and evidence of improved coordination were addressed with concrete failure cases, a comparison against role-specific and hierarchical coordinators, and an analysis showing that CEPO better preserves correct intermediate answers through final synthesis.

Reviewer heqX (2 → 3). Concerns about attributing gains and about controlled comparison were addressed with consistent baseline / optimizer / optimizer+CEPO reporting, the missing DocAgent+PE2 baseline, and a common benchmark on which CEPO improves all four optimizers.

Reviewer bX3n (score maintained).

The reviewer acknowledged our additional experiments and found the proposed designs reasonable.

  • Experimental validation (W2/Q1, Q2, Q3): CEPO outperforms the base optimizer even with a lower budget, is robust to the choice of optimizer LLM, and is stable across three independent runs.

  • Contribution (follow-up on W2): CEPO is a general framework that separates the roles of prompts into shared coordination and agent-specific execution, and optimizes each with guidance matched to its role. The design lesson is to separate the optimized functions and match the feedback scope to each function (Section 3.2). CEPO is not a dedicated credit assignment method, but it further improves credit assignment–based optimizers such as MASPO on all five benchmarks.

  • Generalizability (follow-up on W1): CEPO is not restricted to a particular MAS structure, conceptually or empirically. It only adds a shared prompt to the standard MAS prompt optimization setting, and it improves four different MAS structures, including dynamic orchestration.

We thank the AC and reviewers again for their time and constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.