acceptodds
Under review as a conference paper at ICLR 2027

TCP-MCP: Landscape-Guided Co-Evolution of Prompts and Communication Topologies for Multi-Agent Systems

Abstract

Effective multi-agent systems cannot be designed by selecting prompts or communication graphs in isolation. Agent behavior depends on the information an agent receives, while the usefulness of a communication edge depends on how the receiving agent interprets and uses that information. We propose TCP-MCP (Topology-Coupled Prompting for Multi-Agent Collaborative Problem-Solving), a co-evolution framework that searches agent prompts and communication topologies as a unified genome. TCP-MCP couples this executable representation with a structure-aware evolutionary search procedure, using an initialization-time landscape probe to calibrate early search behavior and Pareto-front diagnostics to adapt exploration under three objectives: task performance, token cost, and structural complexity. Using the same DeepSeek-V3.2 backbone across all methods, TCP-MCP achieves 82.66%, 89.96%, 96.61%, and 77.31% accuracy on MMLU-Pro, MMLU, GSM8K, and WorkBench, respectively. Across the four benchmarks, it consistently outperforms automated graph-generation baselines and achieves competitive accuracy relative to debate-style systems, while using up to 5.71× fewer tokens than those systems at the reported operating points. These results show that jointly evolving prompts and communication structure provides a practical route to cost-aware and task-adaptive multi-agent system design in controlled evaluations.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, multi-agent systems, evolutionary optimization, automated agent design, communication topology, prompt search, Pareto optimization

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_FqinOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TCP-MCP, a framework for jointly optimizing agent prompts and communication topologies in multi-agent systems (MAS). It employs evolutionary search with structure-aware crossover, mutation, and landscape-guided adaptive control, while considering task accuracy, token cost, and structural complexity through Pareto-based optimization. Experiments on four benchmarks demonstrate competitive performance against existing MAS baselines.

Critical strengths
  1. Relevant problem and interesting approach. Jointly optimizing prompts and communication topologies is a meaningful problem in MAS design. Applying evolutionary algorithms to search over agent roles, prompts, and graph structures is an interesting and reasonable direction.

  2. Multi-objective optimization. Considering accuracy, token cost, and structural complexity together is a valuable design choice. The Pareto-based formulation provides a practical way to explore performance-efficiency trade-offs.

Critical weaknesses
  1. Limited methodological novelty and questionable necessity of the complex design. The proposed framework appears to largely combine established techniques, including LHS initialization, evolutionary crossover and mutation, NSGA-II, and heuristic adaptive control. The extensive terminology makes the method appear more sophisticated than its underlying mechanisms suggest. The paper does not sufficiently distinguish its core algorithmic novelty or justify the necessity of these components.
    In particular, the claimed landscape guidance mainly relies on a one-time FDC estimation and subsequent pareto-front statistics. Meanwhile, prompt evolution appears largely tied to role mutation and prompt regeneration, raising questions about the extent of actual prompt-topology co-evolution. Furthermore, restricting communication structures to DAGs excludes common iterative interactions, such as feedback loops between workers and verifiers, limiting the expressiveness of the searched MAS architectures.

  2. Experimental comparisons and evaluation limitations. First, the paper overlooks a highly relevant prior work, MASS (Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies, ICLR 2026), which also studies joint prompt-topology optimization. Its absence from both related work and baseline comparisons makes it difficult to assess the effectiveness of TCP-MCP relative to existing approach. Second, Table 1 shows that the relatively simple LLM-Debate baseline already achieves higher accuracy on MMLU and MMLU-Pro, while remaining competitive on GSM8K and WorkBench. Although Debate incurs considerably higher inference costs, these results raise questions about whether the elaborate evolutionary framework offers sufficient practical benefits. Moreover, Table 1 excludes the substantial offline search cost. Appendix F reports nearly 1 billion on WorkBench alone, the overall computational efficiency therefore deserves more careful analysis. Finally, MMLU and MMLU-Pro share a substantial number of underlying questions, which limits the diversity of the evaluation. More independent benchmarks would better support the generality of the method.

Questions for authors
  1. What concrete advantages does the proposed evolutionary approach offer over existing multi-stage prompt-topology optimization?
  2. How much do the landscape-guided mechanisms improve final performance over a standard evolutionary search baseline with the same search space and budget?
  3. Can the authors provide a more comprehensive efficiency comparison that accounts for offline search costs?
  4. Can the authors evaluate on other tasks, like coding or general reasoning, to better support the generality of the method?
5 comments

We thank the reviewer for finding joint prompt–topology optimization a meaningful problem and the accuracy–cost–complexity formulation valuable. We follow the reviewer's numbering below.

W1: Novelty

Limited methodological novelty and questionable necessity of the complex design. The proposed framework appears to largely combine established techniques, including LHS initialization, evolutionary crossover and mutation, NSGA-II, and heuristic adaptive control. The extensive terminology makes the method appear more sophisticated than its underlying mechanisms suggest. The paper does not sufficiently distinguish its core algorithmic novelty or justify the necessity of these components.

We agree that LHS, crossover and mutation, NSGA-II, and heuristic adaptive control are established techniques. The contribution lies in the object being optimized and in how its parts are updated together.

  1. One executable candidate. M = (G, P) encodes the communication graph and the node-level role–prompt configurations as one system that is executed and evaluated as a whole.

  2. Coupled updates. Structural operators change what each node receives, and role mutation regenerates the prompt of the mutated node. Both kinds of change are selected in the same population.

  3. System-level objectives. Accuracy, token cost, and graph complexity |V| + |E| are evaluated on the complete system.

  4. The ablations show that the components are needed (Table 4). Removing crossover lowers MMLU-Pro accuracy by 2.51 points, removing mutation lowers WorkBench success by 8.27 points, fixing the operator rates lowers it by 4.23 points, and staged optimization (AB4) is 2.48 and 5.96 points lower on MMLU and WorkBench.

We have simplified the terminology in Sections 3 and 4 and now describe each operation directly.

W1 and Q2: Landscape guidance

In particular, the claimed landscape guidance mainly relies on a one-time FDC estimation and subsequent pareto-front statistics.

How much do the landscape-guided mechanisms improve final performance over a standard evolutionary search baseline with the same search space and budget?

We would like to clarify that landscape guidance also acts at runtime, not only through the one-time FDC estimate. Removing all of it (AB7) lowers performance across all four tasks, by 1.76, 6.45, 3.95, and 5.96 points on MMLU, MMLU-Pro, GSM8K, and WorkBench.

  1. Runtime control uses landscape features. In addition to Pareto-front diagnostics, the controller adjusts topology and role mutation rates from diversity, neutrality, multimodality, and ruggedness features computed from recent populations (Appendix C.6).

  2. AB7 removes both parts within the same framework. AB7 uses neutral initialization and fixed operator rates, with both the large-mutation trigger and targeted injection disabled. It keeps the rest of the TCP-MCP search, including the cost–complexity bins, the structural quota, and the elite archive. It therefore measures landscape guidance in the same search space and budget, but it is not a plain NSGA-II baseline.

  3. Results (Table R4, see our response to Reviewer wSTn, W2 and Q2). Relative to Full, AB7 is 1.76, 6.45, 3.95, and 5.96 points lower on MMLU, MMLU-Pro, GSM8K, and WorkBench. FDC alone is evaluated over three complete runs in Table R6.

We have added AB7 to Section 5.4 and Appendix K. We have revised the abstract to describe landscape guidance as FDC-guided initialization together with runtime adaptive control and to report the effect of FDC only where it is significant.

W1: Extent of prompt evolution

Meanwhile, prompt evolution appears largely tied to role mutation and prompt regeneration, raising questions about the extent of actual prompt-topology co-evolution.

We agree with this description. Prompt bodies change only through role mutation, which regenerates the prompt for the new role, and through inheritance when crossover moves a node into another graph.

  1. No free-form prompt editing. The method does not rewrite the prompt text of every node in each generation. Crossover transfers subgraphs, nodes kept from the recipient keep their templates, and imported nodes bring the templates of the donor.

  2. The effective prompt still depends on the topology. BuildPrompt adds the messages of each node's upstream neighbors to its stored template (Section 3.1), so a structural change alters what downstream nodes receive even when their templates are unchanged.

We have stated this scope in Section 4.2.

W1: Restriction to DAGs

Furthermore, restricting communication structures to DAGs excludes common iterative interactions, such as feedback loops between workers and verifiers, limiting the expressiveness of the searched MAS architectures.

We agree that the search space excludes loops with dynamic termination.

  1. Fixed-round exchanges are expressible. A worker–verifier exchange with a fixed number of rounds can be unrolled into round-specific nodes, which forms a DAG.

  2. Reason for the restriction. A DAG gives every candidate a defined execution order and a bounded cost, which keeps the structure search and the cost comparison consistent. Dynamically terminated loops require iteration state, stopping rules, and cumulative budget control as additional search variables.

We have added this limitation to Section 6.

W2 and Q1: Relationship to MASS

First, the paper overlooks a highly relevant prior work, MASS (Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies, ICLR 2026), which also studies joint prompt-topology optimization. Its absence from both related work and baseline comparisons makes it difficult to assess the effectiveness of TCP-MCP relative to existing approach.

What concrete advantages does the proposed evolutionary approach offer over existing multi-stage prompt-topology optimization?

To answer directly: on MMLU the two methods reach the same accuracy and MASS uses 2.7 times fewer deployment tokens, on MMLU-Pro TCP-MCP is 3.44 points more accurate at 2.6 times the deployment cost of MASS, and on GSM8K and WorkBench TCP-MCP is 2.17 and 1.73 points more accurate while using 28.0% and 12.8% fewer deployment tokens.

  1. Setup. We reimplemented the three stages of MASS (block-level prompt optimization, workflow topology optimization, and workflow-level prompt refinement) and evaluated it on the same held-out sets as TCP-MCP. Following Zhou et al. (2025), our implementation uses MIPRO for block-level and workflow-level prompt optimization after single-agent CoT warm-up, and rejection sampling over the Aggregate, Reflect, Debate, Summarize, and Executor building blocks under an agent-count budget cap for topology search. Aside from replacing the underlying backbone with DeepSeek-V3.2 and using the same evaluation splits as TCP-MCP, all hyperparameters match the official MASS implementation. The total search tokens are in a similar range across the four tasks (378M, 490M, 174.64M, and 946.83M for MASS, and 351.05M, 486.05M, 128.02M, and 998.19M for TCP-MCP), but our MASS accounting does not give per-stage candidate counts, so this is not a strictly budget-matched comparison.

  2. Results (Table R8). On MMLU, MASS and TCP-MCP reach 89.94% and 89.96% at 2,323 and 6,305 tokens/query. On MMLU-Pro they reach 79.22% and 82.66% at 3,081 and 8,085 tokens/query. On GSM8K they reach 94.44% and 96.61% at 6,744 and 4,855 tokens/query, and on WorkBench they reach 75.58% and 77.31% at 17,543 and 15,289 tokens/query.

  3. Difference in approach. MASS optimizes prompts, then topology, then prompts again in sequence. TCP-MCP updates topology and role–prompt configurations together within one population and selects on accuracy, token cost, and graph complexity. AB4 and AB5 (Table R4) test staged and alternating schedules inside our framework but do not replace the comparison with MASS.

We have added MASS to Section 2 and Table R8 to Appendix P, and revised the contribution statement in Section 1. We will release our MASS reproduction code.

Table R8: MASS and TCP-MCP

BenchmarkMethodOptimizationAccuracy (%)Deployment tokens/querySearch tokens
MMLUTCP-MCPJoint evolution of (G, P)89.966,305351.05M
MMLUMASSBlock prompts, topology, global prompts89.942,322.73378M
MMLU-ProTCP-MCPJoint evolution of (G, P)82.668,085486.05M
MMLU-ProMASSBlock prompts, topology, global prompts79.223,081.19490M
GSM8KTCP-MCPJoint evolution of (G, P)96.614,855128.02M
GSM8KMASSBlock prompts, topology, global prompts94.446,744174.64M
WorkBenchTCP-MCPJoint evolution of (G, P)77.3115,289998.19M
WorkBenchMASSBlock prompts, topology, global prompts75.5817,543946.83M

TCP-MCP values are from Table 1 and Appendix F. MASS deployment tokens are 32,427,658 on MMLU and 37,072,826 on MMLU-Pro over 13,961 and 12,032 held-out queries.

W2: LLM-Debate

Second, Table 1 shows that the relatively simple LLM-Debate baseline already achieves higher accuracy on MMLU and MMLU-Pro, while remaining competitive on GSM8K and WorkBench. Although Debate incurs considerably higher inference costs, these results raise questions about whether the elaborate evolutionary framework offers sufficient practical benefits.

We agree that LLM-Debate is more accurate on MMLU and MMLU-Pro, by 0.88 and 1.01 points, and it uses about 4.6 and 5.7 times the deployment tokens of TCP-MCP on these tasks.

  1. TCP-MCP is more accurate on GSM8K and WorkBench, where LLM-Debate uses about 4.1 and 5.7 times the deployment tokens.

  2. The trade-off holds across Debate configurations (Appendix O, Table 18). Smaller Debate configurations cost less than TCP-MCP and are less accurate, and a six-agent, three-round configuration is more accurate at 5.6 times the cost.

We have stated this trade-off in Section 5.2.

W2 and Q3: Offline search cost

Moreover, Table 1 excludes the substantial offline search cost. Appendix F reports nearly 1 billion on WorkBench alone, the overall computational efficiency therefore deserves more careful analysis.

Can the authors provide a more comprehensive efficiency comparison that accounts for offline search costs?

To answer directly: including the one-time search, TCP-MCP becomes cheaper than LLM-Debate after 8,539–15,507 queries, and on WorkBench it breaks even against G-Designer and DyLAN after 23,519 and 28,598 tasks. On the three reasoning tasks, the Table 1 operating point uses more tokens per query than Self-Consistency, DyLAN, and G-Designer, whereas selecting lighter candidates from the same Pareto archive (Table R2) recovers the search cost whenever the selected candidate uses fewer tokens per query than the target baseline.

  1. Search cost per task. The searches use 351M, 486M, 128M, and 998M tokens on MMLU, MMLU-Pro, GSM8K, and WorkBench (Appendix F, Table 7).

  2. Break-even workloads (Table R3, see our response to Reviewer wSTn, W1 and Q1). On WorkBench, TCP-MCP breaks even against G-Designer and DyLAN after about 23,519 and 28,598 tasks while keeping a success-rate advantage of 12.50 and 14.04 points.

We have moved the per-task search costs and the break-even workloads into Section 5.2.

W2 and Q4: Benchmark coverage

Finally, MMLU and MMLU-Pro share a substantial number of underlying questions, which limits the diversity of the evaluation. More independent benchmarks would better support the generality of the method.

Can the authors evaluate on other tasks, like coding or general reasoning, to better support the generality of the method?

We agree on both points. We have not evaluated coding tasks, and MMLU and MMLU-Pro are not independent benchmarks.

  1. WorkBench covers tool use, not code generation. Success is judged from the final environment state on 520 tasks. TCP-MCP reaches 77.31% against 64.81% for G-Designer and 63.27% for DyLAN, and EvoMAS reaches 57.24% with 16.8 times the deployment tokens (Appendix N).

  2. MMLU and MMLU-Pro overlap. The shared questions also limit the transfer result between them in Table 2.

  3. One search backbone. All searches use DeepSeek-V3.2. The GPT-4 results in Table 3 test the reuse of searched structures, not independent search on another backbone.

We have added these limits to Section 6.

Reviewer_a9HLOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TCP-MCP, a framework for jointly optimizing agent prompts and communication topologies in LLM-based multi-agent systems. Instead of searching prompts and graph structures separately, TCP-MCP represents them as a unified executable genome and evolves them together. The search uses structural crossover, topology and role mutations, and Pareto-based optimization over task performance, token cost, and structural complexity. An initialization-time landscape probe and later Pareto-front diagnostics are used to adjust the search process. After search on a development set, one prompt-topology genome is selected and frozen for evaluation on unseen queries. The method is evaluated on MMLU, MMLU-Pro, GSM8K, and WorkBench using DeepSeek-V3.2. The results show improvements over automated topology-design baselines, especially G-Designer, and competitive performance compared with more expensive multi-agent methods. The paper also includes cross-dataset and cross-backbone transfer, repeated evolutionary runs, ablations, and a separate comparison with EvoMAS.

Critical strengths
  1. The prompt/topology coupling is a good motivation for the method, and the staged-optimization comparison makes it more than a conceptual argument. A communication edge may be useful under one receiving-agent prompt and useless under another, so optimizing the two in isolation can miss workable configurations. I found the comparison against prompt-first, topology-second search particularly relevant; the joint method is better across the four reported tasks, although the size of the gain varies.
  2. The evaluation goes further than I expected for an evolutionary MAS paper. There are operator ablations, repeated searches, transfer tests, and a comparison with EvoMAS. Importantly, the authors distinguish the one-time search from the frozen system used at test time, and Appendix F gives search-token and wall-clock figures. Those costs are substantial, but reporting them is the right choice and makes the performance numbers easier to interpret.
Critical weaknesses
  1. My main concern is the role of 'landscape-guided' search in the final result. FDC is computed once, at initialization, while the rest of the evolutionary procedure continues to adapt. Appendix J isolates FDC, but only for the first three generations: the final reported hypervolume there is 0.903400 versus 0.883945 without the guidance. That is an early-search signal, not evidence that FDC improves the final selected genome. Since landscape guidance is prominent in the title, I would want a complete-run comparison, preferably over more than one seed. The current results make a stronger case for joint co-evolution than for this particular initialization mechanism.
  2. Practical efficiency needs more careful qualification. Appendix F reports about 1.96 billion search tokens in total. The headline token comparisons, however, are largely for the already-selected system at deployment. The paper does include a break-even analysis against EvoMAS on MMLU and MMLU-Pro, which addresses part of this issue. Still, the result will depend heavily on how many similar queries are processed before the prompt/topology genome has to be searched again. A useful offline design method need not have low search cost, but its intended operating regime should be clearer.
Questions for authors
  1. Could you run the FDC-versus-neutral initialization comparison through to the end of evolution and report the final accuracy/cost or hypervolume over several seeds? I would be interested to know whether the early advantage in Appendix J survives later search.
  2. How do the conclusions change under a smaller search budget? The current experiments make joint search look promising, but the relative benefit when the number of candidate evaluations is sharply reduced would matter in practice.
  3. Beyond the two EvoMAS break-even examples already provided, what workload sizes would make TCP-MCP economical, and when would a change in task distribution or backbone require a new search?
4 comments

We thank the reviewer for the discussion of prompt–topology coupling, for finding the comparison with prompt-first, topology-second search relevant, and for noting that the paper separates the one-time search from the frozen system and reports the search cost.

W1 and Q1: Final effect of FDC

My main concern is the role of 'landscape-guided' search in the final result. FDC is computed once, at initialization, while the rest of the evolutionary procedure continues to adapt. Appendix J isolates FDC, but only for the first three generations: the final reported hypervolume there is 0.903400 versus 0.883945 without the guidance. That is an early-search signal, not evidence that FDC improves the final selected genome. Since landscape guidance is prominent in the title, I would want a complete-run comparison, preferably over more than one seed. The current results make a stronger case for joint co-evolution than for this particular initialization mechanism.

Could you run the FDC-versus-neutral initialization comparison through to the end of evolution and report the final accuracy/cost or hypervolume over several seeds? I would be interested to know whether the early advantage in Appendix J survives later search.

To answer directly: yes. We ran the comparison to generation 21 with three runs per setting on all four tasks. The early advantage persists in final accuracy on MMLU-Pro and WorkBench (p = 0.004 and 0.005) and is not significant on MMLU and GSM8K (p = 0.08 and 0.18).

  1. Complete-run results (Table R6, see our response to Reviewer wSTn, W2 and Q3). FDC-guided initialization is 1.11, 2.28, 1.4, and 1.7 points more accurate than neutral initialization on MMLU, MMLU-Pro, GSM8K, and WorkBench. Mean final HV differences (FDC-guided minus neutral) are +0.0099, −0.0012, −0.0085, and +0.0171 on the four tasks, none of which are significant (Welch t-test p = 0.53, 0.95, 0.69, and 0.47).

  2. We agree with the reviewer's reading. The evidence for joint co-evolution is stronger than the evidence for FDC initialization. FDC only sets the initial crossover probability within 0.31–0.39 (Appendix M), and its effect on the final system is limited to MMLU-Pro and WorkBench.

  3. Landscape guidance beyond FDC. The runtime controller adjusts topology and role mutation rates from diversity, neutrality, multimodality, and ruggedness features of recent populations (Appendix C.6). Its contribution is tested by AB1 (fixed rates) and AB6 (fixed decay schedule), which are 0.18–1.73 points lower than Full on MMLU, MMLU-Pro, and GSM8K.

  4. Appendix J values. The generation-3 HV values in Appendix J are computed on MMLU-Pro under the default normalization in Appendix M (clipping the three-objective vector of accuracy, negative token cost, and negative log(1 + complexity) to fixed bounds, computing HV against the reference point (0, −2,500,000, −log 1001), and dividing by the fixed box volume 2,500,000 log 1001) and are not on the same scale as the final HV in Table 11 and Table R6.

We have added Table R6 to Appendix J and separated FDC-guided initialization from runtime adaptive control in Section 4.5. We have revised the abstract to describe landscape guidance as FDC-guided initialization together with runtime adaptive control and to report the effect of FDC only where it is significant.

W2 and Q3: Operating regime and when to search again

Practical efficiency needs more careful qualification. Appendix F reports about 1.96 billion search tokens in total. The headline token comparisons, however, are largely for the already-selected system at deployment. The paper does include a break-even analysis against EvoMAS on MMLU and MMLU-Pro, which addresses part of this issue. Still, the result will depend heavily on how many similar queries are processed before the prompt/topology genome has to be searched again. A useful offline design method need not have low search cost, but its intended operating regime should be clearer.

Beyond the two EvoMAS break-even examples already provided, what workload sizes would make TCP-MCP economical, and when would a change in task distribution or backbone require a new search?

To answer directly: TCP-MCP is economical when the alternative costs more per query and the workload exceeds the break-even points in Table R3, which are 669–4,122 queries against EvoMAS, 8,539–15,507 against LLM-Debate, and 23,519–28,598 tasks against G-Designer and DyLAN on WorkBench. Against cheaper baselines such as Self-Consistency, the Table 1 operating point does not recover the search cost, though deploying lighter candidates from the same Pareto archive (Table R2) yields finite break-even points whenever the candidate uses fewer tokens per query than the target baseline.

  1. Break-even workloads (Table R3, see our response to Reviewer wSTn, W1 and Q1). At an illustrative 1,000 similar tasks per day, the WorkBench thresholds correspond to 4–29 days.

  2. Reuse without a new search. Transferring a genome between MMLU and MMLU-Pro lowers accuracy by 1.20–1.54 points (Table 2), although these two benchmarks share questions. Structures searched with DeepSeek-V3.2 and reused under GPT-4 outperform the other methods in Table 3 on all four tasks.

  3. When to search again. We evaluate the frozen system on a fixed monitoring set from the target task and keep it while it meets the performance floor, the cost ceiling, and the tool and output requirements. Otherwise we first evaluate archive candidates or a limited adaptation and run a full search only if these fail. Table R7 gives examples based on the transfer results.

We have added the break-even workloads to Section 5.2 and the reuse rule to Section 6.

Table R7: Examples of reuse and re-search decisions

Transfer resultRequirement and decision
MMLU to MMLU-Pro: 81.12% (native search 82.66%)Floor 80%: reuse. Floor 82%: evaluate archive candidates or adaptation, then search if needed.
GPT-4 on WorkBench: 73.85%Floor 73%: reuse. Floor 75%: adapt or search if cheaper options fail.
Monitoring and costMonitoring set of, for example, 100 random target queries. Cost ceiling of, for example, 20,000 tokens/query. Tool and output checks must pass.

The floors and the cost ceiling are illustrative requirements, not experimental results.

Q2: Smaller search budgets

How do the conclusions change under a smaller search budget? The current experiments make joint search look promising, but the relative benefit when the number of candidate evaluations is sharply reduced would matter in practice.

To answer directly: we have not completed independent reduced-budget searches evaluated on the held-out sets, so we cannot yet report how the gap between joint and staged search changes under a sharply reduced budget.

  1. What the 21-generation runs show. The HV checkpoints in Appendix L (Table 11) come from the full runs, not from separate smaller-budget searches. By generation 11, MMLU and GSM8K reach 98.3% and 93.5% of their HV gain between generations 1 and 21, while MMLU-Pro reaches only 55.8%. MMLU-Pro is therefore the task most likely to be affected by a smaller budget.

  2. Convergence statement. The checkpoints do not show convergence at generation 21 on MMLU-Pro. We have removed the statement in Section 5.3 that the Pareto front stabilizes by generation 21 and now state only that results are reported at generation 21.

Thank you again for your constructive feedback. We have directly addressed your central concerns by extending the FDC-versus-neutral comparison through all 21 generations, with three independent runs per setting across four benchmarks. The results establish statistically significant final-accuracy improvements on MMLU-Pro and WorkBench, while we have carefully distinguished the nonsignificant results on the other tasks. We have also added comprehensive break-even analyses across multiple baselines and a concrete framework for deciding when an optimized genome can be reused or requires a new search.

We sincerely appreciate the time and thought you have invested in evaluating our work. We would be very grateful if you could take another look at these additional experiments and let us know whether they address your concerns. Thank you again for your thoughtful feedback and consideration.

Thank you again for your time. We would like to kindly follow up on our previous response and the additional experiments we have provided.

We have made substantial efforts to address the concerns raised in your review, including additional multi-seed evaluations, comprehensive ablation studies, and expanded analyses of search costs and deployment efficiency. We believe these new results provide considerably stronger evidence addressing the questions you raised.

We would sincerely appreciate your feedback on whether the additional evidence addresses your concerns and whether there are any remaining issues that you believe require further clarification. Your assessment would be particularly valuable to us as the discussion approaches its conclusion.

Thank you again for your consideration and for taking the time to evaluate our work.

Reviewer_wSTnOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes TCP-MCP, a framework for jointly optimizing agent prompts and communication topologies in LLM-based multi-agent systems. Rather than treating prompt design and communication structure as independent optimization problems, TCP-MCP represents them as a unified executable genome and employs evolutionary search to optimize task accuracy, token consumption, and structural complexity simultaneously.

The framework combines structure-aware crossover, topology and role mutation, prompt inheritance and regeneration, and NSGA-II-based multi-objective selection. It additionally introduces an initialization-time Fitness Distance Correlation (FDC) probe and Pareto-front-based adaptive control to regulate exploration and mitigate premature convergence. The resulting prompt–topology configuration is selected on development data and then reused for unseen queries without additional evolutionary optimization.

The authors evaluate TCP-MCP on MMLU, MMLU-Pro, GSM8K, and WorkBench using DeepSeek-V3.2 as the primary backbone. The method achieves 89.96%, 82.66%, 96.61%, and 77.31% accuracy or task success, respectively. The paper also provides comparisons with existing multi-agent methods, ablation studies, robustness experiments, and cross-dataset and cross-model transfer evaluations. The results suggest that jointly optimizing agent behavior and communication structure can improve performance, although the advantages vary across tasks and involve nontrivial computational costs.

Critical strengths
  • Clear reporting of computational costs: One strength of the paper is that it reports not only inference costs but also the costs of evolutionary search, including token usage, API calls, and running time (Tables 6–8). The authors also analyze when the initial search cost can be recovered compared with EvoMAS (Table 17). In addition, the LLM-Debate comparison (Table 18) shows that debate can achieve higher accuracy with a larger budget. These analyses help readers better understand the practical costs and trade-offs of the proposed method.

  • Comprehensive experiments and reproducibility: The paper evaluates TCP-MCP against 14 baselines using the same DeepSeek-V3.2 model across four benchmarks, including WorkBench, which tests practical tool-use tasks. The authors also report results from multiple search runs, repeated evaluation, and transfer experiments, providing some evidence of stability and generalization. The supplementary materials include code and supporting files, which are helpful for reproducibility.

Critical weaknesses
  • Unclear performance–efficiency trade-off: While TCP-MCP outperforms G-Designer across all four benchmarks, its advantages over other strong baselines are less convincing. For example, Self-Consistency achieves slightly higher accuracy on MMLU (90.36% vs. 89.96%) with roughly 7× fewer tokens. Similarly, the improvements over DyLAN on MMLU-Pro and PHP on GSM8K are relatively small compared to the additional inference cost. Moreover, the main comparisons do not fully account for the substantial evolutionary search cost reported in Appendix F, making it difficult to assess the overall efficiency of the proposed approach.

  • Limited evidence for the effectiveness of key components: The ablation studies do not fully establish the benefits of joint co-evolution and landscape-guided adaptation. AB4 considers only one staged optimization strategy without matching actual computational costs, and the improvements on MMLU-Pro (+0.36 percentage points) and GSM8K (+0.34 percentage points) are relatively small. Additionally, the FDC ablation only examines the first three generations, leaving its impact on final performance unclear. More controlled comparisons and multi-seed evaluations would help clarify the contributions of these components.

Questions for authors
  • Can the authors provide accuracy–cost comparisons with strong baselines (e.g., Self-Consistency, DyLAN, and G-Designer) under matched token budgets? How many deployment queries are needed for TCP-MCP to offset its initial search cost compared with these methods?

  • The AB4 experiment compares joint optimization against a particular sequential strategy, but it is unclear whether the improvement specifically results from simultaneous co-evolution or from differences in search behavior and computational expenditure. Can the authors compare joint co-evolution with other optimization strategies, such as alternating prompt and topology optimization, under matched computational budgets? Would the improvements remain consistent across multiple random seeds?

  • Appendix J shows an early-generation hypervolume improvement from FDC-guided initialization, but does not establish its effect on the final generation-21 solution. Can the authors clarify whether FDC guidance consistently improves final accuracy, Pareto hypervolume, or convergence efficiency across independent runs? Additionally, how much of the improvement from adaptive control can be attributed specifically to landscape-based guidance rather than a simpler fixed or scheduled mutation strategy?

8 comments

We thank the reviewer for the careful reading of the paper and supplementary material, and for noting the reporting of search-time tokens, API calls, and runtime (Tables 6–8), the break-even analysis against EvoMAS, and the comparison with 14 baselines on four benchmarks.

W1 and Q1: Performance–cost trade-off and search cost

Unclear performance–efficiency trade-off: While TCP-MCP outperforms G-Designer across all four benchmarks, its advantages over other strong baselines are less convincing. For example, Self-Consistency achieves slightly higher accuracy on MMLU (90.36% vs. 89.96%) with roughly 7× fewer tokens. Similarly, the improvements over DyLAN on MMLU-Pro and PHP on GSM8K are relatively small compared to the additional inference cost. Moreover, the main comparisons do not fully account for the substantial evolutionary search cost reported in Appendix F, making it difficult to assess the overall efficiency of the proposed approach.

Can the authors provide accuracy–cost comparisons with strong baselines (e.g., Self-Consistency, DyLAN, and G-Designer) under matched token budgets? How many deployment queries are needed for TCP-MCP to offset its initial search cost compared with these methods?

To answer directly: at matched cost, TCP-MCP has a clear advantage on MMLU-Pro and WorkBench, while Self-Consistency and G-Designer match or exceed it on MMLU and GSM8K. At the Table 1 operating points, the one-time search cost is recovered on WorkBench against G-Designer and DyLAN (and against LLM-Debate on all four tasks and against EvoMAS on the three evaluated tasks), while deploying lighter candidates from the same Pareto archive (Table R2) also recovers the search cost on the reasoning tasks whenever the selected candidate uses fewer tokens per query than the target baseline.

  1. At the reported operating points, the gains are concentrated on MMLU-Pro and WorkBench. On MMLU, Self-Consistency reaches 90.36% with 12.47M tokens and TCP-MCP reaches 89.96% with 88.03M tokens (Table 1), so Self-Consistency is the more cost-effective choice on this task. The gains over DyLAN on MMLU-Pro (0.30 points) and over PHP on GSM8K (0.40 points) also come at a higher deployment cost. TCP-MCP outperforms Self-Consistency by 6.54 points on MMLU-Pro and 30.00 points on WorkBench, and it outperforms DyLAN and G-Designer on WorkBench by 14.04 and 12.50 points with 69.5% and 73.5% fewer deployment tokens.

  2. Baselines reconfigured toward TCP-MCP's per-query cost (Table R1). The reference costs are 6,305, 8,085, 4,855, and 15,289 tokens/query. Self-Consistency uses 11, 11, 15, and 46 samples, and the realized costs of the baselines range from 33.7% to 185.5% of the reference. TCP-MCP leads all three baselines by at least 2.63 points on MMLU-Pro and has the highest success rate on WorkBench. On MMLU, Self-Consistency (91.00%) and G-Designer (93.32%) exceed TCP-MCP (89.96%). On GSM8K, Self-Consistency ties TCP-MCP at 96.61% and G-Designer reaches 97.74%. DyLAN does not improve with the larger budget on the three reasoning tasks.

  3. TCP-MCP candidates at each baseline's per-query cost (Table R2). For each baseline cost in Table 1, we take the candidate from the development-set Pareto archive of TCP-MCP whose cost is closest to the target and evaluate it on the held-out set. At these costs TCP-MCP is more accurate than all three baselines on GSM8K (the candidate matched to DyLAN uses 80.2% more tokens), more accurate than Self-Consistency on MMLU-Pro (79.00% and 76.12% at similar cost), and less accurate than all three baselines on MMLU. The main operating points were fixed on the development set by the rule in Appendix I, and held-out results were not used for selection. Some archive candidates therefore score higher on the held-out set than the reported operating point, for example 97.98% at 1,337 tokens/query on GSM8K.

  1. Break-even workloads (Table R3). With the baseline setup cost set to zero, the break-even workload is the search cost of TCP-MCP divided by the difference in per-query cost, and it exists whenever the deployed candidate costs less per query than the baseline. At the Table 1 operating points, Self-Consistency, DyLAN, and G-Designer cost less per query than TCP-MCP on the three reasoning tasks, whereas selecting lower-cost candidates from the Pareto archive (Table R2) yields finite break-even points without re-running search, for example about 156,311 queries against G-Designer on GSM8K (2,762 vs. 3,581 tokens/query) while retaining higher accuracy. On WorkBench, TCP-MCP breaks even after about 23,519 tasks against G-Designer and 28,598 against DyLAN while keeping its success-rate advantage. The break-even points are 8,539–15,507 queries against LLM-Debate and 669–4,122 against EvoMAS. At an illustrative 1,000 similar tasks per day, the WorkBench thresholds correspond to 4–29 days, after which all subsequent queries yield net token savings.

We have added Tables R1–R3 to a new Appendix P, moved the per-task search costs (Appendix F, Table 7) and the break-even workloads into Section 5.2, and stated there that Self-Consistency is more cost-effective on MMLU. We have also restricted the statements in the abstract and Section 5.2 that TCP-MCP outperforms G-Designer and other automated graph-generation baselines to the operating points in Table 1.

Table R1: Baselines reconfigured toward the per-query cost of the TCP-MCP main configuration

BenchmarkMethodAccuracy (%)Tokens/queryCalls/queryCost / reference (%)
MMLUTCP-MCP89.966,3054100.0
MMLUSelf-Consistency91.005,0171179.6
MMLUDyLAN86.657,89116125.2
MMLUG-Designer93.326,46915102.6
MMLU-ProTCP-MCP82.668,0854100.0
MMLU-ProSelf-Consistency80.016,8041184.2
MMLU-ProDyLAN80.0310,09016124.8
MMLU-ProG-Designer80.017,5281293.1
GSM8KTCP-MCP96.614,8553100.0
GSM8KSelf-Consistency96.614,6941596.7
GSM8KDyLAN93.314,90115101.0
GSM8KG-Designer97.744,5911594.6
WorkBenchTCP-MCP77.3115,2897100.0
WorkBenchSelf-Consistency76.7328,36846185.5
WorkBenchDyLAN60.005,154733.7
WorkBenchG-Designer50.0018,84629123.3

Table R2: Closest-cost TCP-MCP candidates at each baseline's per-query cost

DatasetTarget baselineBaseline: acc. (%) / tokensTCP-MCP candidate: acc. (%) / tokensCost differenceCalls
MMLUSelf-Consistency90.36 / 89390.20 / 3,246+263.5%4
MMLUDyLAN90.27 / 95089.70 / 1,552+63.4%2
MMLUG-Designer87.90 / 2,00884.00 / 2,399+19.5%3
MMLU-ProSelf-Consistency76.12 / 3,79779.00 / 3,675−3.2%4
MMLU-ProDyLAN82.36 / 2,29479.50 / 1,777−22.5%2
MMLU-ProG-Designer79.32 / 2,73277.00 / 2,781+1.8%3
GSM8KSelf-Consistency96.05 / 1,58197.98 / 1,337−15.4%2
GSM8KDyLAN95.78 / 74297.98 / 1,337+80.2%2
GSM8KG-Designer95.81 / 3,58195.97 / 2,762−22.9%4

Baseline values are from Table 1, where Self-Consistency uses k = 5 samples. Tokens are per query.

Table R3: Token break-even workloads (queries or tasks)

BaselineMMLUMMLU-ProGSM8KWorkBench
Zero-shotNoneNoneNoneNone
PHPNoneNoneNoneNone
Complete GraphNone166,28297,389None
GPTSwarmNoneNoneNoneNone
LLM-BlenderNoneNoneNoneNone
G-DesignerNoneNoneNone23,519
DyLANNoneNoneNone28,598
Complex CoTNoneNoneNoneNone
Self-ConsistencyNoneNoneNoneNone
LLM-Debate15,50712,8218,53913,867
Random Graph29,34224,97118,122None
AutoGenNoneNoneNoneNone
ChatDevNoneNoneNone276,095
MetaGPTNoneNoneNone95,066
EvoMAS2,784669Not evaluated4,122
"None" means the baseline costs less per query than TCP-MCP. Baseline setup cost is set to zero. EvoMAS on MMLU and MMLU-Pro uses the matched 500-query subsets (Appendix N). A break-even point does not imply equal accuracy.

W2 and Q2: Joint and staged optimization

Limited evidence for the effectiveness of key components: The ablation studies do not fully establish the benefits of joint co-evolution and landscape-guided adaptation. AB4 considers only one staged optimization strategy without matching actual computational costs, and the improvements on MMLU-Pro (+0.36 percentage points) and GSM8K (+0.34 percentage points) are relatively small.

The AB4 experiment compares joint optimization against a particular sequential strategy, but it is unclear whether the improvement specifically results from simultaneous co-evolution or from differences in search behavior and computational expenditure. Can the authors compare joint co-evolution with other optimization strategies, such as alternating prompt and topology optimization, under matched computational budgets? Would the improvements remain consistent across multiple random seeds?

In short, joint optimization is ahead of both staged schedules on MMLU and of the alternating schedule on MMLU-Pro while using fewer search tokens, and the GSM8K differences are within 0.34 points. All ablations are single runs.

  1. AB5 adds the alternating schedule. AB4 optimizes the role–prompt configurations P for 11 generations under the topology generated at initialization and then optimizes the topology G for 10 generations with P fixed. AB5 starts from the same topology and alternates every generation, optimizing P in odd generations and G in even generations, so P is updated in 11 generations and G in 10. Both use the same backbone, development split, population and elite sizes, NSGA-II selection, and search-budget cap as Full.

  2. Full uses less search compute than AB4 and AB5 (Table R5). Search tokens are 351M for Full against 378M and 386M on MMLU, 486M against 532M and 527M on MMLU-Pro, and 128M against 157M and 158M on GSM8K. On WorkBench the three are within 1% of each other (998M, 994M, and 1,003M). The accuracy gap is therefore not explained by a larger search expenditure for Full.

  3. Accuracy (Table R4). Full exceeds AB4 by 2.48, 0.36, 0.34, and 5.96 points and AB5 by 2.30, 1.06, 0.16, and 0.19 points on MMLU, MMLU-Pro, GSM8K, and WorkBench. The MMLU-Pro difference against AB4 and both GSM8K differences are within the run-to-run variation of Full (standard deviation 0.47 on MMLU-Pro in Table 5 and 1.2 on GSM8K in Table R6), so we do not read them as gains.

  4. Seeds. AB4–AB7 are single runs, and we could not complete additional ablation seeds within the rebuttal period, so Table R4 supports point comparisons only.

We have added AB5 and the search-token accounting to Section 5.4 and Appendix K, and Section 5.4 now states that the ablations are single runs.

Table R4: Accuracy of the staged, alternating, and guidance ablations (single runs)

SettingMMLUMMLU-ProGSM8KWorkBench
Full (main run, Table 4)89.9682.6696.6177.31
AB4: P then G (Table 4)87.4882.3096.2771.35
AB5: alternating P and G87.6681.6096.4577.12
AB6: fixed mutation-rate decay88.6781.6595.7375.38
AB7: without landscape guidance88.2076.2192.6671.35

MMLU, MMLU-Pro, and GSM8K report accuracy. WorkBench reports raw success on 520 tasks.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.