acceptodds
Under review as a conference paper at ICLR 2027

EvolveMAS: A Unified Self-Evolution Framework for Decentralized Multi-Agent Systems

Abstract

Decentralized multi-agent systems (MAS) enable large language model agents to interact without centralized control. Agents need to adapt to other agents at test time and reuse accumulated experience to refine their interaction strategies for future interactions across cooperative, mixed-motive, and adversarial settings. However, existing approaches often develop adaptation mechanisms for particular interaction settings. This shared requirement raises a central research question: how can a decentralized MAS framework support both test-time adaptation and reliable experience reuse across various interaction settings? To this end, we first propose EvolveMAS, a unified self-evolution framework that supports experience-driven adaptation within and across episodes. EvolveMAS pairs each main agent with a dedicated MemoryAgent for test-time memory governance, which maintains isolated Structured Memory outside the main agent’s reasoning context. EvolveMAS organizes adaptation into two complementary processes: Within-Episode Online Adaptation and Cross-Episode Knowledge Evolution. During Within-Episode Online Adaptation, MemoryAgent updates episodic, semantic, and working memory from interaction evidence to inform subsequent actions. During Cross-Episode Knowledge Evolution, the main agent uses memory and episode outcomes to distill reusable interaction knowledge comprising partner profiles and interaction strategies for subsequent episodes. Across five benchmarks and six evaluation configurations spanning the three interaction settings, EvolveMAS achieves the highest normalized average score of 77.2 among the compared methods. Ablations show that test-time memory governance and cross-episode knowledge reuse both improve performance. Interaction traces further illustrate how updated memory and distilled knowledge guide subsequent interactions. Overall, the experimental results support a unified self-evolution framework for experience-driven adaptation in a decentralized MAS environment without updating model parameters. To facilitate future research, our code is available at https://anonymous.4open.science/r/artifact_7f3a/.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Decentralized Multi-Agent Systems, Large Language Models, Self-Evolution, Test-Time Adaptation, Memory Governance

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_76r2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces EvolveMAS, a unified self-evolution framework for decentralized multi-agent systems. Motivated by the fact that existing approaches build adaptation mechanisms for one particular interaction regime, and that accumulating long interaction histories inside the reasoning context causes context bloat and memory pollution, the authors ask how a single mechanism can support both test-time adaptation and reliable experience reuse across cooperative, mixed-motive, and adversarial settings. EvolveMAS pairs each main agent with a dedicated MemoryAgent that maintains Episodic, Semantic, and Working memory outside the agent's reasoning context and updates it turn by turn (Within-Episode Online Adaptation), while at episode end the main agent distills reusable Interaction Strategies and Partner Profiles conditioned on a task-native outcome signal (Cross-Episode Knowledge Evolution) — all without updating model parameters. Its contributions are the resulting Execute–Evaluate–Evolve framework, the external MemoryAgent governance design, and an evaluation across five benchmarks and six configurations spanning the three regimes, where EvolveMAS attains the highest normalized average score of 77.2.

Critical strengths
  • The evaluation breadth is stronger than most work in this line. Five benchmarks spanning cooperative, mixed-motive, and adversarial regimes — BBH, HumanEval, SOTOPIA, LLM-Deliberation, and Werewolf — are genuinely heterogeneous rather than variants of a single paradigm.
  • The supplementary system-level analyses are practically useful. The team-size sweep (stable at 83–86 across 1–6 agents), the comparison against static MoA aggregation (84.0 vs. 81.0), and the cost/context-occupancy trajectories provide information that reviewers typically want but most papers omit.
  • The code is released via an anonymous repository, which partially supports reproducibility. This is preferable to the common "available upon acceptance" practice, though reproducibility remains partial because the MemoryAgent prompts and the distillation procedure are not specified in the paper.
Critical weaknesses
  • Weak statistical support and no repeated-trial statistics. Each cell of the main ablation comes from a single execution, and column means are reported without SDs or confidence intervals. The paper's own precision analysis gives a standard error of about 4 points on BBH, yet the BBH gain over the native baseline is only +4, and the reported McNemar tests find almost nothing significant. The one confidence interval that is reported treats 50 sessions of a single evolving stream as independent samples, so the effective sample size is one, not fifty.

  • Baselines miss the paper's own contribution class. All five baselines are MAS architecture families transferred across settings; no memory or self-evolution method is compared, even though ReasoningBank, A-Mem, Mem0, ExpeL, and Reflexion are all cited. The result therefore shows EvolveMAS beats transferred scaffolds, not that it beats alternative memory mechanisms. Two stated contributions are also left unsupported: "unified" is never tested by transferring knowledge across regimes, and the "plug-in extension" claim is never tested by attaching the memory layer to any baseline family.

Questions for authors
  • Could the author report mean ± SD over at least three seeds for every cell of Table 1, with paired significance tests for the Werewolf and SOTOPIA columns? Given the standard errors quoted in Appendix A.1 (≈4 points on BBH, ≈5 on each Werewolf audit), how much of the headline +9.7 survives? Also, could the LLM-Deliberation ablation be repeated over independent streams, so that the stream rather than the session is the replication unit?
  • All five baselines are MAS architecture families; no memory or self-evolution method is compared despite ReasoningBank, A-Mem, and Mem0 being cited. Could the author add one such baseline, and attach the EvolveMAS memory layer to a baseline family (e.g., AgentNet) to separate the contribution of memory governance from that of native team structure?
5 comments
AuthorsOpenReview ↗

Comment 1|Overview

We thank the reviewer for the detailed assessment, particularly the recognition of our evaluation breadth and supplementary system-level analyses. We appreciate the concerns regarding statistical reliability, baseline selection, and the interpretation of our Unified and Plug-in claims. We would like to clarify several important points supported by both the original manuscript and our additional experiments.

First, we have strengthened the evidence for experimental reproducibility. We conducted two additional independent complete evolution runs across all six evaluation configurations, giving three runs in total. The normalized average scores are 77.2, 79.1, and 77.2, consistently exceeding the strongest transferred MAS baseline (67.5). We distinguish these independent-run statistics from the task-level uncertainty and within-stream variability originally reported in Appendix A.1, without treating dependent sessions as independent evolution runs.

Second, EvolveMAS does not introduce a different native MAS team architecture. As specified in Sec. 4.1, Self-Memory preserves the same interaction structure, while Figure 4 and Table 6 evaluate memory governance and cross-episode Knowledge Evolution under controlled configurations. These comparisons directly examine the proposed mechanisms without changing the underlying team organization.

Third, Unified and Plug-in have specific architectural meanings. Unified refers to applying the same agent-local evolution mechanism across heterogeneous interaction regimes, not transferring task-specific knowledge between unrelated environments. Plug-in refers to a modular, training-free extension operating through local interaction trajectories, memory-conditioned decisions, and task-native feedback, without imposing centralized policy control or a fixed team topology.

Our principal contribution remains a unified framework for continuous, interaction-driven adaptation among autonomous decentralized agents. The following responses provide the detailed statistical results, controlled comparisons, and methodological reasoning supporting these distinctions.

AuthorsOpenReview ↗

Comment 2|Response to Weaknesses (1/2): W1

W1 (statistical support):

We clarify that our paper already reports uncertainty analyses, and that our protocol follows prior test-time self-evolution studies. Appendix A.1 (L871–879) reports the sampling precision of each column, and Appendix E.1 further reports paired McNemar tests.

Evaluating a single evolution run is the common practice in self-evolving agent studies. For example, ReasoningBank (ICLR 2026) evaluates test-time self-evolution on sequential task streams and reports point estimates without run-to-run variance, and Unified Context Evolution (EMNLP 2026) reports a single evolution process with task-level McNemar and paired bootstrap tests rather than repeated evolution runs. Because each task in our setting is encountered only once along a continuous experience stream, one complete run over the benchmark is the natural unit of evaluation.

Nevertheless, to further address your concern, we conducted two additional independent full evolution runs for all six configurations (three runs in total), including complete LLM-Deliberation streams, so that the evolution stream rather than the session is the replication unit. Across the three runs, the normalized average is 77.9 ± 1.1 (77.2, 79.1, and 77.2), and every run exceeds the strongest transferred MAS baseline (67.5) by 9.7–11.6 points. The per-benchmark results (mean ± SD) are listed in our response to Q1. We have reported these results in Appendix H of the revised manuscript.

AuthorsOpenReview ↗

Response to Weaknesses (2/2): W2

Response to Weaknesses (2/2)

W2 (memory baselines, Unified, and Plug-in):

We would like to clarify that the central contribution of EvolveMAS is a unified self-evolution framework for decentralized MAS, not a standalone memory method, and both claims should be read in this sense.

  1. Unified refers to the mechanism, not to knowledge transfer. Across settings, the self-evolution process, memory representation, and update schedule remain unchanged (Sec. 4.3, L404). Our paper never claims that Knowledge evolved in one regime transfers to another; prior self-evolution methods such as Reflexion likewise reuse experience within the evaluated tasks.
  2. Plug-in follows from the architecture. EvolveMAS preserves the native decentralized organization and pairs each main agent one-to-one with a dedicated MemoryAgent (Sec. 3.1, L187–191). In every column of Table 1, it runs on the same team as the native system and only adds per-agent memory governance and evolution. This is why we describe it as a plug-in extension to existing agents (Conclusion, L481), consistent with modular memory augmentation in prior work such as Proactive Memory Agent.
  3. Memory baselines target a different task. Mem0 and A-MEM are designed and evaluated for question answering over accumulated conversation sessions, whereas our benchmarks require memory to be updated during ongoing multi-agent interactions. Their published implementations therefore cannot be transferred as equivalent baselines without redesigning their workflows. Our memory-governance claim is instead supported by controlled comparisons with the same interaction structure: Self-Memory (Sec. 4.1, L316), Figure 4, Table 6, and Appendix F.
Reviewer_r87KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors discuss LLM agents in a decentralized setting, where agents need to adapt to other agents at test time. The work primarily focuses on designing mechanisms to enable experience reuse. The paper proposes EvolveMAS, a self-evolution framework that uses a dedicated MemoryAgent to organize experience for each agent in the system. There are two ways of considering the experience: 1) Within Episode Online Adaptation, where the structured memory is partitioned into 3 parts, including the episodic, semantic, and working memory, and is organized by the MemoryAgent; and 2) Cross-Episode Knowledge Evolution, where the main agent distills knowledge based on interaction and episode outcome. EvolveMAS is tested across 5 benchmarks with 6 configurations and 3 interaction settings (cooperative, mixed-motive, and adversarial), and achieves better performance compared to baselines.

Critical strengths
  1. Using MemoryAgent to maintain agent memory is inspiring and seems generalizable.
  2. The experiments tested across 5 benchmarks with 6 configurations and 3 interaction settings are comprehensive.
Critical weaknesses
  1. Does MemoryAgent prepare information every time the main agent is triggered? This seems to add latency to the system.
  2. The episodic memory feels like it's going to grow untrackably, as new events keep getting appended to Episodic Memory. Both MemoryAgent and the main agent seem to read from it. Do they always read the entire episodic memory, or is some additional filtering performed beforehand?
  3. A lot of the consolidation/retrieval seems entirely determined by the LLM. For example, the main agent distills reusable regularities from its structured memory. It is fine if agents do the job, but I think the paper would benefit from evidence showing how agent-distilled knowledge is used and actually changes decisions as expected. For example, in the case study, the strategies in panel b include applying negotiation, partial commitments, and flexible offers. In future rounds with a similar scenario, would the main agent actually use the strategy, and would that affect performance?
Questions for authors
  1. Could the authors further elaborate on the definition of a Turn in the system? Can multiple agents act concurrently in a turn, or can a single agent perform multiple tool calls in a turn? What defines the transitions? This will affect how agents can possibly collaborate with each other, especially in a decentralized setting. Further, line 161 seems to suggest that the task determines which agents interact and when: does this mean there is some predefined collaboration protocol that agents follow? Could the authors clarify what decentralization means in terms of adapting to other agents at test time?
  2. Does a MemoryAgent have local memory beyond the main agent's memory that it manages?
  3. In the paper, all MemoryAgents share the same memory schema and governance protocol. Would different interaction settings allow different agents to benefit from different schemas or protocols? It would be nice to hear some discussion from the authors.
5 comments
AuthorsOpenReview ↗

Comment 1|Overview

We thank the reviewer for the accurate understanding of our decentralized test-time adaptation setting and for recognizing the value of our MemoryAgent design and comprehensive cross-regime evaluation. The reviewer's questions provide an opportunity to clarify several important aspects of the execution mechanism and agent-level adaptation.

First, EvolveMAS preserves decentralized decision-making and independent experience evolution. Each Main Agent acts based on locally observed interactions, maintains its own Structured Memory through a paired MemoryAgent, and independently evolves Partner Profiles and Interaction Strategies. The environment determines task-native interaction rules and episode/turn boundaries, but does not centrally determine agents' evolving knowledge or behavioral policies. The framework does not impose a universal agent-role assignment or fixed collaboration topology beyond the requirements of the underlying tasks.

Second, we provide concrete implementation and behavioral evidence. We report the measured overhead of synchronous MemoryAgent updates; our setting also supports an asynchronous MemoryAgent implementation, which we tested and which removes main-thread blocking in a controlled BBH evaluation. We also clarify the separation between persistent memory storage and actively consumed reasoning context. Importantly, our existing case studies and execution traces demonstrate the subsequent use of accumulated knowledge, complemented by controlled performance ablations.

Finally, we appreciate the suggestion regarding agent-specific specialization. Although the governance protocol is shared, agents differ in their capabilities, accumulated knowledge, and interaction strategies. Appendix D and additional cooperative execution traces illustrate how these individual differences can contribute to complementary capabilities at the MAS level.

We address the corresponding implementation details, evidence, and conceptual distinctions individually below.

AuthorsOpenReview ↗

Comment 2|Response to Weaknesses (1/2)

Response to Weaknesses (1/2)

W1 (MemoryAgent latency):

No. The MemoryAgent is not invoked before every main-agent action. It is a lightweight GPT-4o-mini model paired with each main agent (Sec. 4.1, L313), and it updates memory after new interaction evidence is produced, following each task's native protocol: BBH and SOTOPIA update after completed turns and at episode boundaries, HumanEval and Werewolf at episode boundaries, and LLM-Deliberation after each session.

In the synchronous implementation used for the main experiments, the median time per update is 2.25 s on BBH, 2.76 s on HumanEval, 2.87 s on SOTOPIA, 3.47 s on WW-V, 2.89 s on WW-W, and 2.0 s on LLM-Deliberation.

Moreover, since the MemoryAgent runs outside the main agent's reasoning context (Sec. 3.1, L190), our setting supports an asynchronous MemoryAgent implementation, and we tested this asynchronous version. Turn-level updates are queued and applied between turns, so memory updates do not block the main agent; episode-end evolution remains synchronous. On five BBH tasks under the same configuration:

MetricSynchronousAsynchronous
Median MemoryAgent update time1.98 s2.00 s
Recorded main-thread blocking / enqueue time397.9 s0.0012 s (enqueue only)
Average task wall-clock time169.6 s83.7 s (−50.6%)
Correct answers5/55/5

With asynchronous updates, main-thread blocking drops from 397.9 s to 0.0012 s (enqueue only), so the main agent proceeds without waiting for the MemoryAgent; the average task wall-clock time is roughly halved, with the same accuracy.


W2 (growth of Episodic Memory):

No. The main agent does not read the entire Episodic Memory, and the stored Episodic Memory is bounded by a sliding window. Structured Memory is maintained outside the main agent's reasoning context (Sec. 3.1, L190). The main agent acts on Working Memory, which holds the current focus and task-relevant views of accumulated memory and Knowledge (Sec. 3.2, L242), rather than the full interaction history. Since the update rules in Sec. 3.2 do not prescribe a specific storage implementation (L246–247), our implementation keeps Episodic Memory in a sliding window of the latest 100 lines. Each MemoryAgent update reads this window together with the newly observed interaction evidence (Eq. 2), so its input does not grow with the full interaction history.

This selective use is enabled by the structure of memory (Sec. 3.2): Episodic Memory records concrete interaction experiences, Semantic Memory consolidates reusable knowledge with confidence and provenance annotations, and Working Memory continuously updates the current task-relevant state. Cross-episode Interaction Knowledge further consolidates reusable partner and strategy regularities. As reported in Appendix C, this keeps acting-context occupancy lower than retaining raw interaction histories (Sec. 4.6, L468).

Regarding growth, our execution logs show how the memory evolves over time. The following table reports the median MEMORY.md size over the first and last 10 memory updates for one representative agent per setting, together with the corresponding Episodic entry counts:

SettingInitial size (chars)Final size (chars)Episodic entries (initial → final)
BBH2,35024,2975 → 50
HumanEval3,67048,3406 → 50
SOTOPIA1,47491,5482 → 46
WW-V5,67419,45220 → 72
WW-W3,03014,2266 → 50

Within this window, the retained Episodic entries never reach 100: they number 46–72 at the end of these runs and at most 86 at any point across all runs. Working Memory is refreshed to reflect the latest interaction state rather than indefinitely accumulating previous states. Moreover, current mainstream LLMs support very long context windows, e.g., 1M tokens for GPT-4.1 and Gemini 2.5 Pro. Even the largest memory store in our long-horizon runs (about 190K characters, roughly 50K tokens, for the SOTOPIA seller after 80 episodes) remains far below these limits.

Therefore, EvolveMAS does not claim constant-size persistent storage or lossless preservation of every historical detail. Its contribution is a structured memory-governance mechanism that organizes growing interaction experience and makes it selectively reusable, without requiring the main agent to repeatedly process the full interaction history. Further archival or retrieval policies could support substantially longer deployments.

AuthorsOpenReview ↗

Comment 3|Response to Weaknesses (2/2)

Response to Weaknesses (2/2)

W3 (whether distilled Knowledge changes later decisions):

Yes. Our paper provides evidence for both the later use of distilled Knowledge and its effect on performance.

The case study in Sec. 4.5 (Figure 5b, L457–461) follows exactly the scenario you describe. Experiences with a USD 10 deposit and a counteroffer revised from USD 44 to USD 45 are distilled into negotiation Strategies and a Partner Profile, which then guide subsequent offer formulation. In the later execution, the agent offers "38 and 10 down" (USD), applying the learned flexible-offer and partial-commitment regularities.

The same holds in cooperative settings, where knowledge about partners changes who does the task. In an additional BBH run, an agent handed a task to a partner with a stronger record on that task family, and the partner corrected the original answer. The complete trace is provided as Case 1 at the end of this response.

Regarding performance, under the same Deliberation protocol, adding Knowledge raises unanimity from 74.0% to 86.0% with the MemoryAgent (Figure 4; Sec. 4.4, L422–424), supporting the value of reusing negotiation regularities across episodes. Knowledge also does not keep growing the context: Working Memory only provides task-relevant views of it (Sec. 3.2, L242), and the full Knowledge files are retrieved when needed.

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

This paper introduces EvolveMAS, a self-evolution framework for decentralized multi-agent systems. The framework equips each agent with a dedicated MemoryAgent that maintains structured episodic, semantic, and working memory outside the agent's reasoning context, mitigating context growth and supporting effective experience reuse. Experiments across five benchmarks demonstrate that EvolveMAS achieves good performance.

Critical strengths
  1. EvolveMAS provides a consistent self-evolution mechanism across cooperative, mixed-motive, and adversarial multi-agent interactions.
  2. The dedicated MemoryAgent separates memory management from agent reasoning, mitigating context growth while enabling within-episode adaptation and cross-episode knowledge reuse.
  3. Easy to comprehend with figures.
Critical weaknesses
  1. In Table 6, adding cross-episode knowledge to Self-Memory reduces HumanEval accuracy from 97.0% to 95.7%, while full EvolveMAS improves over Self-Memory without knowledge by only 0.6%.
  2. Lack of uncertainty measure in experiments. Since some of the reported improvements are smaller than 1%, reporting statistical uncertainty and variance (e.g., standard deviations across multiple runs or confidence intervals) is essential to establish whether these gains are statistically significant and reliably demonstrate the effectiveness of the proposed method.
  3. Although EvolveMAS claims to mitigate memory pollution, it lacks explicit mechanisms for filtering unreliable or outdated information and experiments validating robustness to corrupted memories or deceptive interactions, particularly in adversarial settings.
  4. Missing Ablations for Interaction Strategies and Partner Profiles alone. They are ablated together, without independently evaluating the contributions of these components
  5. Limited Novelty. The core components of EvolveMAS substantially overlap with existing approaches: dedicated external memory management has been explored in [1], cross-episode experience distillation and reuse in [2,3], and decentralized agent evolution with persistent memory in [4]. While EvolveMAS integrates these ideas across different interaction settings, the paper does not clearly establish what fundamentally new algorithmic mechanisms are introduced beyond combining existing techniques, nor does it provide direct comparisons with these closely related methods to substantiate its methodological novelty.

[1] Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents. arXiv 2026. https://arxiv.org/abs/2607.08716

[2] ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory. ICLR 2026. https://arxiv.org/abs/2509.25140

[3] Agent Workflow Memory. ICML 2025. https://proceedings.mlr.press/v267/wang25bx.html

[4] AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems. NeurIPS 2025. https://arxiv.org/abs/2504.00587

Questions for authors
  1. What task-level feedback rer_e is available during cross-episode evolution, particularly on BBH and HumanEval? Does the framework rely on ground-truth correctness or other privileged evaluation signals?

  2. Since episodic and semantic memories may persist across episodes even when KeK_e is disabled, how can you isolate the contribution of cross-episode knowledge evolution from persistent memory reuse?

  3. How does MemoryAgent identify and filter incorrect, outdated, or deceptive information, particularly in adversarial interactions? Is there empirical evidence that it prevents memory pollution rather than merely compressing interaction history?

6 comments

We thank the reviewer for recognizing the consistency of our self-evolution mechanism across heterogeneous interaction settings and the clarity of our presentation. We appreciate the detailed concerns and would like to clarify the central contribution of EvolveMAS, which is important for interpreting both its novelty and empirical evidence.

EvolveMAS establishes a unified paradigm for interaction-driven self-evolution in decentralized multi-agent systems. Unlike conventional memory-centric approaches, our research focuses on how independently acting agents can continuously accumulate experience and adapt to other autonomous agents with potentially cooperative, partially conflicting, or adversarial objectives. The proposed Execute–Evaluate–Evolve framework connects within-episode online adaptation with cross-episode Interaction Knowledge evolution, supported by independent, agent-local memory governance. Our contribution lies in establishing and evaluating this unified adaptation mechanism across heterogeneous interaction regimes, rather than claiming the individual invention of external memory or experience distillation.

We also provide additional evidence addressing the reviewer's empirical concerns. Specifically, we conducted two additional independent evolution runs across all six evaluation configurations, demonstrating consistent overall performance. We further examined actual execution trajectories to clarify how task-native feedback is incorporated into experience evolution, how structured memory governs accumulated observations, and how accumulated experience informs subsequent decisions. The existing matched-structure comparisons and mechanism ablations provide complementary evidence for the contributions of memory governance and Interaction Knowledge evolution.

The following responses address each concern with specific methodological clarifications, experimental evidence, and references to the corresponding sections and appendices.

AuthorsOpenReview ↗

Response to Weaknesses (1/2): W1–W3

Response to Weaknesses (1/2)

W1 (Knowledge in the HumanEval ablation):

We clarify that our claim on cross-episode Knowledge is grounded in the controlled ablation and the full-configuration results. Adding Knowledge raises unanimity from 74.0% to 86.0% with the MemoryAgent (Figure 4; Sec. 4.4, L422), and the full configuration attains the highest score on LLM-Deliberation, BBH, and SOTOPIA (Appendix E.2, L1140–1141).

On HumanEval, the 1.3-point drop corresponds to only two tasks. HumanEval is near saturation (95.7–97.6% across the four variants), and as already noted in Appendix E.2 (L1143–1145), the four conditions differ by at most three of 164 tasks, which a single run per condition cannot distinguish from noise. It therefore does not indicate a systematic negative effect of Knowledge.

This result also reflects why we introduce the MemoryAgent. Self-Memory keeps memory inside the main agent's reasoning context (Sec. 4.1, L316), where accumulated histories cause context bloat that can affect task performance (Sec. 1, L76). Decoupling memory from reasoning allows Knowledge to be distilled from governed Structured Memory (Sec. 3.3, L261), which is why the full configuration performs best.


W2 (uncertainty measure):

We clarify that our paper already quantifies uncertainty, and that our protocol follows prior test-time self-evolution studies. Appendix A.1 (L871–879) reports the sampling precision of each column; on LLM-Deliberation, the gains of EvolveMAS over both the native baseline (+0.48) and Self-Memory (+0.50) have 95% intervals that exclude zero. Appendix E.1 further reports paired McNemar tests.

Evaluating a single evolution run is the common practice in self-evolving agent studies. For example, ReasoningBank (ICLR 2026) evaluates test-time self-evolution on sequential task streams and reports point estimates without run-to-run variance, and Unified Context Evolution (EMNLP 2026) reports a single evolution process with task-level McNemar and paired bootstrap tests rather than repeated evolution runs. Because each task in our setting is encountered only once along a continuous experience stream, one complete run over the benchmark is the natural unit of evaluation.

Nevertheless, to further address your concern, we conducted two additional independent full evolution runs for all six configurations (three runs in total):

BenchmarkRun 1 (Table 1)Run 2Run 3Mean ± SD
BBH (%)83.083.083.083.0 ± 0.0
HumanEval (%)97.097.097.097.0 ± 0.0
SOTOPIA0.6130.6350.6410.630 ± 0.015
LLM-Deliberation5.405.405.545.45 ± 0.08
WW-V (%)62.065.061.062.7 ± 2.1
WW-W (%)70.076.066.070.7 ± 5.0
Normalized Avg.77.279.177.277.9 ± 1.1

All three runs exceed the strongest transferred MAS baseline (67.5) by 9.7–11.6 points. We have reported these results in Appendix H of the revised manuscript.


W3 (memory pollution):

We would like to clarify the role of memory pollution in our paper. Memory pollution is a problem identified by prior work, and it motivates the design of the MemoryAgent (Sec. 1, L76–83; Sec. 2, L142–143). As stated in our paper:

"Retaining inaccurate accumulated experience introduces memory pollution that can misguide subsequent interactions (Xiong et al., 2026)." (Sec. 1, L76)

"However, in long-horizon decentralized interactions, retaining accumulated memory in the main agent's context can lead to context expansion, attention dilution, and memory pollution (Liang et al., 2026; Zhang et al., 2026; Wu et al., 2026; Wan et al., 2026)." (Sec. 2, L142–143)

EvolveMAS therefore pairs each main agent with a dedicated MemoryAgent that maintains Structured Memory outside the reasoning context (Sec. 1, L83).

Our experiments show that this decoupling is effective. Beyond the main results (Table 1), replacing Self-Memory with the MemoryAgent under the same interaction structure raises unanimity from 24.0% to 74.0% (Figure 4; Sec. 4.4, L421). On the LoCoMo memory benchmark, online MemoryAgent updates also achieve higher accuracy than offline updates and self-managed memory (Appendix F; Sec. 4.6, L471).

Designing dedicated filters for specific types of corrupted or deceptive information is a different research direction from our focus, which is a unified self-evolution framework for decentralized MAS.

AuthorsOpenReview ↗

Comment 3|Response to Weaknesses (2/2): W4–W5

Response to Weaknesses (2/2)

W4 (separate ablations of Strategies and Profiles):

We clarify that Interaction Strategies and Partner Profiles are jointly modeled as two complementary components of a single Interaction Knowledge K_e=(G_e,P_e) (Sec. 3.3, L263). Profiles capture how partners tend to behave, and Strategies capture how to act toward them; both are distilled together from the same interaction outcomes. Our claim therefore concerns the effectiveness of this Interaction Knowledge for adaptation in decentralized interactions, and we do not claim independent gains from either component.

The ablation tests exactly this claim. Figure 4 toggles the joint Knowledge (K=∅ vs. K_e=(G_e,P_e); Sec. 4.4, L415), and adding it raises unanimity from 74.0% to 86.0% with the MemoryAgent (L422). Ablating Strategies and Profiles separately is a finer-grained representation-design question, while the core goal of this paper is a unified, interaction-driven adaptation mechanism across heterogeneous settings. Separate ablations are therefore not required to support our claims.


W5 (novelty):

Thank you for pointing out these related works. We would like to clarify that the novelty of EvolveMAS lies in the research problem and the paradigm it establishes, rather than in any individual component. Our paper asks how a decentralized MAS framework can support both test-time adaptation and reliable experience reuse across interaction settings (Abstract, L17). It answers this question with a unified self-evolution framework that couples Within-Episode Online Adaptation and Cross-Episode Knowledge Evolution (Contribution 1, Sec. 1, L98–100).

The cited works address different problems. As discussed in our Related Work (Sec. 2, L123–143), existing memory, self-evolution, and MAS adaptation methods are designed for specific tasks or interaction settings, whereas EvolveMAS keeps the same decentralized organization and evolution framework across cooperative, mixed-motive, and adversarial regimes.

Our results support this paradigm. EvolveMAS achieves the highest normalized average score of 77.2 across the six configurations (Table 1), and the ablations (Figure 4; Table 6) isolate the contributions of memory governance and Knowledge evolution. Regarding direct comparisons, we further evaluated our MemoryAgent against dedicated memory systems, including Mem0 and A-MEM, in an independent evaluation under the LoCoMo protocol (F1/BLEU-1). MemoryAgent outperforms both in all four question categories with two backbones. We provide the results table in a separate comment.

Comments on the paper

1 comment

Dear Reviewers,

Thank you very much for the time and effort you have devoted to reviewing our paper. We sincerely appreciate your helpful comments and suggestions.

We have carefully responded to each of your comments under your respective reviews. For each concern, we provide detailed explanations and clarifications grounded in the manuscript, and, where requested, additional experimental results. In particular, we conducted two additional independent evolution runs for all six evaluation configurations, and all three runs exceed the strongest transferred MAS baseline (Appendix H of the revised PDF).

As the discussion deadline approaches, we would be very grateful if you could review our responses and let us know whether any concerns remain. We would welcome the opportunity to discuss them further and provide any additional clarification that may be helpful.

Thank you again for your time and consideration.

Best regards, The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.