acceptodds
Under review as a conference paper at ICLR 2027

DynaWorld: Towards Self-Evolving Multi-Agent Collaboration in Dynamic Multimodal Environments

Abstract

Autonomous agents powered by large language models (LLMs) are increasingly deployed in long-horizon interactive environments that evolve over time. However, most existing agents are developed and evaluated on **_isolated tasks_** under static conditions, which limits their ability to remain effective as environments evolve. To address this challenge, we introduce **DynaWorld**, a dynamic multimodal environment derived from **_real-world_** customer-service interactions, comprising eight long-horizon scenarios and 1,400 user-agent interaction episodes. By connecting successive episodes through controlled environmental changes, DynaWorld enables systematic evaluation of **_how agents adapt as the environment evolves_**. To further strengthen agent adaptation under such dynamics, we propose **DynEvo**, a self-evolving multi-agent framework that continually refines externalized experience and strategies from environmental feedback without updating model parameters. Extensive results on DynaWorld show that existing agents struggle under changing environments, whereas **DynEvo** consistently achieves stronger task performance than standard LLM agents and strong self-evolving baselines, while remaining interaction-efficient with fewer dialogue turns. Further analyses reveal shift-specific degradation and recovery patterns that static settings cannot capture, highlighting the need to evaluate and evolve multimodal agents in dynamic environments.

Primary area:
datasets and benchmarks
Keywords:
Dynamic Environments, Self-Evolving Agents, Multi-Agent Collaboration, Agent Evaluation

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

The paper proposed a new agent benchmark DynaWorld, where agents interact with user simulators under evolving conditions. To adept the evolving environment, the paper proposed DynEvo which combines multi-agent collaboration with online self evolution through a tool workbench. After each episode, the service agent reflects on the trajectory and episode-level feedback update notes and skills, while strategies are reviewed periodically. The experiments shows DynEvo achieves consistent gains across dynamic scenarios and outperforms existing evolving agents.

Critical strengths

S1. DynaWorld is the first interactive service benchmark where the environment itself changes across episodes. The event design (policy, user population, objective, time-limited rules) and the fixed event schedules make it possible to measure, in a repeatable way, how an agent’s performance drops and recovers when conditions change.

S2. The analysis by event type (Figures 4–5) gives results that static benchmarks cannot: changes in user population cause the largest and slowest-recovering drop, while time-limited changes are absorbed quickly. These findings are useful on their own, regardless of how DynEvo compares to other methods.

Critical weaknesses

W1. The claim that DynEvo beats “strong self-evolving baselines” is not supported. None of the four baselines keeps and reuses experience across episodes, while the methods that do, and that the paper itself cites (ACE, Memento, ExpeL, Mem0), are not compared. At least one such method should be run with the same base model and agent setup.

W2. The judge-based Overall Score is used both to evaluate all methods and to guide DynEvo’s updates between episodes. This gives DynEvo a feedback signal the baselines do not get, and it risks DynEvo learning to please the judge rather than to serve users better. No confidence intervals are reported, although they could be computed from the existing episodes without new runs.

Questions for authors

Is the judge (and the user simulator) also GPT-5.2? EvoAgentX is said to use episode-level feedback. Is that the same Overall Score DynEvo uses, and do the other three baselines get anything between episodes? How many runs are behind Table 3? Temperature alone moves the score by about a point (Table 11), and the main gap is 4.5.

3 comments
AuthorsOpenReview ↗

Author Response to Reviewer CWWJ

Dear Reviewer CWWJ:

We sincerely thank you for the insightful and constructive feedback! We appreciate your recognition of DynaWorld as the first interactive service benchmark where the environment itself changes across episodes, its reproducible event design, and the event-level insights that static benchmarks cannot provide. We address the concerns and clarify the experimental protocol below.


W1: Comparison with Experience-Learning Methods.

We thank the reviewer for this valuable suggestion. Among the existing baselines, AutoAgents and EvoAgentX both include memory and experience-reuse components (Appendix C.1, Table 10), and we have clarified these mechanisms in the revised manuscript. To compare with methods specifically designed for memory-based experience learning, we additionally evaluate ExpeL (AAAI 2024)[1] and ACE (ICLR 2026)[2] using the same GPT-5.2 backbone and settings as DynEvo, with results shown below.

Dynamic environments require both experience accumulation and timely strategy adjustment when conditions change. Memory-based methods emphasize updating and reusing experience and rules, but such reuse may not readily translate into changes in overall handling principles. DynEvo combines experience management across different timescales with explicit Strategy Reconstruction, supporting longer-term adaptation while preserving useful experience, and achieves higher Overall Score and Res. Rate in these comparisons. We have added these results to Table 3 and clarified the baseline settings in Appendix C.1 (Baselines) of the revised manuscript.

MethodMetricDaily OpsDouble-11Fraud After.Spring FestivalVIP Anniv.Defective BatchNewcomer SubsidyAdv. StormAvg.
ExpeLOverall0.7470.7340.7320.7200.7230.7040.7380.6460.718
ExpeLRes. Rate0.6440.5940.7960.7720.6940.6710.6830.6650.690
ACEOverall0.7720.8060.7730.7630.7790.7760.7790.7360.773
ACERes. Rate0.7250.8350.8110.8960.8180.8350.8000.7950.814
DynEvoOverall0.8230.8290.8110.8360.8270.8070.8400.7950.821
DynEvoRes. Rate0.9000.9060.9170.9670.9240.9180.9310.9000.920

W2 & Q2: Fairness of Scoring Feedback and Evaluation Metrics.

We thank the reviewer for pointing out this ambiguity. EvoAgentX and the other self-evolving baselines receive the same per-episode Overall Score feedback as DynEvo to guide their respective updates. Thus, DynEvo has no additional advantage in access to this feedback. We have clarified this shared feedback protocol in Appendix C.1 (Baselines) of the revised manuscript.

This design draws on feedback mechanisms in real-world dynamic interactive service environments, where user feedback is commonly used both to evaluate service quality and to guide future service improvement. We also report Res. Rate, which is determined by the User Simulator based on task completion, does not rely on the Judge’s scores, and is not used for evolution updates. DynEvo achieves the highest average Overall Score and Res. Rate across the eight scenarios, demonstrating advantages in both service quality and task completion.


Q1: Model Settings of the Judge Agent and User Simulator.

We thank the reviewer for pointing out this ambiguity. Drawing on the evaluation design of prior agent benchmarks such as τ-bench[3], we use GPT-5.2 with temperature = 1.0 and top-p = 1.0 for both the User Simulator and the Judge Agent. We have explicitly reported these settings in Appendix C.1 (Experimental Setup) of the revised manuscript.


W2 & Q3: Number of Runs and Confidence Intervals.

We thank the reviewer for raising the question of statistical reliability. Following τ-bench[3], each method in our main experiments was evaluated over three independent runs. We report 95% Student’s t confidence intervals across runs for the average Overall Score and Res. Rate, as shown below. We have clarified the number of runs in Appendix C.1 (Experimental Setup) and added the confidence intervals and statistical protocol to Appendix C.3.1 (Robustness) of the revised manuscript.

MetricDirectCoTDSPyEvoAgentAutoAgentsEvoAgentXDynEvo
Overall[0.7382, 0.7632][0.7462, 0.7721][0.7372, 0.7689][0.7627, 0.7891][0.6355, 0.6728][0.7528, 0.7793][0.8118, 0.8300]
Res. Rate[0.7534, 0.7956][0.7580, 0.8092][0.7305, 0.7859][0.8093, 0.8572][0.5657, 0.6336][0.7695, 0.8196][0.8991, 0.9408]

  • [1] ExpeL: LLM Agents Are Experiential Learners. AAAI, 2024.
  • [2] Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models. ICLR, 2026.
  • [3] τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains.

The author's rebuttal has fully addressed my concerns. The advantage over memory-based baselines is solid especially on Res. Rate. However, the authors admitted to use GPT-5.2 with temperature = 1.0 and top-p = 1.0 for both the User Simulator and the Judge Agent implies reasoning_effort is none. I would doubt it may not be an ideal setting for Judge Agent. Therefore, I keep my rating.

We sincerely thank the reviewer for the insightful and constructive feedback! Your comments have helped us improve the paper. During the rebuttal period, we conducted human validation with three trained annotators to further assess the reliability of the User Simulator and Judge Agent under the current settings. The protocol and results have been added to Appendix C.1 (Experimental Setup).

(a) Judge Agent. The annotators independently assessed the trajectories using the same evaluation criteria as the Judge Agent. The Spearman correlation between the averaged human Overall Scores and the Judge Agent’s Overall Scores is 0.814, indicating strong alignment with human judgments.

(b) User Simulator. The annotators reviewed the user profiles and trajectories for persona fidelity, goal consistency, factual consistency, and response and termination behavior. With approval from at least two annotators required to pass, the majority-vote pass rates are 99.29%, 100%, 100%, and 97.86%, respectively, supporting the User Simulator’s consistency with the specified profiles and objectives and the appropriateness of its conversational behavior.

We appreciate your feedback on the Judge Agent’s reasoning configuration and will take this suggestion into consideration in future work. Thank you again for your time and valuable feedback throughout the review process!

Reviewer_nEd8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DynaWorld, a benchmark for evaluating agents across changing customer-service environments, and DynEvo, a multi-agent framework that updates external notes, skills, and strategies without changing model parameters. DynaWorld contains eight scenarios and 1,400 interaction episodes. DynEvo outperforms the evaluated baselines on average Overall Score and Resolution Rate.

Critical strengths
  1. I find the benchmark contribution useful. Linking episodes through changes in policies, user populations, and operational objectives provides a concrete setting for studying adaptation beyond isolated tasks.
  2. The experiments cover several baselines, component ablations, inference costs, and different types of environmental changes. These results give a reasonably broad picture of the system's performance.
Critical weaknesses
  1. The evidence for faster adaptation is less convincing than the evidence for better average performance. Figure 4 reports Drop, Recovery, Loss, and RecoverSteps across event types, while Figure 5 shows a trajectory for DynEvo alone. Neither provides a direct comparison with competing methods or the memory ablations. Reporting these metrics and trajectories for the baselines, especially variants without individual memory components, would help establish whether DynEvo recovers faster after a shift. This seems central to the benchmark's stated purpose.
  2. Some details of the memory updates remain unclear. Appendix B.3 specifies the update schedule and explains how Overall Score enters the tension signal that triggers strategy reconstruction. However, I could not determine exactly how skill success is defined or how credit is assigned when several skills contribute to an episode. The appendix also describes revising matching skills and generating new strategies at a fairly high level. The prompts used for these updates, together with an example showing how a skill or strategy changes across episodes, would make the method easier to reproduce.
  3. Changes in Overall Score need careful interpretation. Appendix C.2 specifies that the scoring weights change with the active objective, and Figure 5 includes the component scores. Even so, an aggregate score can change because the weights change, without a corresponding change in agent behavior. Population shifts also change task difficulty. The event-level analysis would be easier to interpret if it separated these effects from adaptation, for example through a fixed-weight diagnostic alongside the operational score.
  4. I see the main contribution as the benchmark. DynEvo brings together reflection, reusable skills, external memory, and strategy updates in a sensible way, but the paper makes a less convincing case for the novelty of the evolution mechanism itself. A clearer comparison with the closest self-evolving frameworks would help identify what is new beyond their integration.
Questions for authors

see above weakness

4 comments
AuthorsOpenReview ↗

Author Response to Reviewer nEd8 (Part 1/2)

Dear Reviewer nEd8:

We sincerely thank you for the insightful and constructive feedback! We appreciate your recognition of DynaWorld as a useful benchmark for studying adaptation beyond isolated tasks, as well as the breadth of our experimental evaluation. We provide detailed responses below and have incorporated the corresponding changes into the revised manuscript.


W1/Q1: Comparing Dynamic Adaptation with Self-Evolving Baselines and Ablation Variants.

We thank the reviewer for this valuable suggestion. To make differences in dynamic adaptation easier to assess, we have added trajectories for the four self-evolving baselines and variants without Notes, Skills, or Strategy (Figs. 14 and 15), alongside comparisons of Drop, Recovery, Loss, and RecoverSteps across event types (Figs. 16 and 17) and accompanying analyses, to Appendix C.3.3.

The trajectories and ablation results illustrate the components’ roles. Without Notes, performance declines more sharply around some environmental changes, such as Frustrated Wave in Fig. 15, suggesting that reflections on recent cases may preserve useful experience for subsequent interactions. Without Skills, Res. Rate decreases across all eight scenarios, supporting the role of reusable skills in successful task execution. Without Strategy, performance declines more sharply and remains depressed longer around policy changes such as Anti-Fraud On. The event-level analysis further shows that this variant takes longer to recover from population shifts than the variants without Notes or Skills, supporting the role of strategy updates in adapting to changing user populations.


W2/Q2: Details and Examples of Memory Updates.

We thank the reviewer for this valuable suggestion. In Appendices B.2 and B.3, we have clarified the definition of skill success and the accounting rules for episodes involving multiple skills. An application succeeds when the skill is actually used, its key steps are correctly executed, and tool outputs or user feedback in the trajectory support completion of its predefined local objective. Each skill’s usage and success are recorded separately, with at most one usage count per episode and one success count only if all its applications in that episode succeed. To make the method easier to understand and reproduce, we have provided prompt templates for Notes Reflection, Skills Management, Strategy Review, and Strategy Reconstruction, together with concrete skill and strategy update examples, in Appendix B.4.

AuthorsOpenReview ↗

Author Response to Reviewer nEd8 (Part 2/2)

W3/Q3: Analysis of Scoring Weights.

We thank the reviewer for this valuable suggestion. To distinguish changes in component performance from the effects of switching scoring weights, we recompute the composite score from the same logged components using fixed, equal weights [1/3, 1/3, 1/3]. As shown in the table, the ranking by average score across the eight scenarios remains unchanged under fixed equal weights, indicating that DynEvo’s average advantage is not solely attributable to weight switching.

The two scoring schemes offer complementary perspectives: equal weighting provides a common comparison scale, while dynamic scoring reflects the priorities of the current operational objective. For example, during the Cost-First phase of Double-11, which prioritizes cost efficiency, DynEvo’s advantage over EvoAgent increases from 3.56 pp under equal weights to 4.08 pp under the corresponding operational weights. We have added these results and analyses to Appendix C.3.4 of the revised manuscript.

MethodDaily OpsDouble-11Fraud After.Spring FestivalVIP Anniv.Defective BatchNewcomer SubsidyAdv StormFix Avg.Original Avg.
DSPy0.76110.77060.80350.75310.80590.81150.76520.76580.77960.7530
EvoAgent0.75610.78270.82220.77690.80670.81800.77980.79310.79190.7759
AutoAgents0.71630.70320.68370.62370.71050.71530.69880.69680.69350.6541
EvoAgentX0.76930.76680.80060.79880.80910.80240.77350.78920.78870.7661
DynEvo0.81630.82690.81770.83640.82580.81470.83890.80450.82260.8209

W4/Q4: Contributions and Comparison with Self-Evolving Frameworks.

We thank the reviewer for recognizing the value of DynaWorld for studying adaptation beyond isolated tasks. DynaWorld provides a unified, controllable, and reproducible interactive environment for studying online adaptation to cross-episode changes. Within this setting, DynEvo’s contribution centers on experience management and strategy updates across different adaptation timescales. Compared with approaches primarily optimizing prompts, memories, or workflows, DynEvo separately manages recent experience, reusable procedures, and longer-horizon guidelines for dynamic tasks: Notes preserve recent case experience through episode-level reflections; Skills distill reusable procedures from past interactions to support task execution in dynamic environments; Strategy adapts longer-horizon behavior through periodic review and performance-triggered reconstruction, retaining past strategies for reuse when similar conditions recur. In Appendix C.1 (Baselines), we have clarified differences in update targets, timing, triggers, and experience reuse, making DynEvo’s contribution and connection to the requirements of dynamic environments more explicit.

AuthorsOpenReview ↗

A Gentle Reminder

Dear Reviewer nEd8,

Thank you again for your careful review and for recognizing DynaWorld’s value for studying adaptation beyond isolated tasks, as well as the breadth of our experimental evaluation. Your comments have helped us improve the paper. When your time permits, could you please review our response and let us know whether your concerns have been addressed?

During the rebuttal period, we made every effort to address your concerns as follows:

  • Dynamic adaptation comparisons: We added trajectories for four self-evolving baselines and three DynEvo variants without Notes, Skills, or Strategy, together with event-level comparisons of Drop, Recovery, Loss, and RecoverSteps. The analyses clarify the components’ complementary roles in preserving recent experience, supporting task execution, and adapting longer-term strategies (Appendix C.3.3, Figs. 14–17).

  • Memory-update details and examples: We clarified skill-success criteria and how usage and success are recorded when multiple skills are applied within an episode (Appendices B.2–B.3). We also provided prompt templates for the memory-update processes and concrete skill and strategy update examples (Appendix B.4).

  • Scoring-weight analysis: We recomputed scores using fixed equal weights. The ranking by average score across the eight scenarios remains unchanged, indicating that DynEvo’s average advantage is not solely attributable to weight switching. We also explained the complementary roles of equal weighting and operational weighting (Appendix C.3.4).

  • Contributions and framework comparison: We clarified how Notes, Skills, and Strategy support adaptation across different timescales, and explained the differences in update targets, timing, triggers, and experience reuse relative to existing frameworks (Appendix C.1).

We sincerely appreciate your valuable feedback. We hope these additions and clarifications address your concerns and would be grateful if you could reconsider your recommendation in light of our response and the revised manuscript. If any concern remains, we would be happy to provide further clarification.

Warm regards, Authors

AuthorsOpenReview ↗

Gentle Follow-up as the Discussion Period Ends Soon

Dear Reviewer nEd8,

Thank you again for your time and thoughtful feedback!

As the discussion period is drawing to a close, we would like to kindly ask whether you have had a chance to review our responses and additional results.

We hope these additions have addressed your concerns and would be grateful if you could reconsider your evaluation in light of them. If any concerns remain, we would be happy to provide further clarification before the discussion ends.

Thank you again for your time and consideration.

Warm regards, Authors

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

The paper introduces DynaWorld, a customer-service environment in which successive episodes are linked through scheduled environmental events. Agents interact with an LLM user simulator and are scored by an LLM judge on satisfaction and compliance. These scores are combined with an efficiency term into an Overall Score whose weights depend on the active objective. The paper further proposes DynEvo, a two-agent system made up of a service agent and a multimodal evidence agent. With GPT-5.2 as the backbone, DynEvo outperforms prompting baselines and several self-evolving frameworks on every scenario while using fewer dialogue turns.

Critical strengths
  1. Treating environmental change across episodes as an explicit part of evaluation addresses a real gap between static agent benchmarks and deployment. The analysis of degradation and recovery by shift type is something static benchmarks cannot provide.
  2. DynEvo is easy to deploy on closed models. The paper reports time, calls, tokens, and turns per episode alongside quality, showing that DynEvo resolves issues in fewer turns at moderate overhead.
Critical weaknesses
  1. DynEvo uses the episode-level Overall Score as its evolution signal, and that score is the evaluation metric itself. The paper does not specify which feedback signal each baseline receives. It also does not state that the agent is told the active objective. If it is not, a method that observes this score has a structural advantage on exactly the metric being reported.
  2. Most self-evolving baselines are workflow-generation or prompt-optimization frameworks that were not designed for online cross-episode adaptation. In the case studies, DSPy and EvoAgentX respond to entirely unrelated tasks, and AutoAgents performs below direct prompting. This suggests integration problems rather than a lack of adaptation, and casts doubt on the baseline numbers.
  3. Every ablated variant still outperforms all baselines, and removing the multimodal evidence agent hurts the most. So it is unclear how much of the gain comes from self-evolution rather than from within-episode design and prompting.
Questions for authors
  1. Do the baselines receive the same per-episode Overall Score feedback as DynEvo? How do memory-based online learners such as ExpeL, ACE, or Mem0 perform under identical feedback?
  2. Can you add a no-evolution variant of DynEvo, results with other backbones, and repeated runs? Can you also say which models serve as the simulator and judge, and provide some human validation of both?
5 comments
AuthorsOpenReview ↗

Author Response to Reviewer xXmt (Part 1/3)

Dear Reviewer xXmt:

We sincerely thank you for recognizing that DynaWorld addresses a real gap between static agent benchmarks and deployment, and that DynEvo is easy to deploy on closed models while providing quality and efficiency measurements. We appreciate the constructive suggestions and address the concerns below.


W1 & Q1-1: Clarification of Scoring Feedback and Evaluation Fairness.

We thank the reviewer for pointing out this ambiguity in the feedback protocol. All self-evolving baselines receive the same per-episode Overall Score feedback as DynEvo to guide their respective updates. Thus, DynEvo has no additional advantage in access to this feedback. Furthermore, the active operational objective and future event schedules are not provided to any method; all methods must infer environmental changes from public interaction context and feedback and adapt their behavior accordingly. We have clarified this shared setting in Appendix C.1 of the revised manuscript.

In dynamic interactive service environments, user feedback is commonly used both to evaluate interaction quality and to improve future interactions. Following this paradigm, DynaWorld uses Overall Score as feedback for online updates. We also report Res. Rate, which is determined by the User Simulator based on task completion, does not rely on the Judge Agent’s scores, and is not used for evolution updates (lines 1305–1308). DynEvo achieves the highest average Overall Score and Res. Rate across the eight scenarios, demonstrating advantages in both service quality and task completion.


W2: Baseline Implementations and Failure Cases.

We thank the reviewer for raising concerns about the baselines. All self-evolving baselines are implemented based on their official code and receive the same task information and episode-level feedback. We have clarified these implementation settings in Appendix C.1 (Baselines) and expanded the corresponding failure-case analysis in Appendix D of the revised manuscript.

These failure cases highlight the risk of context confusion and inappropriate experience reuse during long-horizon, multi-user interactions in dynamic environments. For example, in the reported failure cases, some baselines apply a previous return-handling procedure to the current user’s new order inquiry and even conflate user identities across episodes. Historical experience supports procedural reuse, but its applicability depends on the current user, task, and rules. As these conditions change, inappropriate reuse may interfere with current task execution.


Q1-2: Comparison with Memory-Based Online Learners.

We thank the reviewer for this valuable suggestion. To provide a more direct comparison with memory-based online learners, we additionally evaluate ExpeL (AAAI 2024)[1] and ACE (ICLR 2026)[2], with results shown below. Dynamic environments require agents to accumulate useful experience and adjust their handling principles as conditions change. Memory-based methods emphasize updating and reusing experience and rules, but this reuse may not readily translate into longer-term strategy adjustment. DynEvo combines experience management across different timescales with explicit Strategy Reconstruction, supporting such adjustment while retaining useful experience, and achieves higher Overall Score and Res. Rate in these comparisons. We have added these results to Table 3 and clarified the baseline settings in Appendix C.1 of the revised manuscript.

MethodMetricDaily OpsDouble-11Fraud After.Spring FestivalVIP Anniv.Defective BatchNewcomer SubsidyAdv. StormAvg.
ExpeLOverall0.7470.7340.7320.7200.7230.7040.7380.6460.718
ExpeLRes. Rate0.6440.5940.7960.7720.6940.6710.6830.6650.690
ACEOverall0.7720.8060.7730.7630.7790.7760.7790.7360.773
ACERes. Rate0.7250.8350.8110.8960.8180.8350.8000.7950.814
DynEvoOverall0.8230.8290.8110.8360.8270.8070.8400.7950.821
DynEvoRes. Rate0.9000.9060.9170.9670.9240.9180.9310.9000.920

  • [1] ExpeL: LLM Agents Are Experiential Learners. AAAI, 2024.
  • [2] Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models. ICLR, 2026.
AuthorsOpenReview ↗

Author Response to Reviewer xXmt (Part 2/3)

W3 & Q2-1: Contribution of Online Self-Evolution, Backbone Generalization, and Stability across Runs.

We thank the reviewer for these valuable suggestions. We have added a no-evolution control and experiments with another backbone, clarified the number of runs, and reported confidence intervals in the revised manuscript.

(a) Ablation without online self-evolution. To directly assess the gains from online self-evolution, we additionally evaluate DynEvo w/o Online Self-Evolution, a variant that uses the same initial settings as full DynEvo but disables cross-episode updates to Notes, Skills, and Strategy. As shown below, disabling all cross-episode updates leads to a larger decline in Res. Rate than any individual component ablation, indicating that online self-evolution provides additional gains under the same within-episode architecture and prompting. We have added these results to Table 4 and expanded the corresponding analysis in Section 5.3.1 of the revised manuscript.

MethodDaily OpsDouble-11Fraud After.Spring FestivalVIP Anniv.Defective BatchNewcomer SubsidyAdv. StormAvg.
DynEvo0.9000.9060.9170.9670.9240.9180.9310.9000.920
w/o Notes0.8810.8630.8830.9260.8940.9020.9060.8500.888
w/o Skills0.8940.8740.8610.9370.9180.9130.9250.8900.902
w/o Strategy0.8630.8960.9060.9590.9180.9040.9060.8700.903
w/o Online Self-Evolution0.8420.8410.8390.9060.8730.8820.8850.8300.862
w/o MM Sub-Agent0.8500.8510.8510.9090.8820.8870.8810.8350.868
Single-Agent w/ MM Reasoning0.8750.8780.8840.9380.9030.9020.9060.8650.894

(b) Backbone generalization. As a parameter-free inference-time self-evolution framework, DynEvo is not tied to a specific LLM backbone. To further assess its backbone generalization, we additionally implement DynEvo on Gemini-3-Flash. Given its API requests-per-minute constraints and the limited rebuttal period, we report results on four randomly sampled scenarios (670 instances) and compare DynEvo with a strong self-evolving baseline. As shown below, DynEvo achieves a higher average Res. Rate, suggesting that its performance advantage extends beyond GPT-5.2. We will include the complete results in the appendix of the revised manuscript once experiments across all scenarios are completed.

MethodDaily OpsSpring FestivalDefective BatchNewcomer SubsidyAvg.
EvoAgent0.8880.8970.9470.8520.896
DynEvo0.9000.9060.9240.8880.905

(c) Stability across runs. We thank the reviewer for raising the question of statistical reliability. Following τ-bench[3], each method in our main experiments was evaluated over three independent runs. We report 95% Student’s t confidence intervals for the average Overall Score and Res. Rate, as shown below. The broadly consistent results across repeated runs support the repeatability of agent evaluation in DynaWorld, while DynEvo maintains its lead in both metrics across all three runs. We have clarified the number of runs in Appendix C.1 (Experimental Setup) and added the confidence intervals and statistical protocol to Appendix C.3.1 (Robustness) of the revised manuscript.

MetricDirectCoTDSPyEvoAgentAutoAgentsEvoAgentXDynEvo
Overall[0.7382, 0.7632][0.7462, 0.7721][0.7372, 0.7689][0.7627, 0.7891][0.6355, 0.6728][0.7528, 0.7793][0.8118, 0.8300]
Res. Rate[0.7534, 0.7956][0.7580, 0.8092][0.7305, 0.7859][0.8093, 0.8572][0.5657, 0.6336][0.7695, 0.8196][0.8991, 0.9408]
AuthorsOpenReview ↗

Author Response to Reviewer xXmt (Part 3/3)

Q2-2: Model Settings and Human Validation of the User Simulator and Judge Agent.

We thank the reviewer for pointing out this ambiguity in the model settings. Both the User Simulator and the Judge Agent in DynaWorld use GPT-5.2, which we have explicitly reported in Appendix C.1 (Experimental Setup) of the revised manuscript. To further assess the reliability of automated evaluation and user simulation, we recruited three annotators with at least a bachelor’s degree and relevant research experience. After standardized training, they independently evaluated a randomly sampled 10% of the interaction trajectories to validate both the Judge Agent and the User Simulator.

(a) Human validation of the Judge Agent. The three annotators assessed the interaction trajectories using the same evaluation criteria as the Judge Agent. We computed human Overall Scores from the mean valid ratings using the same scoring rules and compared them with the Judge Agent’s Overall Scores. The Spearman correlation is 0.814, indicating that the overall scoring trends align well with human judgments and supporting the reliability of the Judge Agent’s scoring.

(b) Human validation of the User Simulator. For the same simulated user instances, the three annotators reviewed the user profiles and interaction trajectories along four dimensions: persona fidelity, goal consistency, factual consistency, and response and termination behavior. Each dimension was assessed by majority vote, with approval from at least two annotators required to pass. The pass rates are 99.29%, 100%, 100%, and 97.86%, respectively, indicating that the User Simulator generally follows the specified personas, task objectives, and known facts while producing appropriate dialogue and termination behavior. Both validation results have been added to Appendix C.1 (Experimental Setup) of the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all reviewers for their thoughtful reviews, insightful comments, constructive suggestions, and the time and effort devoted to evaluating our work. We are encouraged that the reviewers found that DynaWorld addresses a real gap between static agent benchmarks and deployment (Reviewer xXmt), considered the benchmark contribution useful as it provides a concrete setting for studying adaptation beyond isolated tasks (Reviewer nEd8), and recognized it as the first interactive service benchmark where the environment itself changes across episodes (Reviewer CWWJ). The reviewers also noted that our analysis of performance degradation and recovery yields findings that static benchmarks cannot provide (Reviewers xXmt and CWWJ). For DynEvo, we appreciate the recognition that it is easy to deploy on closed models and resolves issues in fewer turns (Reviewer xXmt), and that our experiments give a reasonably broad picture of its performance (Reviewer nEd8).

Based on the reviewers' valuable feedback, we have conducted additional experiments and analyses, and further improved the manuscript. Notably, Items 1–3 address the concerns shared by Reviewers xXmt and CWWJ, and we are glad that Reviewer CWWJ found that our rebuttal fully addressed them. The key revisions and additional results are summarized below:

  1. Evaluation protocol (xXmt, CWWJ): All baselines are implemented from their official code and receive the same task information and per-episode Overall Score feedback as DynEvo. Res. Rate, which is independent of the Judge Agent and not used for evolution updates, further confirms DynEvo's advantage.
  2. Comparison with more memory-based methods (xXmt, CWWJ): DynEvo outperforms the newly added ExpeL and ACE under identical settings on both Overall Score and Res. Rate.
  3. Statistical robustness and evaluation reliability (xXmt, CWWJ): Clarified model configurations, reported 95% confidence intervals over three runs showing that DynEvo's advantage is robust to run-to-run variance, and conducted human validation supporting the reliability of the Judge Agent (Spearman 0.814) and the User Simulator.
  4. Gains from self-evolution and generalization (xXmt): Added a no-evolution ablation, which leads to the largest drop among all ablations, and Gemini-3-Flash backbone results, where DynEvo maintains its advantage.
  5. Adaptation evidence and score interpretation (nEd8): Added adaptation comparisons with baselines and memory ablations, and a fixed-weight diagnostic under which the method ranking remains unchanged.
  6. Method details and positioning (nEd8): Provided memory-update prompts and examples, and clarified differences from related self-evolving frameworks in addressing the challenges of dynamic environments.

Beyond this submission, we hope DynaWorld can serve as a reproducible testbed for the community to study how agents adapt as their deployment environments evolve, moving agent evaluation beyond static and isolated tasks toward continual, real-world service settings. We truly appreciate the reviewers' time and thoughtful engagement, and we warmly welcome any further thoughts or questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.