acceptodds
Withdrawn from ICLR 2027

EvoMCP: Co-Evolving MCP Agents and Training Data

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Supervised fine-tuning of Model Context Protocol (MCP) agents typically relies on static training data that does not adapt to the student’s evolving capabilities. We propose EvoMCP, an agent–data mutual evolution framework that evolves complete task–trajectory pairs through a closed loop of task evolution, trajectory evolution, and supervised fine-tuning. EvoMCP uses student rollout feedback to assess established capabilities and capability gaps, guiding the selection of learning targets and compatible verified seed tasks. Task evolution translates these targets into executable tasks by adapting task instructions, observable environment states, or both, while maintaining consistency among instructions, environments, and verifiers. Trajectory evolution uses reusable skills to recover failed teacher rollouts and refines successful demonstrations for efficiency. Accepted demonstrations train the next student, while their associated tasks become seeds for subsequent evolution. Fresh feedback from the updated student then guides the next round, allowing the training distribution to evolve alongside the agent’s capabilities. Experiments on Toolathlon, MCPMark, and MCP-Atlas show consistent improvements across three evolution rounds. Starting from Qwen3.6-27B, EvoMCP achieves absolute gains of 16.5, 10.2, and 12.6 percentage points on the three benchmarks, respectively.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Data Synthesis, LLM Agents, Self-Evolution

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_bC9WOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes EvoMCP, a framework that combines iterative training-data construction with supervised fine-tuning to improve MCP-based tool-use agents. In each round, feedback from the current student’s rollouts guides the selection of learning targets and verified seed tasks. The framework then modifies task instructions, environment states, or both, while maintaining consistency with the corresponding verifiers. A teacher model executes the resulting tasks to produce demonstrations. For each fixed task, model-related execution failures are addressed through reusable recovery skills, while successful trajectories are audited and potentially regenerated to reduce tool-use errors and avoidable work. Selected verifier-passing demonstrations are used to fine-tune the student after removing teacher-only guidance. Accepted tasks serve as seeds for subsequent task generation, and the updated student provides fresh feedback for the next round.

Experiments using Qwen3.6-27B on Toolathlon, MCPMark, and MCP-Atlas report improvements over the initial student and a static SFT baseline across three evolution rounds. The paper also presents component ablations and analyses of task validity, teacher-trajectory quality, and skill-guided recovery to assess the contributions of its data-construction pipeline.

Critical strengths
  1. Practical attention to training-data validity. The pipeline explicitly checks consistency between generated task instructions, environment states, and verifiers. It also distinguishes model-behavior failures from task or infrastructure failures before applying recovery. These are sensible design choices for constructing executable MCP training data.

  2. Controlled evidence for guided recovery. The reported improvement supports the usefulness of recovery guidance for obtaining additional successful demonstrations beyond simply allocating more execution attempts.

Critical weaknesses

1. The main evaluation does not separate training feedback from held-out testing.

Section 5.1 and Appendix E.1 state that the same benchmark tasks are used both to guide the next round of data generation and to evaluate the updated student. My concern is that the model is being tested on tasks that have already shaped its training data. The gains therefore show benchmark-guided adaptation, but do not tell us how well the method generalizes beyond that feedback pool. I would like to see results on a separate held-out test set that is excluded from both feedback and data construction.

2. The baseline comparisons do not sufficiently establish the advantage of the proposed adaptive training strategy.

The public-model scores in Table 1 use different evaluation setups and do not directly compare training methods. In Table 2, the only baseline followed through all three rounds is Static SFT, which keeps reusing the same seed data. Feedback-free generation is only tested at R1, so it remains unclear how much the later gains depend on refreshed student feedback beyond the benefit of new training examples. Matching SFT steps alone also does not ensure comparable data-generation costs. I would like to see the feedback-free control extended across rounds, alongside few existing adaptive data-generation baselines with comparable student, teacher, and compute settings.

3. The incremental novelty over existing adaptive data-generation and trajectory-quality methods is insufficiently established.

Much of the pipeline has close precedents. Confucius [1] uses student uncertainty to guide data generation and iterative SFT, while CoEvolve [2] updates training tasks or data based on model feedback. The paper acknowledges most of these connections, but I am less convinced about what its specific MCP integration adds. The ablations show that the components help within this pipeline, without establishing an advantage over existing alternatives. For example, comparing skill-guided recovery with per-task reflection, and trajectory regeneration with step-quality-filtered SFT, would help clarify the contribution. As it stands, I see the work mainly as an integration of established ideas, with the value of its particular design choices still needing stronger evidence1.

References

[1] Confucius: Iterative Tool Learning from Introspection Feedback by Easy-to-Difficult Curriculum. AAAI 2024.

[2] CoEvolve: Training LLM Agents via Agent-Data Mutual Evolution. ACL 2026.

Questions for authors
  1. Comparison with related methods. Could the authors compare EvoMCP with a representative prior data-construction approach under a comparable MCP training setup? This would help clarify what the proposed pipeline adds. The comparison should use the same student and teacher, with comparable generation and training budgets.

  2. Held-out evaluation. Appendix A.3 and Figure 14 mention held-out environment variants, but I could not find aggregate student results on that split. Could the authors report these results, specify the split size, and clarify whether it was excluded from both feedback generation and training-data construction? If a held-out student evaluation is outside the intended scope, could the authors explain why benchmark-guided evaluation is sufficient to support the paper’s main claims?

Reviewer_hyoXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

EvoMCP uses student rollout feedback to select learning targets, co-evolves MCP tasks, environments, and verifiers, and generates verified SFT trajectories through teacher recovery and refinement. Updated students and accepted tasks feed subsequent rounds. Three-round gains over R0 are 16.5, 10.2, and 12.6 percentage points on Toolathlon, MCPMark, and MCPAtlas. Validation includes step-matched Static SFT, R1 component ablations, target-realization checks, and recovery comparisons with matched teacher-rollout budgets.

Critical strengths
  1. Jointly modifying tasks, environments, and verifiers makes a selected learning target an environment-supported, verifier-checked behavior rather than only a textual instruction. This addresses a concrete mismatch between instructions and executable supervision in synthetic agent data. Table 3's target-realization improvement from 27.8% to 85.7% provides direct evidence for the construction.
  2. Distilling failure experience into reusable recovery skills obtains successful demonstrations from tasks that direct teacher executions fail, rather than discarding those tasks. Under matched teacher-rollout budgets, skill-guided recovery achieves recovered-task shares of 26%/40% at R1/R3 versus 9%/14% for unguided retries. This comparison supports the recovery mechanism's value for acquiring usable data, independently of the headline benchmark score.
Critical weaknesses
  1. Reusing benchmark tasks for feedback and checkpoint evaluation prevents the main results from separating feedback-pool adaptation from improvement on unseen tasks. The protocol is disclosed; the concern is missing evaluation excluded from all feedback, not concealed data use. For a paper centered on agent/data co-evolution, this is an important evidence gap and the main reason for leaning toward rejection.
  2. Table 2 already includes an R1 no-student-feedback ablation matched in demonstration count and SFT settings, supporting feedback's first-round contribution. However, R2/R3 mainly compare against Static SFT on repeated seed data, without extending that no-feedback data-evolution control. The widening advantage therefore does not fully separate continued feedback planning from access to fresh data. Matching accepted demonstrations and training steps also does not match total collection cost, and missing independent training repeats limits interpretation of smaller differences.
Questions for authors
  1. Is there an independent test set excluded throughout from learning-target planning, seed selection, and checkpoint selection?
  2. Can Table 2's R1 no-student-feedback variant be continued through R3? Does the full method retain an advantage with matched teacher and environment collection budgets, including rejected-data cost?
  3. Are independent training repeats and paired task outcomes available for full and key baselines? Is the loss of 34 of 189 initially successful tasks in Appendix E.2 stable across repeated runs?
Reviewer_UdqsOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes EvoMCP, a framework for co-evolving MCP agents and their training data through iterative task evolution, trajectory refinement, and supervised fine-tuning. The framework leverages student feedback to identify capability gaps and generate targeted training tasks, while improving teacher demonstrations through failure recovery and efficiency refinement. Experiments across three MCP benchmarks demonstrate consistent performance improvements.

Critical strengths
  1. The central idea of dynamically adapting training data to the evolving capabilities of student agents is intuitive and well-motivated. The proposed framework provides a coherent approach to improving MCP agents through iterative data generation and model training.
  2. The authors conduct extensive experiments across multiple MCP benchmarks, including comparisons with baselines, ablation studies, and analyses of task and trajectory evolution. The results provide encouraging evidence for the effectiveness of the proposed framework.
Critical weaknesses
  1. Section 4.1 introduces a student assessment profile (Line 209) to identify the student's established capabilities and capability gaps. However, the paper does not adequately explain how this profile is constructed. This is a critical component of the proposed framework, as accurately identifying student weaknesses is the foundation for all subsequent task evolution and training steps. It is possible that the authors employ an LLM to analyze and summarize student trajectories, but the technical implementation of this process remains unclear. More details about the assessment procedure, including the prompts, inputs, and output format, would significantly improve the clarity and reproducibility of the work.
  2. The paper does not provide sufficient information about the origin and construction of the initial verified seed task–trajectory pairs. Since these seed tasks serve as the foundation for subsequent task evolution, understanding how they are collected, constructed, and verified is essential. The authors should clarify whether the seed data are manually created, derived from existing datasets, or automatically synthesized.
  3. The notation throughout the paper could be improved for clarity. For example, Equation (2) introduces Ts\mathcal{T}_s and Em\mathcal{E}_m, but their mathematical definitions and concrete implementations are not sufficiently explained. It is unclear how these operators correspond to the actual task evolution procedures described later in the paper, especially since the notation is not subsequently reused. Furthermore, the meanings of subscripts such as TT and mm are not sufficiently intuitive or consistently clarified. A more systematic notation system, together with explicit connections between the mathematical formulations and implementation, would make the method easier to follow.
Questions for authors
  1. Could the authors provide a detailed description of the capability assessment pipeline? In particular, is the profile generated by an LLM-based analysis of student trajectories, a rule-based procedure, or a combination of both? Providing the corresponding prompts or implementation details would be helpful.
  2. Are these data manually constructed, collected from existing benchmarks or datasets, or automatically synthesized? Could the authors provide more information about the data construction process and the size of the initial seed pool?
  3. The main text refers to a Task Generator, whereas the appendix provides separate prompts for Task Builder and Task Realization. It is unclear whether these components refer to the same underlying model operating with different prompts or to distinct modules within the framework. Clarifying their respective responsibilities and relationships in the main text would help readers better understand the overall pipeline.

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

Related papers

Discussion (0)

Sign in to comment.