acceptodds
Under review as a conference paper at ICLR 2027

CoEvoSkill: Niche-Aware Co-Evolution of Skill Content and Graph Topology for Agents

Abstract

Skill libraries have emerged as a lightweight adaptation for LLM agents, providing deployable tools without modifying model weights. However, existing methods optimize skill libraries along two isolated lines: refining individual skill content, or evolving graphs that organize skills. Moreover, neither line addresses the execution layer itself—refined skills can still be realized as invalid, repetitive, or premature steps. We argue that skill content, graph topology, and execution are mutually dependent. We present CoEvoSkill, a framework that closes this loop: skill texts are refined by failure attribution from execution, the dependency graph is updated by success-conditioned synergy scores, and a Monte Carlo Tree Search (MCTS) controller searches over graph-supported skills, with a guard that validates every language realization before it changes the state. The two layers exchange evidence: guard rejections and search statistics drive content edits, edge updates, and detection of missing skills, while evolved libraries reshape the candidate space the search sees. We prove a fixed-point theorem showing that the alternating optimization converges under bounded edits and contractive topology changes. Experiments across various domains such as web navigation, mathematical reasoning, financial analysis, and code generation show that CoEvoSkill outperforms state-of-the-art methods, at the same time exhibiting strong generalization capability.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Self-Evolving Agents, Skill Discovery, Skill Library, Library Learning, Monte Carlo Tree Search, Policy Learning

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

CoEvoSkill jointly evolves skill content and graph topology, using niche-conditioned edits, success-based graph updates, and a guarded MCTS controller. The paper presents a convergence and dominance theorem and evaluates the framework on five benchmarks.

Critical strengths

Jointly adapting skill content and its organization is a reasonable direction. The evaluation covers three domains and includes content-only and graph-only ablations.

Critical weaknesses
  1. The central dominance guarantee is not established. Theorem 1 argues that a larger proposal set and monotonic acceptance ensure that joint evolution reaches a stationary point at least as good as either single-layer alternative. This does not follow: different update paths can reach different local optima. Being able to replay a single-layer trajectory does not mean the proposed algorithm will do so. The stated assumptions do not resolve this gap.

  2. Experimental validation is insufficient. Section 4.4 reports evolution and convergence statistics without supporting curves or systematic analysis. The claimed cross-domain transfer lacks a held-out-domain evaluation. Although SFT and GRPO budgets are shared, the paper does not clearly establish matched inference and controller-training budgets across baselines or report total computational cost. These omissions limit what the reported gains establish.

Questions for authors
  1. What additional assumptions justify the dominance claim for the actual update procedure, rather than merely the existence of a favorable trajectory?

  2. Do all baselines use the same inference controller and budget, and what is the total computational cost of each method?

Reviewer_e12YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

CoEvoSkill jointly updates skill text and a graph of skill relationships, using each skill’s structural role to guide revisions. A guarded MCTS controller executes the skills and supplies feedback. Experiments on five benchmarks report improvements over content-only and graph-only alternatives, alongside a theorem claiming convergence and dominance over these alternatives.

Critical strengths

The author proposed that skill edits are checked in the context of their dependencies, rather than evaluated independently. The graph-only and content-only ablations are useful, and the improvements across embodied, mathematical, and financial tasks suggest that coordinating the two layers has practical value.

Critical weaknesses
  • The dominance claim in Theorem 1 is not established. Appendix E argues that the joint search must outperform either single-layer search because it can make all their edits. However, access to those edits does not mean the algorithm follows either successful sequence: an accepted edit can lead to a worse local optimum and make another path inaccessible under the strict improvement gate. Accepting one coupled edit also does not establish strict superiority. The finite-acceptance bound is valid, but it does not prove these stronger claims.

  • The computational fairness of the comparisons is unclear. Table 3 gives CoEvoSkill 12 MCTS simulations per decision, while Table 1 specifies shared SFT and GRPO budgets and one episode per task. It does not establish whether the baselines receive equivalent search, model-call, or token budgets. Without this information and a search-controller ablation, the headline gains cannot be cleanly attributed to skill co-evolution.

Questions for authors

See Weakness.

Reviewer_62YPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces CoEvoSkill, a niche-aware framework for jointly optimizing skill content and graph topology in LLM-based agents. The framework uses structural roles, or niches, to guide localized text edits, while updating graph relations according to feedback from execution trajectories. These two processes are coordinated by a co-evolution scheduler. During execution, CoEvoSkill employs a guarded Monte Carlo tree search controller, together with strict validation gates to filter candidate revisions. Experiments on five benchmarks spanning embodied intelligence, mathematical reasoning, and financial analysis show that CoEvoSkill consistently outperforms content-only and structure-only baselines, with particularly strong improvements on compositional tasks.

Critical strengths
  1. The paper provides a concrete and operational definition of niche. Rather than treating the concept as a loose metaphor, it maps structural roles—such as dependency depth, influence, diversity, and abstraction—into measurable edit budgets and stratified validation scopes, as formalized in Equations 2 and 3. This design makes the niche mechanism directly actionable: local text edits are explicitly constrained by the topology of the skill graph.

  2. The empirical results provide strong motivation for joint evolution. Across five benchmarks, CoEvoSkill consistently outperforms both static skill libraries, such as SkillRL, and content-only evolution methods, such as EvoSkill. The gains are particularly notable on compositional tasks, for example on ALFWorld, where CoEvoSkill achieves 92.5 compared with 88.0 for the strongest baseline. The ablation study in Table 5 is also informative: removing either the micro-level or macro-level component leads to the largest performance degradation, suggesting that both layers are important to the overall framework.

Critical weaknesses
  1. The independent contribution of the joint loop is not fully isolated. The paper shows that using both components together outperforms using either component alone. However, it remains unclear whether the improvement comes from the actual interaction between the micro and macro layers, or simply from additional computation and validation. A joint-but-no-feedback baseline with a matched compute and validation budget would be necessary to determine whether the continuous feedback loop itself is responsible for the reported average score of 81.7, rather than the effect of stacking content edits and graph updates.

  2. The theoretical assumptions are not sufficiently supported by empirical evidence, and the statistical analysis is relatively limited. Theorem 1 relies heavily on the contraction assumption A3 and on the condition of a fixed executor. However, the paper does not report the observed trajectories of niche drift Δt\Delta_t or scheduler probability ftf_t, which makes it difficult to assess whether the assumed contraction actually occurs during training. In addition, the experiments are conducted only with Qwen2.5-7B-Instruct, and the results are reported as means without standard deviations or significance tests. This limits the reader’s ability to evaluate robustness with respect to random variation and model choice.

Questions for authors
  1. Could the authors include a joint-but-no-feedback ablation? In this setting, both content evolution and graph evolution would remain active, but the feedback channel between them would be disabled. For example, graph changes would no longer affect the edit budget bt(v)b_t(v), and content edits would no longer update or refresh graph evidence. This baseline should use the same compute and validation budget as the full CoEvoSkill model, so that the effect of the coupled feedback loop can be isolated more cleanly.

  2. How do the authors justify the contraction assumption A3 in practice? It would be helpful to report the empirical trajectories of niche drift Δt\Delta_t and scheduler probability ftf_t throughout training, together with per-phase accept and reject counts. Moreover, the proposed symbiosis score σ\sigma appears to be correlational rather than causal. A placebo test, such as randomizing co-activation edges while keeping the skill contents fixed, would help clarify whether the observed gains are truly driven by graph-topology changes rather than by additional validation or search.

  3. Could the authors provide stronger statistical and cross-model evidence? In particular, please report per-seed results, standard deviations, and confidence intervals for the main comparisons. Since the current experiments rely entirely on Qwen2.5-7B-Instruct, a small-scale replication using a different backbone or model size would help demonstrate that the method is not specific to the formatting behavior or instruction-following characteristics of this particular checkpoint.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.