acceptodds
Withdrawn from ICLR 2027

EvoGenome: Genotype-Phenotype Evolution for Compositional Agent Skills

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language model agents increasingly use reusable skills that encode procedural knowledge for tool use, artifact manipulation, verification, and failure recovery. Existing skill-evolution methods search over free-form text, files, or complete skill packages. As a result, a single edit can modify unrelated behaviors, which complicates the attribution and recombination of successful changes. We introduce EvoGenome, a genotype–phenotype framework that represents compositional agent skills as typed, functionally defined, and independently heritable units. Its SkillGenome encodes activation triggers, preconditions, planning logic, tool-use policies, executable modules, postconditions, verification procedures, recovery behavior, and adaptation interfaces. A materialization compiler translates each genome into a conventional executable skill package, separating structured search from deployment while maintaining compatibility with existing agent harnesses. Type- and interface-aware operators perform constrained mutation, homologous crossover, module specialization, pruning, and cross-domain transplantation while preserving dependencies among skill components. EvoGenome uses execution traces and lineage differences to attribute behavioral and fitness changes to inherited genome segments, allowing subsequent variation to prioritize genes that provide consistent benefits. We compare EvoGenome with free-form rewriting and file-level crossover on a skill benchmark and a dynamic tool-use environment under fixed execution budgets. We evaluate validity, executability, behavioral innovation, fitness improvement, regression, and cross-domain transfer. Ablation studies assess the effects of type constraints, lineage-based credit, crossover, non-textual genes, and genome granularity.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Skill Evolution, Genotype–Phenotype Framework, LLM Agents

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

The paper proposes EvoGenome, a framework that represents agent skills as typed modules for planning, tool use, execution, verification, and recovery. It combines constrained mutation, crossover, and module transfer with trace- and lineage-based credit assignment to guide skill evolution. The authors report improved task success and executability, with fewer regressions, on SkillsBench and a dynamic tool-use environment.

Critical strengths

The paper studies a practically relevant problem: interference between edits and difficulties in reusing skill modules. Its proposed decomposition of planning, verification, and recovery into independently optimizable modules is a potentially useful design choice.

Critical weaknesses
  1. Missing strong baselines: The paper lacks direct comparisons with GEPA, EvoSkill, and other closely related methods, making its advantages over existing approaches difficult to establish.
  2. Weak credit assignment: Allocating overall gains based on module activation and verifier evidence does not isolate individual contributions when multiple modules are modified jointly.
  3. Questionable search strategy: Prioritizing edits to high-credit modules assumes that past contributions predict further improvement, an assumption the paper does not adequately justify.
  4. Poor writing and presentation: The figures, in particular, look templated and hastily assembled, giving the paper a pronounced “vibe-coded” feel rather than clearly communicating the method.
  5. Inconsistent experimental protocol: Unresolved inconsistencies in data splits, sampling sizes, and budget accounting undermine the credibility and reproducibility of the results.
Questions for authors
  1. Why are GEPA, EvoSkill, and other strong baselines missing? How do you rule out localized editing and additional checks as the primary sources of improvement?
  2. When multiple modules are modified jointly, how does your credit assignment identify harmful edits rather than incorrectly reward them for an overall gain?
  3. Why should high-credit modules be prioritized for further modification rather than preserved? What experimental evidence supports this choice?
  4. If the evolution split contains only 20% of the development tasks, how can you sample 32 tasks without replacement? Please provide the dataset version, split sizes, and task lists.
  5. What are the separate budgets for generated candidates, actual environment executions, and model calls? Why do the failure rates and reductions reported in Table 11 and the main text not align?
Reviewer_Hx8yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for evolving agent skills. It represents skills as typed functional modules covering activation, planning, tool use, executable code, verification, and recovery. It includes the Type- and interface-aware operators perform multiple operations, including constrained mutation, homologous crossover, module specialization, pruning, and cross-domain transplantation. The method also uses execution traces with the skills to guide subsequent edits. Experiments on SkillsBench and a dynamic tool-use environment demonstrate the effectiveness of the proposed method.

Critical strengths
  • The idea of evolving skills through functional modules with explicit interfaces is interesting and reasonable.

  • The studied problem is practical and interesting. This paper is well-motivated.

  • This paper is well written.

Critical weaknesses
  • Most of the experiments use Qwen2.5-Coder-32B-Instruct for both task execution and skill evolution. Qwen2.5-Coder-32B-Instruct is an outdated and small model given the fast pace in this field. Evaluation with additional more recent and larger backbone models would help establish whether the benefits generalize across models (especially recent ones) with different capabilities.

  • The framework contains many components, but the ablation studies do not fully supports their necessity. Although the paper evaluates the contributions of type constraints, lineage credit, crossover, and genome granularity, it does not separately examine the contributions of the eight gene types or other operators such as specialization and pruning. The proposed credit-assignment mechanism is also only compared with removing lineage credit, rather than other alternatives such as uniform credit assignment.

  • The paper does not directly compare with existing skill-evolution methods such as EvoSkill, CoEvoSkills, and SkillEvolver, although these methods are discussed as closely related work. The experimental baselines mainly consist of self-constructed strategies, including free-form rewriting, file-level crossover, and AST-level evolution. The experients are not sufficient for demonstrating the advantage over existing skill evolution methods. Comparisons with representative existing methods would strengthen this paper.

Questions for authors

see above

Reviewer_JkNiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EvoGenome represents an agent skill as eight types of genes and compiles them into executable skill packages. Typed mutation and crossover operate on compatible components, while lineage credit based on paired execution traces guides future edits. A frozen Qwen2.5-Coder-32B backbone is evaluated on SkillsBench and a dynamic command-line environment. Reported results improve task pass rates and candidate validity over rewriting, file-level crossover, and AST-level evolution, with additional component and transfer studies.

Critical strengths
  1. The representation covers triggers, tool policies, executable code, verification, and recovery in one compositional interface. This is a useful way to organize edits that need to preserve dependencies across instructions and executable resources, and the reported validity and regression measurements directly address that motivation.

  2. The empirical study includes five seeds, matched parent/offspring evaluation, several baselines, and paired statistical comparisons in Table 12. Gains over AST-level evolution are reported in both settings: 64.5 versus 62.1 on SkillsBench and 54.5 versus 49.7 in the dynamic environment. The component and transfer analyses provide more evidence than a single aggregate pass rate.

Critical weaknesses
  1. The evaluation population and budget accounting cannot be reconstructed precisely. Section 4.1 and Appendix B describe a 20% development-task evolution split and batches of 32 tasks sampled without replacement, but do not give the benchmark version, task counts/IDs, or exact generated-episode counts. Likewise, 128 candidate evaluations exclude initialization, and parents are reevaluated for credit updates; the relation between this count and total execution/model-call budgets needs an explicit ledger. The supplied code is clearly labeled as a synthetic CPU demonstration rather than the reported experiments, so it cannot resolve these details. This limits assessment of the reported empirical advantage, despite the useful aggregate statistics.

  2. The specific benefit attributed to trace-weighted gene credit is not isolated. Equation 11 allocates one joint fitness difference with nonnegative responsibility weights, so every edited gene receives an increment with the same sign, even when one edit helps and another hurts. This can be a useful search heuristic, but it does not identify separate gene effects. The no-credit ablation changes proposal sampling to uniform, which tests the whole credit mechanism rather than the added value of trace weighting. The acknowledged noise from interacting genes does not resolve this structural ambiguity.

Questions for authors
  1. Please provide task/version manifests, counts for each split and generated configuration, and explain whether the 32-item batch counts distinct tasks or episodes. Can a per-method budget ledger include initialization, rejected candidates, paired parent reevaluations, environment executions, and model calls?

  2. Does trace-weighted allocation outperform allocating the same paired fitness difference equally among edited genes, with the rest of the proposal rule fixed? A small analysis of multi-gene edits with opposing effects would clarify when the proposed credit is informative.

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

Related papers

Discussion (0)

Sign in to comment.