acceptodds
Under review as a conference paper at ICLR 2027

SkillNEO: Non-Elitist Evolutionary Optimization of Agent Skills

Abstract

Agent skills enable LLM agents to extend their capabilities without updating model weights. Skill evolution builds on this flexibility through continual skill refinement. Current methods use execution feedback to propose revisions, then select versions for further development based on current performance. However, a partially repaired skill may remain low-scoring until later edits resolve the remaining failures. To address this limitation, we introduce SkillNEO (Non-Elitist Evolutionary Optimization), which develops intermediate versions using comparative execution evidence. Its ordinary update selects parents from the current generation’s evaluated children without requiring parent-relative gains or a place among the best historical versions. Comparisons of repeated, parent-child, and sibling executions identify unresolved failures and useful procedures. Meanwhile, archived versions can be revisited with later evidence, and their traces remain available for subsequent edits. We evaluate task-conditioned skill development on 86 SkillsBench tasks across eight domains, using GPT-5.4 Mini in every model-based development role. Selected bundles raise Mini’s average reward from 49.1% to 66.4%, within GPT-5.5 and GPT-5.6 Luna’s original-skill performance range. On the 41 modified tasks, frozen-skill reuse raises Luna’s average reward by 10.5 percentage points; GPT-5.5 exceeds its original-skill reference by 10.6 points. Recorded searches show lower-scoring versions leading to later deployed bundles and supplying procedures to other branches. A weaker executor can thus develop skills reusable by stronger executors on the same tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Agent Skills, Skill Evolution

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_g1Y9OpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SkillNEO, a framework for automatically improving LLM agent skills through non-elitist evolutionary optimization. The paper addresses the problem that conventional performance-driven skill refinement may prematurely discard intermediate skill versions whose current performance is poor but which contain useful partial improvements.

The proposed framework combines offspring-only parent selection, comparative analysis of execution trajectories, archive revisiting, and cross-branch procedure reuse. Unlike approaches that require revisions to improve upon their parents or enter a historical top-performing set, SkillNEO allows lower-scoring intermediate versions to remain eligible for further development. An LLM-based proposer and editor use execution feedback to revise skill instructions, scripts, and supporting resources.

The authors evaluate the framework on 86 SkillsBench tasks across eight domains, using GPT-5.4 Mini for task execution, proposal generation, and skill editing. The reported average reward improves from 49.1% to 66.4%. The authors further demonstrate that the evolved skills can benefit stronger models, including GPT-5.5 and GPT-5.6 Luna. Additional analyses of recorded search trajectories illustrate how lower-performing intermediate versions can lead to successful descendants.

Critical strengths
  1. An important and well-motivated problem.

The paper addresses a meaningful limitation of performance-driven agent skill optimization: current task reward may not adequately reflect the potential of an intermediate skill version for future improvement. The distinction between immediate performance and subsequent development potential is relevant to iterative agent optimization, particularly when improvements require multiple coordinated modifications.

  1. Encouraging empirical results and detailed analysis of evolutionary trajectories.

The reported improvements across 86 tasks are substantial, and the cross-model evaluation provides useful evidence that skills developed by a weaker model can benefit stronger executors. The paper also presents concrete examples of partial repairs, cross-branch procedure reuse, and archived-version revisiting. These analyses provide useful insights into the behavior of iterative skill optimization beyond aggregate performance metrics.

Critical weaknesses
  1. Limited methodological novelty and insufficient experimental evidence supporting the core contribution.

The main methodological contribution relies on offspring-only non-elitist selection, which is an established strategy in evolutionary optimization. Other components, including trajectory-based feedback, archive revisiting, and cross-branch reuse, are also closely related to mechanisms explored in prior skill optimization and evolutionary search frameworks.

Although their combination constitutes a particular system design, the paper does not sufficiently establish what new algorithmic capability this combination provides beyond existing approaches.

More importantly, the paper lacks controlled comparisons against alternative selection strategies under matched computational budgets. In particular, comparisons with elitist parent-offspring selection, historical top-k selection, and simpler iterative refinement approaches are necessary to establish whether the reported improvements are attributable to non-elitist selection rather than repeated LLM-based editing and evaluation.

The paper also lacks component-level ablations that isolate the effects of comparative execution evidence, archive revisiting, and cross-branch procedure reuse. The trajectory analyses demonstrate that lower-scoring candidates can produce successful descendants, but they do not establish that the proposed selection strategy is more effective or efficient than competing approaches.

This substantially weakens the central methodological claim and makes it difficult to assess the actual contribution of SkillNEO beyond its overall engineering implementation.

  1. Limited generalization evidence and potentially instance-specific optimization.

A significant limitation is that skills are independently optimized for individual benchmark tasks and evaluated on the same task instances used during development.

Although the paper conducts fresh executions after skill evolution and introduces safeguards against direct leakage of private evaluator answers, these procedures do not establish that the resulting skills generalize to unseen task inputs or variations.

Repeated optimization using feedback from a fixed task instance may encourage highly specialized instructions or procedures. Consequently, the reported performance improvements could partly reflect adaptation to specific benchmark instances rather than the acquisition of broadly reusable skills.

This distinction is particularly important because the paper emphasizes skill evolution and cross-model transfer. While cross-model reuse is demonstrated on the same tasks, the experiments do not establish whether the evolved skills remain beneficial under different task inputs.

The absence of held-out evaluation limits the strength of the paper's conclusions about reusable skill improvement.

Questions for authors
  1. How can the authors isolate the effectiveness of non-elitist selection from the benefits of repeated skill editing and evaluation?

Could the authors provide matched-budget comparisons with conventional elitist selection, historical top-k selection, and simpler iterative skill refinement? Ideally, these comparisons should use identical developer models, initial skill bundles, task execution budgets, and LLM-call budgets. Ablations of archive revisiting and comparative trajectory feedback would also help clarify which components are responsible for the observed gains.

This evidence would materially affect my assessment of the paper's methodological contribution.

  1. To what extent do the evolved skills generalize beyond the specific task instances used during development?

Could the authors evaluate the evolved skills on held-out inputs or task variations without further skill modification? Such an experiment would help distinguish reusable procedural improvement from instance-specific optimization. It would also clarify whether the observed cross-model benefits persist beyond the development instances.

This is important for assessing the significance of the claimed skill evolution and transferability.

Reviewer_T8niOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SkillNEO, a framework for improving agent skills through non-elitist evolutionary search. It allows lower-scoring versions to undergo further revision and uses comparisons of execution traces to guide edits and reuse procedures from other branches. Experiments on 86 SkillsBench tasks show that GPT-5.4 Mini can improve its supplied skills and that the resulting skills also benefit stronger models on the same tasks. The authors further analyze cases where unsuccessful intermediate versions contribute to later improvements.

Critical strengths
  1. The topic is important, and the motivation for continuing to improve partially repaired skills is reasonable.
  2. The experiments cover several task domains, and the execution analyses provide useful examples of how skill revisions change agent behavior.
Critical weaknesses
  1. The paper lacks comparisons with alternative search strategies under the same budget. The gains could mainly come from repeated execution feedback and skill editing rather than non-elitist selection. It would be helpful to compare with single-path refinement and selection of the best historical versions, and to ablate archive revisiting and cross-branch reuse. The examples of successful descendants from lower-scoring parents do not establish which strategy works better overall.

  2. Each task's skills are evolved separately through repeated feedback on the same task and input, then evaluated on that same instance. This supports task-specific adaptation, but does not show that the resulting skills are broadly reusable, and the evaluation is limited to SkillsBench.

  3. The full skill evolution process is tested only with GPT-5.4 Mini. The other models are from the same GPT family and only execute the resulting skills, so these experiments do not show whether the method works with different LLMs. In particular, weaker models may struggle to diagnose failures and make useful edits. Experiments with models of different capabilities and families would help clarify how broadly the method applies.

Questions for authors

What is the actual cost of evolving skills per task, compared with directly completing the task using the original skills? If evolution is much more expensive than task execution, when would the gains justify that cost in practice?

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

This paper proposed an interesting idea: the intermediate update without relying only on long-term rollout. It can also make use of the potential abilities of the historical parent version. It seems very useful and can improve performance. The contribution is a bit incremental since there are already works on parent-child comparison, but the experimental results show good performance.

Critical strengths

This paper proposes an interesting idea: an intermediate update that doesn't rely just on long-term rollout. It also leverages the potential capabilities of the historical parent version. It seems very useful and can improve performance.

Critical weaknesses

The contribution is a bit incremental since there are already some works on parent-child comparison, but the experimental results show good performance. And the efficiency of the method is unclear.

Questions for authors
  1. How can the method guarantee the edits are going in the right evolutionary direction? I think it's a bit like an exploration strategy, but because this work doesn’t require competition and improvement with parents, that means the exploration space would be very huge; from this perspective, it's very hard to get a signal to support frequent comparisons. Should the selection of the child version have some prediction results in advance?
  2. This method designs comparisons on repeated execution on the same version, parents-child,....., which is very comprehensive, but I’m curious: would it be very inefficient? I notice there is a limitation on parents, is it the number of parents that should be stored in memory or the number of updates?
  3. How to design the rewards? Does it only have global rewards?
  4. As the method needs to store and repeatedly compare different versions, how large are the cache requirements? would it be large memory cost?
  5. The experiments part uses only GPT; as the problem I raised above, the comparison experiments might cost a lot. It might be good to show more base model results because it will give a good evaluation of the method's performance stability, and will provide a cost and efficiency evaluation on different methods.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers and Area Chairs (1/2): What Reviewers Value and Two Shared Concerns

We thank all the reviewers for their careful and constructive feedback. SkillNEO starts from one observation: a skill version's current reward can miss its development value, because a partially repaired skill can score low until later edits finish the repair. SkillNEO keeps such a version editable through non-elitist parent selection, and uses execution comparisons to guide the next edit. We are encouraged that the reviewers value the problem, the idea, the analysis and the results.

What the reviewers valueIn their words
A meaningful distinction between immediate performance and future development potential"The paper addresses a meaningful limitation of performance-driven agent skill optimization: current task reward may not adequately reflect the potential of an intermediate skill version for future improvement." (g1Y9) "The distinction between immediate performance and subsequent development potential is relevant to iterative agent optimization, particularly when improvements require multiple coordinated modifications." (g1Y9) "The topic is important, and the motivation for continuing to improve partially repaired skills is reasonable." (T8ni)
Insight into how skills evolve and change agent behavior, beyond aggregate scores"These analyses provide useful insights into the behavior of iterative skill optimization beyond aggregate performance metrics." (g1Y9) "The paper also presents concrete examples of partial repairs, cross-branch procedure reuse, and archived-version revisiting." (g1Y9) "the execution analyses provide useful examples of how skill revisions change agent behavior." (T8ni)
Evidence that skills developed by a weaker model benefit stronger executors"the cross-model evaluation provides useful evidence that skills developed by a weaker model can benefit stronger executors." (g1Y9)
An interesting idea that leverages the potential of historical versions"This paper proposes an interesting idea: an intermediate update that doesn't rely just on long-term rollout. It also leverages the potential capabilities of the historical parent version." (ogPK)
Substantial improvements across 86 tasks in several domains"The reported improvements across 86 tasks are substantial" (g1Y9) "The experiments cover several task domains" (T8ni) "It seems very useful and can improve performance." (ogPK)

Two shared concerns, both answered with new experiments

Our individual responses address every point raised and contain the full result tables. The reviews share two concerns:

ConcernRaised byWhat we did and foundDetails in
Do the gains come from non-elitist selection or from repeated feedback and editing?T8ni and g1Y9 asked for a comparison with alternative selection rules under the same budget, with component ablations. ogPK asked about the cost of different methods.Given the same budget and with everything else held fixed, SkillNEO achieves the best mean of the three parent rules on 8 tasks: 77.0, against 64.8 for historical top-3 and 59.4 for single-path, a lead of 12.2 and 17.6 points. Removing archive revisiting or comparison & grafting lowers the mean by 11.3 and 14.3 points on 5 of these tasks.T8ni W1, g1Y9 W1/Q1, ogPK W1/Q5
Do the gains hold beyond the development instance?T8ni and g1Y9 noted that each skill is evaluated on its development instance. g1Y9 asked for held-out inputs.On unseen task variants and without any modification, the evolved skills raise GPT-5.4 Mini from 49.3% to 82.1% over 8 tasks and GPT-5.6 Luna from 37.8% to 84.4% over 5 tasks.T8ni W2, g1Y9 W2/Q2, ogPK Q5
AuthorsOpenReview ↗

General Response to All Reviewers and Area Chairs (2/2): Other Questions and Key Takeaways

Other questions

QuestionRaised byOur answerDetails in
What does the design add beyond existing approaches?g1Y9 noted that offspring-only selection is an established strategy, and ogPK that earlier work also compares parents and children.SkillNEO's components are designed together around one idea: a version's current reward can miss its development value. Existing approaches mainly admit revisions that beat their parent or a historical top-k set. SkillNEO keeps a low-scoring version editable and uses comparisons to direct the next edit, not to admit children. It draws on offspring-only selection, which to our knowledge has not been used in skill evolution, and builds the rest of the method around it.g1Y9 W1, ogPK W1
Does the method depend on the developer model?T8ni asked about weaker developers and other model families, and ogPK about other base models.GPT-5.4 Mini, the weakest model in the study, develops the skills from its own executions, and the same skills help stronger executors. SkillNEO is not tied to a model family: its model-dependent steps are prompted LLM calls for diagnosis, proposal and editing.T8ni W3, ogPK Q5
Is the method efficient?T8ni and ogPK asked about cost, efficiency and memory.Evolving one task's skills is a one-time cost of about $20, after which reward per execution on the modified tasks rises from 31.5 to 67.8. Comparisons add no task executions, and disk and memory use are small.T8ni Q1, ogPK W2/Q2/Q4
How are edits directed and rewarded?ogPK asked about edit direction and reward design.The search is small and evidence-driven, and every deployment is checked, so none of Mini's tasks ends up worse. SkillNEO uses each task's verifier reward and takes its finer-grained signal from execution evidence.ogPK Q1/Q3

Key takeaways from the rebuttal

  • The selection rule matters. With the same budget, feedback and editing, changing only the parent rule lowers the mean by 12 to 18 points on the 8 tasks compared.
  • The improvement is reusable. Without any modification, the evolved skills raise GPT-5.4 Mini from 49.3% to 82.1% on unseen variants of 8 tasks, and GPT-5.6 Luna, which only executes them, from 37.8% to 84.4% on 5 tasks.

Taken together, the two experiments point back to the question behind SkillNEO. Improving a skill often takes several edits, and the reward along the way does not show whether a version is a dead end or one edit from working. SkillNEO offers a concrete way to keep developing such versions, with evidence that doing so pays off. As agents rely more on automatically developed skills, we believe this question will only grow in importance.

We thank the reviewers and Area Chairs for their time and hope this response addresses the shared concerns. We would be glad to answer any further questions.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.