acceptodds
Withdrawn from ICLR 2027

Skills and Challenges Co-Evolve: Learnability-Aware Autocurricula for Agent Skill Evolution

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language model agents use reusable skills to solve multi-step tasks that require reasoning, tool use, and executable procedures. Existing methods typically optimize skills against largely fixed objectives or generate evaluation tasks separately from skill learning. This separation can favor verifier-specific solutions and produce challenges that provide little learning signal or exceed the skill population's capabilities. We introduce SCOPE, Skill and Challenge Optimization through co-Population Evolution, a learnability-aware autocurriculum that jointly evolves candidate skills and executable challenges near the changing capability boundary of the skill population. SCOPE evaluates each challenge by how well it discriminates among skills, provides actionable signals for attainable improvement, and reveals previously uncovered failures. A solvability gate filters invalid or unsupported challenges through executable verification, reference procedures, or independently verified model consensus. The surviving challenges guide skill mutation and selection. Structured mutations vary tools, schemas, information availability, dependencies, and feedback signals. Hall-of-fame replay maintains pressure from previously informative challenges and reduces regressions on acquired capabilities. We compare SCOPE with fixed-task optimization, random domain perturbation, and adversarial challenge generation without learnability control under hidden tool versions, unseen difficulty ranges, and composed perturbations. Controlled analyses examine evolutionary dynamics, post-evolution transfer, failure coverage, and the individual contributions of learnability, solvability validation, replay, and mutation diversity.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Large Language Model Agents, Skill Evolution, Learnability-Aware Autocurriculum

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

SCOPE co-evolves reusable agent skills and executable challenges using discriminativeness, probe-based learning progress, failure novelty, solvability validation, and hall-of-fame replay. Model parameters remain frozen. Experiments with Qwen2.5-72B-Instruct cover four professional domains and several distribution shifts. The reported average pass rate improves from 52.8% with fixed-task optimization to 62.1%, with ablations supporting learnability estimation, validation, replay, and mutation diversity.

Critical strengths
  1. A useful curriculum design principle. Combining measured learning progress with executable solvability checks addresses a concrete limitation of failure-maximizing generation: difficult challenges need not provide productive training feedback. The learnability and validation ablations support this distinction.

  2. Meaningful reported generalization gains. SCOPE improves over VALIDATED-ADV by 3.1 percentage points on average and 8.1 points on compositional shifts (Table 1). Freezing skills during evaluation makes these results relevant to persistent procedural improvement rather than continued adaptation at deployment.

Critical weaknesses
  1. Inconsistencies in method specification and experimental reporting. Equation (1) minimizes distance over the current task population and archive. Since the evaluated task itself belongs to the current population, its self-distance makes novelty zero, disabling its contribution to task fitness and archive admission through the novelty criterion. The supplementary demo instead compares against archive descriptors only. This may be a notation error, but leaves a central component of the implemented method unclear. Additionally, Appendix B.2 specifies random seeds {13, 42, 2026}, whereas Table 9 reports results for seeds {17, 42, 73}. These inconsistencies raise concerns about the precision of the method specification and experimental reproducibility.

  2. The main experiments are insufficiently reproducible. Appendix B.1 describes task counts and categories without identifying the source artifacts or adequately specifying seed-task construction. The construction and contents of the eight initial skills are also missing. The supplement explicitly provides a synthetic simulator rather than the four-domain experimental pipeline and data. Consequently, it does not resolve these omissions or allow independent assessment of the custom benchmark and reported improvements.

Questions for authors
  1. Which novelty reference set was used in the main experiments: the population excluding the current task, the archive alone, or something else? Does correcting Equation (1) require any experimental results to change?

  2. How were the eight initial skills constructed, and what procedures and source artifacts generated the benchmark? Can the actual initial skill packages and task-construction/evaluation artifacts be provided?

  3. What incremental benefit does discriminativeness provide beyond measured learnability? Does an ablation removing both its task-fitness term and its archive-admission criterion preserve the gains, with other components held fixed?

  4. What absolute execution limit and total GPU hours apply to each baseline under the accounting in Appendix B.3? Table 2 reports costs only for ablations, while Table 5 varies generations; neither establishes the comparison at matched total inference cost.

Reviewer_h9yFOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SCOPE, a co-evolutionary framework that jointly evolves reusable skills and executable challenges while keeping the underlying language model fixed. At each generation, SCOPE mutates challenges and selects a training curriculum based on population discriminativeness, short-horizon learning progress from temporary skill probes, behavioral failure novelty, and independent solvability checks. These selected challenges then guide persistent skill refinement, while a hall-of-fame archive replays previously informative challenges and blocks updates that cause regressions on acquired capabilities. Across four professional task domains evaluated on in-distribution, hidden-tool, unseen-difficulty, and compositional test suites, SCOPE achieves an average pass rate of 62.1%, outperforming validated adversarial generation at 59.0% and achieving its largest gain on compositional shifts.

Critical strengths
  1. Sections 3.2 through 3.4 clearly separate whether a generated challenge is executable and solvable, whether it differentiates current skills, and whether its feedback supports attainable improvement. Combining these task-selection criteria with archived regression checks addresses two competing pressures in curriculum co-evolution: new challenges must expose unaddressed weaknesses, while skill updates must preserve procedures that already work. By assigning each pressure an explicit role in curriculum selection and offspring retention, SCOPE controls co-evolution more effectively than methods that simply generate tasks to maximize failure.

  2. In Table 1, SCOPE improves pass rates on the Compositional suite from 50.9% under VALIDATED-ADV to 59.0% under an evaluation protocol that freezes the evolved skill and disables test-time adaptation. This transfer claim is reinforced by Appendix B, which enforces disjoint source artifacts, groups templates by underlying operation graphs, and samples test perturbations independently of the training curriculum. At the same time, the paper transparently notes that SCOPE shows no advantage over VALIDATED-ADV on Unseen-Difficulty, which appropriately locates the method's benefit in reusable adaptation under composed shifts instead of claiming uniform gains across all distribution shifts.

Critical weaknesses
  1. The mathematical definition of failure novelty in Equation 1 collapses to zero for every evaluated challenge as written. Specifically, Equation 1 defines the novelty of a challenge xx in the active task population TgT_g as the minimum failure-descriptor distance between xx and all members of the union of TgT_g and the archive HgH_g. Because xx itself belongs to TgT_g, this minimization includes the zero self-distance between xx and itself, causing the novelty term to vanish identically. Moreover, Appendix A.1 and Algorithm 1 repeat this exact reference set without excluding xx or specifying an alternative comparison group. Since failure novelty directly governs both task fitness in Equation 5 and hall-of-fame admission in Equation 9, the paper must reconcile this formulation with the actual implementation.

  2. The matched execution budget claim in Table 1 is unverifiable because SCOPE incurs heavy hidden overheads from solvability checks, probe refinements, and archive replays that baselines omit. Without a clear cap on total environment executions or method-specific execution counts, I cannot tell whether methods were matched by generation count or actual compute work. This leaves a critical ambiguity as to whether SCOPE's performance edge stems from its core design or simply from consuming more computational resources.

  3. Severe inconsistencies in random seeds and standard deviations across the appendix undermine the reliability and reproducibility of the results. Appendix B.2 claims that all runs use seeds 13, 42, and 2026, but Table 9 reports them under 17, 42, and 73. Moreover, aggregate standard deviations fluctuate unpredictably across tables for identical conditions, such as SCOPE shifting between 0.8 and 0.7. These contradictions break the direct mapping between the protocol and the reported data, making it impossible to interpret the experimental variability with confidence.

Questions for authors
  1. In the implementation of Equation 1, does the novelty calculation exclude the current challenge xx from TgT_g, or does it compare xx against a different reference population altogether? Please specify the exact comparison set and failure-descriptor distance metric used in code, and clarify whether correcting the written definition changes any reported result.

  2. Across the methods compared in Table 1 and Table 2, how many environment executions and model calls does each method spend on challenge generation, solvability verification, persistent skill refinement, archive replay, and checkpoint validation? When the computation saved by removing learnability probing is reallocated to additional persistent skill updates, does SCOPE still outperform the variant without learnability, and how is the learning-progress indicator in the unsolvability penalty of Equation 4 handled in that variant?

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

This paper is about reusing skills that an LLM has and how to improve the reusable skills. It looks at tasks like spreadsheet editing, document and question answering etc... The key here is that model weights are frozen and the skill text and the artifacts considered are the things which evolve. Like many "evolving agents" its co-evolution are the skills which are basically mutated by teh LLM and executable tasks that are mutated to probe the skills’ current weaknesses.

Critical strengths

Some strenghts of the paper are: that it treats agent skill evolution as two coupled populations, skills and executable challenges, so the training distribution follows the skill population’s changing capability boundary. A solvability gate checks specification consistency, executability, and solvability, using a reference program or verified consensus from a three-model committee which is very good. Also, Skills evolve by rank selection, refinement from failed traces and appreciated is that there is a retention constraint against an archive/trace of past challenges. Also the "Hall-of-fame replay" reduces forgetting.

Critical weaknesses

The major concerns are tha there is very little if no comparison with existing skill-evolution methods (co-evolving verification, multi-round skill refinement) or established curriculum methods. The adversarial baselines were designed by the authors. Also, The weights , thresholds , and beta are fixed, and only the replay size is varied. It is unclear how fragile the ranking of methods is. All four domains and the shift suites are built by the authors. There are no results on standard skill or agent benchmarks, so comparability and the realism of the shifts are hard to judge. Double check the references in the paper and use the same format for references...pls.

Questions for authors

Is the frozen checkpoint evaluation separate from Table 1? If so, why are the differences identical? What are the pass rates of the initial seed skills, of an agent with no skill, and of a zero-shot best-of-N skill baseline? How does SCOPE compare with existing skill-evolution methods and with prioritized level replay, ACCEL, or learnability-sampling curricula, why was a standard baseline not used? The appendix says seeds {13, 42, 2026}, but the per-seed table uses seeds 17, 42, and 73. (Did I miss something?)

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

Related papers

Discussion (0)

Sign in to comment.