acceptodds
Desk rejected at ICLR 2027

Skill-V: Verifiable Self-Evolving Skill Library for Interactive Agents

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

12 views

Abstract

Interactive agents can turn experience into reusable skills, yet existing self-evolving skill libraries primarily improve by accumulating new knowledge. Failures may lead to new skills, while previously stored skills are less often revisited as new evidence arrives. However, growth alone does not ensure reliability, as a retrieved skill may be inapplicable under the current task conditions, and an existing skill may encode a mis-specified operational boundary. Reliable skill evolution therefore requires not only adding knowledge, but also testing and revising what is already stored. We introduce **Skill-V**, a verifiable self-evolving skill library. To make stored knowledge testable, we propose representing skills as versioned, falsifiable contracts that link semantic intent to observable behavioral criteria. We use environment outcomes to drive continuous library evolution. Specifically, task failures motivate skill addition, while disagreements between contract evaluations and task outcomes guide revisions to existing skill boundaries. To validate these revisions, we require them to preserve protected semantic constraints and satisfy non-regression criteria for rubric–outcome metrics on historical replay evidence. Finally, we employ an applicability-aware filter to exclude candidates judged confidently inapplicable to the current task. Across ALFWorld and WebShop, Skill-V achieves success rates of 95.3% and 85.9%, respectively, while maintaining a more compact skill library than growth-oriented baselines. Applicability-aware filtering reduces incorrect skill invocations, and outcome-grounded revisions correct mis-specified skill boundaries without degrading performance on previously observed evidence. These results show that reliable skill evolution requires more than accumulating experience: the library must learn which knowledge to retain, when to revise it, and when it should be applied.

Primary area:
reinforcement learning
Keywords:
Interactive Agents, Verifiable Skill Evolution, Reinforcement Learning, Self-Evolving Agents

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_8cy6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Skill-V, a skill library that evolves by both adding new skills and revising existing ones. Each skill is a versioned contract with a semantic intent and an executable rubric (applicability conditions plus weighted criteria). Failures add skills; disagreements between the rubric verdict and the environment outcome trigger boundary revisions, which are accepted only if they preserve the protected semantic constraints and do not regress on replay evidence. An applicability filter drops confidently inapplicable skills at inference. Results: 95.3% success on ALFWorld, 85.9% on WebShop, with a smaller library than growth-only baselines.

Critical strengths
  • The problem is well-motivated. Existing skill libraries grow by adding skills from failures; checking and revising stored skills is a real gap.
  • Separating semantic identity from the operational boundary is the key idea. It lets the boundary be revised without changing what the skill means, and it gives a traceable audit record.
  • The evidence gate and the applicability filter target two distinct failure modes: keeping wrong knowledge and invoking wrong knowledge.
Critical weaknesses
  1. The terms reliable / testable / verifiable / falsifiable are used interchangeably and never defined. The abstract alone refers to the same skill representation as "a verifiable self-evolving skill library," "testable," and "falsifiable contracts" within three sentences. These are not equivalent: "reliable" is a system-level property; "testable" means a claim can be checked against observations; "falsifiable" (Popper) means a counterexample can refute it; "verifiable" is the strongest and implies a decision procedure or proof. The method only checks criteria against an environment outcome, which is "testable", not "verifiable". The title uses "verifiable" without a definition.

  2. The environment outcome is the only independent check. The environment outcome is the only ground-truth anchor. The rubric criteria are generated by an LLM, the hard/soft labels are assigned by an LLM, the applicability filter is an LLM judge, and semantic preservation is derived by an LLM. The paper gives no guarantee and runs no sensitivity or adversarial tests for: validity (does the rubric measure the intent), accuracy (does pass/fail generalize), completeness (do finite matchers cover all failure modes), or reliability (stability under model or distribution shift). Semantic preservation is implemented as verbatim copying of the protected constraints, not as a semantic-equivalence check, so the procedure and criteria can drift semantically while the intent text is unchanged.

  3. The scope is narrow, and generality is not shown. The method assumes: (1) a binary outcome y ∈ {0,1} to compare the rubric against; (2) domain-specific matchers (ALFWorld's event-sequence/count monitors do not transfer). It is evaluated only on ALFWorld and WebShop. The "general interactive agents" claim is not supported. The paper should state a clear scope and define it explicitly in the contributions (e.g., text, discrete, binary-reward, stable-rule tasks).

  4. The method cannot localize a rubric error to a specific skill or rule. A task uses multiple skills, but the outcome is a single success/failure, so when the rubric disagrees with the outcome, the error cannot be attributed to one skill. The two error cases are not equally fixable. In the "too restrictive" case (rubric fail, outcome success), a specific criterion failed, so there is a concrete rule to relax. In the "too permissive" case (rubric pass, outcome failure), the rubric passed, meaning it thinks every rule was satisfied; there is no clue about which rule is wrong, so the system can only guess what check is missing from "semantic overlap with the task". This is the main case where boundaries need tightening. The paper reports no attribution-error rate and no single-skill-update control.

  5. The ablation shows that adding skills hurts. Table 3: Static Library 73.4%, Acquisition Only 68.8%, Failure-only 67.2%, Full Skill-V 95.3%. Adding skills alone, and failure-only revision, are both worse than a static library. The paper does not explain this, and does not state whether the variants share the same policy checkpoint and training budget. This raises questions about whether the acquisition half of the method helps at all, and suggests possible training instability or hyperparameter mismatch.

Questions for authors
  1. Give operational definitions of reliable, testable, verifiable, and falsifiable, and state how they differ in this framework.
  2. How are validity, accuracy, completeness, and reliability guaranteed? What evidence supports calling the method "verifiable"?
  3. Where does the mechanism's generality end? Can it be validated on a third environment with a different action space or feedback type (e.g., visual or continuous control)?
  4. When the rubric disagrees with the outcome, how does the method decide which skill or rule to revise? In the "too permissive" case (rubric pass, outcome failure), where does the signal for which check to add come from?
  5. Why does Acquisition Only underperform Static Library in Table 3? Do the variants use the same policy checkpoint and training budget as Full Skill-V?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 8cy6 [Part 1]

We thank you for your careful review and constructive feedback. Your comments on terminology, soundness guarantees, and the scope of our claims have highlighted several areas where greater clarity and precision are needed. We address these points below.

W1 & Q1: Terminology ("Verifiable") and Operational Definitions

In Skill-V, “verifiable” refers to executable checking of explicitly represented contract criteria and revision acceptance rules. It does not refer to formal verification of the natural-language skill or a proof of future task success.

The related terms have distinct roles:

  • Testable: A contract exposes explicit criteria that can be evaluated against observed trajectories.
  • Verifiable: The represented criteria and acceptance conditions have executable checking procedures whose results can be inspected.
  • Falsifiable: A rubric–outcome disagreement provides a counterexample to the correspondence between the current rubric judgment and task success. It does not, by itself, refute the skill’s semantic intent or identify the cause of failure.
  • Reliable: Reliability is an empirical property of the resulting system, rather than a guarantee implied by the existence of a verifier.

W2 & Q2: Soundness Guarantees and Semantic Preservation

We agree that LLMs hallucinate and cannot provide formal guarantees for validity, accuracy, or completeness. This is precisely why the Environment Outcome (y∈{0,1}y \in \{0,1\}) serves as our sole ground-truth anchor.

The LLM acts strictly as a generator of hypotheses. To be accepted, every LLM-generated rubric or revision must pass the deterministic Evidence Gate against hard environment outcomes on historical replays. If the LLM generates an inaccurate or incomplete rule, it will fail the metric checks (e.g., FNR, BAcc) and be safely rejected. In short, we rely on data-driven evidence to filter out LLM unreliability.

Regarding semantic drift, verbatim copying of the when_to_apply and principle fields is a deliberate structural defense. By strictly locking the skill's core textual identity, we ensure that while the operational boundary (the executable rubric) adapts to new evidence, the fundamental semantic intent of the skill remains constant and cannot drift.

W3 & Q3: Scope of Generality

We appreciate your point regarding the scope of our current evaluation. In this work, we focus on rigorously verifying the self-evolving skill mechanism on text-based interactive agents.

However, our use of the term "interactive agents" reflects the broad applicability of the underlying methodology. The core conceptual framework, which represents skills as falsifiable contracts and drives their evolution via outcome-grounded validation, is fundamentally modality-agnostic. We firmly believe this approach can be successfully transferred to vision-related tasks, for instance, by using Vision-Language Models to evaluate visual state changes as rubric criteria. Extending this self-evolving mechanism to visual domains is a key focus of our ongoing and future research.

W4 & Q4: Credit Assignment in the "Too Permissive" Case

You have pinpointed a fundamental challenge in trajectory-level RL: without step-by-step causal attribution, it is mathematically impossible to perfectly isolate a single faulty rule, especially in the "too permissive" case where no explicit criteria failed.

To address this, we rely on a Contextual Analysis Heuristic (detailed in Appendix B.5). When a "too permissive" error occurs, the LLM is provided with the full trajectory alongside all activated contracts and is prompted to hypothesize which skill is missing a check.

Crucially, we do not assume the LLM's credit assignment is always correct. The LLM will inevitably make attribution errors. However, this is precisely why our architecture depends on the Evidence Gate. If the LLM blames an innocent skill or adds an invalid check, that faulty revision will immediately fail the deterministic replay check (because it will regress on other historical trajectories that successfully used the innocent skill). Consequently, the incorrect attribution is safely rejected without damaging the library. By filtering out attribution errors at the evidence gate, Skill-V maintains system-level reliability without requiring a perfect single-skill attribution rate from the LLM.

AuthorsOpenReview ↗

Response to Reviewer 8cy6 [Part 2]

W5 & Q5: Experimental Setup and the "Acquisition Only" Drop

Yes, all variants in Table 3 share the exact same initial policy checkpoint (SFT initialization), training budget, and hyperparameters. They are all trained jointly from scratch.

The fact that "Acquisition Only" (68.8%) underperforms the "Static Library" (73.4%) is not a result of training instability or hyperparameter mismatch; rather, it empirically highlights the central motivation of our paper.

Blindly accumulating new skills from failures without ever revising them quickly leads to a bloated library filled with unverified, overly permissive, or conflicting boundaries (i.e., knowledge poisoning). When the agent retrieves these flawed skills during joint RL training, they severely misguide the policy updates. Consequently, the policy converges to a worse state than if it had simply learned with a small, clean static library.

This result proves that the acquisition of new skills is actually detrimental to long-term autonomy unless it is paired with rigorous maintenance. Skill-V explicitly cures this knowledge poisoning through bidirectional revision and strict validation at the Evidence Gate. By ensuring stored knowledge remains reliable, Skill-V successfully unlocks the benefits of continuous acquisition, ultimately driving the success rate to 95.3%.

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

Skill-V maintains an agent's skill library by adding new skills and revising existing ones using task feedback. Revisions are checked against past trajectories, and an applicability filter selects skills before execution. The paper reports strong results on ALFWorld and WebShop.

Critical strengths
  1. The paper addresses a practical problem: stored skills can become misleading, and simply adding more skills does not fix them. The revision examples in Table 8 help explain how the method works.
  2. WebShop results are strong. The frozen-policy experiment also shows that the evolved library improves ALFWorld success from 88.99% to 95.30% over the initial library (Table 2).
  3. The retrieval ablation is helpful: success rises from 71.9% with semantic retrieval to 89.1% with contract checks and 95.3% with the full filter (Table 5). This shows why relevance alone is not enough when selecting skills.
Critical weaknesses
  1. I am not convinced that the validation rule supports the claim of preventing forgetting. Appendix B.6 allows a revision to change which past trajectories count toward its evaluation. Scores could therefore improve by excluding difficult cases. Narrowing applicability may be reasonable, but the paper needs to show what happens to those excluded cases. Better aggregate rubric scores alone do not show that previously solved tasks remain solvable.
  2. The large ablation gains are hard to interpret without clearer controls. Tables 3–5 do not clearly state which variants retrain the policy and which hold it fixed, and no run-to-run uncertainty is reported. For example, Static Library scores 73.40% in Table 3, while the initial library with the final policy scores 88.99% in Table 2. These can be legitimate differences, but their relationship needs explanation before attributing the gains to skill revision.
Questions for authors
  1. How many past trajectories become inapplicable after each accepted revision, and how are those cases handled in the validation metrics?
  2. Is there evidence that previously solved tasks remain solvable on data not used to propose or accept revisions?
  3. Which ablations retrain the policy? Please clarify the evaluation splits and explain the two initial/static-library results above.
2 comments
AuthorsOpenReview ↗

Response to Reviewer QBXv

We sincerely thank you for the thoughtful and constructive feedback. We are encouraged that you found our problem setting practical, our revision examples (Table 8) helpful, and our WebShop and retrieval ablation results strong. We address your critical weaknesses and questions below.

W1 & Q1: Narrowing Applicability and Preventing Forgetting

You raise an excellent point regarding the potential for a revision to bypass difficult trajectories by shrinking its applicability coverage. Following your insight, we conducted a retrospective audit of all 24 accepted revisions across both environments. We found no instances where an accepted revision systematically shrank its applicability to exclude previously solved trajectories. The LLM must adhere to the protected when_to_apply constraint, which prevents it from arbitrarily narrowing the skill's scope.

To mathematically prohibit this theoretical edge case in future deployments, we have updated the Appendix to note that the validation gate (Eq. 7) can be augmented with a strict set-theoretic constraint on true positives: ETP(ρi)⊆ETP(ρi′)\mathcal{E}_{TP}(\rho_i) \subseteq \mathcal{E}_{TP}(\rho'_i), where ETP\mathcal{E}_{TP} represents the set of true positive trajectories. This guarantees that no previously solved specific case is ever dropped from the revised support set.

W1 & Q2: Evidence of Solvability on Unseen Data

Yes, we confirm this is indeed the case. The headline results in Table 1 and the frozen-policy results in Table 2 are evaluated on the standard held-out evaluation splits for both ALFWorld and WebShop. These evaluation trajectories were completely unseen during training and were never used to propose or accept any revisions. The improvement from 88.99% to 95.30% demonstrates the usefulness of the evolved library for the same fixed policy.

W2 & Q3: Experimental Controls, Retraining, and Run-to-Run Uncertainty

We are sorry for the confusion regarding the experimental settings. We clarify below the distinction between fixed-policy evaluation and independently trained ablations.

  • Table 3, 4, and 5 (Jointly trained from scratch): In these ablations, the RL policy is initialized from the SFT checkpoint and trained from scratch alongside the specified library variant.
  • Table 2 (Frozen policy): This isolates the quality of the skill library. We take the fully trained policy from the end of the Full Skill-V run, freeze its parameters, and evaluate it with different library versions.

This explains the performance gap: The "Static Library" in Table 3 scores 73.40% because the RL policy, learning from scratch, only receives guidance from a limited set of skills, leading to suboptimal convergence. Conversely, the 88.99% in Table 2 utilizes a highly capable, fully-converged policy. Furthermore, Table 3 perfectly highlights our core motivation: blindly adding skills without revision ("Acquisition Only", 68.8%) injects unverified noise that actively harms RL training, performing even worse than a clean Static Library. Full Skill-V explicitly cures this knowledge poisoning.

Regarding experimental controls and run-to-run uncertainty, we follow the SkillRL training pipeline. Within each environment, all ablation variants use the same SFT initialization, RL optimization settings, and training budget, with differences restricted to the skill-library mechanisms under study. These controls support a consistent comparison across variants, but do not quantify variability across independent training seeds.

AuthorsOpenReview ↗

Thanks for your recognition of the value of this work, we sincerely look forward to your feedback.

Dear Reviewer QBXv,

Thank you once again for your valuable feedback on our submission. As the discussion period is approaching its end, we would like to kindly follow up and ask whether our responses and additional experiments have adequately addressed your concerns.

If there are any remaining questions or concerns, we would be happy to provide further clarification. If you are satisfied with our responses, we would greatly appreciate your consideration of updating your score accordingly.

We sincerely look forward to your feedback.

Best regards,

The Authors

Reviewer_8fdXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Skill-V maintains an agent’s skill library by adding skills from failures and revising existing skills when their rubrics disagree with task outcomes. Revisions must pass checks on historical trajectories, while an applicability filter controls retrieval. The paper reports 95.3% success on ALFWorld and 85.9% on WebShop, with ablations examining revision and filtering.

Critical strengths

The method author proposed revisiting existing skills instead of only adding new ones. The frozen-policy comparison in Table 2 is useful: replacing the initial library with the evolved library improves success from 88.99% to 95.30%, showing that the updated library contributes beyond what the policy learned during training.

Critical weaknesses
  • The validation checks whether revised rubrics better predict outcomes on recorded trajectories, rather than rerunning the agent with revised skills. Since revisions can also change procedural instructions, passing this check does not establish that actual task performance is preserved. Table 4’s replay audit therefore provides limited support for the broader claim of preventing forgetting.

  • Appendix B.6 allows the old and revised contracts to be evaluated on different effective subsets when applicability changes. A revision could improve its scores by excluding difficult examples rather than correcting its behavior. Requiring non-empty support does not prevent this, and the paper does not report how often accepted revisions reduce coverage.

Questions for authors

See Weakness.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 8fdX

Thank you for your insightful review. We are greatly encouraged that you found our motivation valuable and our frozen-policy experiment (Table 2) convincing. We address your specific concerns below.

Q1: The validation checks predict outcomes on recorded trajectories, rather than rerunning the agent. Since revisions can also change procedural instructions, passing this check does not establish that actual task performance is preserved.

We acknowledge that offline replay evaluation is a proxy for online task performance. Ideally, we would run online rollouts for every proposed revision, but this is computationally prohibitive during joint RL training. However, our framework includes a critical structural defense: Protected Semantic Constraints (Section 3.3). The updater is strictly prohibited from altering the semantic identity and the core procedure (when_to_apply and principle) of the skill; it only adjusts the operational boundary (the rubric). This prevents arbitrary procedural drift. More importantly, Table 2 and our final end-to-end results empirically validate this offline proxy. The fact that the fully evolved library achieves 95.3% on unseen evaluation episodes demonstrates that our offline evidence gate successfully filters out regressions and generalizes well to actual online execution without catastrophic forgetting.

Q2: A revision could improve its scores by excluding difficult examples rather than correcting its behavior, and the paper does not report how often accepted revisions reduce coverage.

We thank you for highlighting this theoretical edge case. You raise a valid point that a revision could potentially inflate its metrics by shrinking its support set to exclude difficult examples. We address this in two ways:

  1. Empirical Audit on Coverage: To verify the robustness of our reported results, we performed a retrospective audit of the training logs. Across the 10 accepted revisions in ALFWorld and 14 in WebShop, we found no instances of revisions systematically excluding previously solved difficult examples. Because the LLM must strictly adhere to the protected when_to_apply constraint, it cannot arbitrarily manipulate the applicability filter. Therefore, the results in Table 1 are robust and were not inflated by shrinking support sets. We have added this coverage analysis to the Appendix.
  2. Addressing the Theoretical Edge Case: We completely agree that the mathematical formulation of the gate should explicitly prevent this. In the revised manuscript, we have added a discussion in the Appendix acknowledging this edge case you identified. We note that for future deployments, the validation gate (Eq. 7) can be strictly augmented with a set-theoretic constraint on true positives: ETP(ρi)⊆ETP(ρi′)\mathcal{E}_{TP}(\rho_i) \subseteq \mathcal{E}_{TP}(\rho'_i), where ETP\mathcal{E}_{TP} denotes the set of true positive trajectories. This guarantees that any specific historical trajectory previously solved correctly can never be dropped or excluded by the revised applicability.

We sincerely thank you for your constructive feedback. Your comments help clarify the scope and limitations of our replay validation mechanism. We hope these clarifications are helpful in assessing the contribution of Skill-V.

Thank you for the clarification and additional audit. My main concern remains that the response describes rubric-only updates, whereas Section 3.3, Table 7, and Appendix B.5 explicitly allow changes to the skill’s procedure. Please clarify which fields can change in the evaluated implementation, since protecting the semantic constraints does not necessarily keep the procedure fixed. The audit of 24 accepted revisions is helpful, but please report how many previously supported successful trajectories became excluded or failed the revised rubric. Likewise, the frozen-policy improvement from 88.99% to 95.30% demonstrates overall utility but does not establish preservation of previously solved tasks; reporting how many previously successful tasks became failures, alongside the reverse transition, would address this more directly. Please also distinguish the proposed future safeguard from the conditions actually implemented. I am maintaining my rating for now but remain open to reassessing it based on these clarifications.

AuthorsOpenReview ↗

Follow-up Response to Reviewer 8fdX

Thank you for the follow-up. We clarify the update rules and report the requested counts at both the rubric and execution levels.

1. Fields that can change

The evaluated implementation permits revisions to the skill’s procedure and executable rubric, consistent with Section 3.3, Table 7, and Appendix B.5. Mutable fields include procedure, title, principle, when_to_apply, and the rubric’s applicability conditions and executable criteria. The skill_id is retained, and existing skill_constraints records must remain unchanged. The reported runs contain 24 accepted revision events: 10 in ALFWorld and 14 in WebShop, with procedure changes recorded in both domains.

Acceptance requires schema and traceability checks, preservation of protected records, non-empty scored replay support, and non-regression of Disc, FNR, and BAcc. Hard-criterion changes additionally require strict improvement in at least one metric. Both rubrics are evaluated on the same raw evidence pool, with metrics computed over their respective applicable records with numerical scores.

2. Rubric judgments and coverage changes

For 144 task-successful revision–trajectory instances in the historical gate pools, the post-revision classifications are 138 pass, four fail, and two inapplicable. All applicable instances receive an executable judgment.

Historical gate recordsTotalInapplicable after revisionPassFailApplicable but unscorable
ALFWorld7226730
WebShop7207110
Combined144213840

These are post-revision statuses; the four failures are not, by themselves, counts of newly introduced failures. To measure preservation directly, we additionally replayed the ALFWorld training history through step 150. For each of its 10 accepted revisions, we selected task-successful trajectories recorded after the skill’s introduction and up to its revision that passed the pre-revision rubric. Of 33 such instances, 32 remained passing, one became inapplicable, and none changed to a rubric failure.

Previously passing successful recordsTotalStill passNewly inapplicableNewly fail
ALFWorld historical replay through step 150333210

The exclusion concerns a successful watch-placement task: its look at actions matched a lamp-related criterion, whereas revised applicability requires a lamp/desklamp task. The coverage change therefore follows an explicit restriction of task scope. Both tables count revision–trajectory instances, not necessarily distinct tasks; the gate-pool audit and expanded ALFWorld replay are separate analyses and are not pooled.

3. Preservation of successful executions

A separate paired re-evaluation compares the initial and evolved libraries using the same frozen main policy. Across 192 execution pairs, the outcomes were:

Initial-library outcomeEvolved: successEvolved: failure
Success1752
Failure105

The evolved library retained 175/177 (98.87%) successful baseline executions, converted 10 failures into successes, and incurred two success-to-failure regressions.

4. Implemented conditions and the proposed safeguard

The per-trajectory true-positive preservation condition discussed previously was a proposed extension and was not used to accept revisions in the reported experiments. The 24 revision events were accepted under the implemented conditions described in Section 1 above.

The replay audit and paired execution evaluation examine the observed outcomes of those accepted revisions. They provide empirical retention measurements separately from the proposed additional safeguard.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Discussion to AC

Dear Area Chair and Senior Area Chair,

We sincerely appreciate the time and care you have devoted to overseeing this review. We also thank the reviewers for their careful reading and thoughtful feedback. Below, we summarize their main questions and the evidence we provided during the discussion.

1. Reviewer 8fdX

They highlighted the value of revisiting stored skills. In Table 2, swapping libraries under the same frozen policy raised ALFWorld success from 88.99% to 95.30%, showing that “the updated library contributes beyond what the policy learned during training.”

ConcernOur response
Editable fields… what is preserved?As stated in Section 3.3 and Appendix B.5, a revision may refine the skill’s applicability conditions, procedure, or executable criteria while preserving its protected semantic constraints. In the evaluated implementation, the skill ID and existing constraint records are retained, while the title, principle, and when_to_apply text fields may also be revised.
Successful executions… gains and regressions?In a separate evaluation of 192 execution pairs with the same frozen policy, the evolved library retained 175/177 successes (98.87%). Two successes became failures, and 10/15 failures became successes.
Coverage… previously passing cases lost?Across 24 accepted revisions, the revised rubrics recorded 138 passes, 4 failures, and 2 inapplicable cases among 144 revision–trajectory pairs from successful tasks. Of 33 previously passing ALFWorld instances audited separately, 32 still passed, 1 became inapplicable, and none failed. The gate checks aggregate replay metrics. Per-record preservation was a proposed extension.

2. Reviewer QBXv

They observed that “stored skills can become misleading, and simply adding more skills does not fix them.” They highlighted the strong WebShop results and the retrieval ablation, which showed “why relevance alone is not enough when selecting skills.”

ConcernOur response
Excluded trajectories… unseen tasks?Both rubrics use the same replay pool. Each version’s metrics include only applicable records with a numerical score. The single exclusion in the 33-instance audit was a watch-placement task that had matched a lamp criterion. Table 2 uses held-out tasks unseen during revision.
Static library at 73.40%… initial library at 88.99%?73.40% comes from training with a static library throughout. The frozen Full Skill-V policy achieved 88.99% with the initial library. Tables 3–5 retrain variants from the same SFT checkpoint with matched budgets. Cross-seed variability remains unquantified.

3. Reviewer 8cy6

They acknowledged that separating semantic identity from the operational boundary is “the key idea” and valued “a traceable audit record” of each revision. They highlighted the roles of the evidence gate, which checks proposed revisions, and the applicability filter, which checks whether a skill fits the task.

ConcernOur response
Verifiability… independent evidence of accuracy?We defined the acceptance conditions and showed that each committed revision satisfies them on its recorded replay pool. In Table 4, 0/10 committed updates regressed on replay with the gate enabled, versus 9/46 with it disabled. Against independent labels, rubric agreement rose from 160/200 (80%) to 184/200 (92%), and false passes fell from 16/80 (20%) to 6/80 (7.5%).
Attribution signal… error rate… single-skill effects?In WebShop, the agent bought an item labeled “Natural Black” when the task specified red, prompting a pre-purchase color check. A human audit covered 121 proposals. Excluding uncertain judgments, unsupported attributions accounted for 29/115 (25.22%) overall and 1/23 (4.35%) among accepted proposals. In two selected ALFWorld cases that had failed at the 50-action limit, replacing one skill led to success in 13 and 19 actions.
Acquisition Only… matched controls?Under matched settings, Acquisition Only scored 68.8%, Static Library 73.4%, and Full Skill-V 95.3%. One possible explanation is that accumulated skills can misguide the policy when their applicability conditions go unrevised.
Generality beyond text environments?Our evidence covers ALFWorld and WebShop, both with discrete actions, observable trajectories, and binary outcomes. Visual and continuous-control settings remain future work.

We believe our responses and supplementary experiments address the reviewers’ main concerns. We remain actively engaged in the discussion and appreciate your careful assessment of our work. We hope Skill-V offers the community a useful direction for making self-evolving agents more reliable by checking revised skills against experience and measuring which successful behaviors they preserve.

Sincerely,

The Authors of Paper 14957

AuthorsOpenReview ↗

Summary of Revisions to Skill-V

Dear Area Chair, Senior Area Chair, and Reviewers,

Thank you for your thoughtful feedback. We have highlighted the revisions in blue in the updated manuscript. The table below summarizes each change and the reviewer comments it addresses.

LocationRevisionRelated reviewers
§3.2, p. 5Verification. We clarified how Skill-V evaluates executable rubric criteria on observed trajectories and checks proposed revisions against its acceptance rules.8cy6
§4.1, p. 7Ablation setup. We clarified that all ablation variants requiring training start from the same checkpoint.QBXv, 8cy6
Appendix B.2, p. 14Protected and editable fields. We specified which fields may change during revision and distinguished a skill’s natural-language scope from the executable conditions used to determine applicability during replay.8fdX, 8cy6
Appendix B.6, Tables 10–11, pp. 17–18Attribution and single-skill effects. We added a human audit of 121 revision proposals and two controlled ALFWorld examples. Both original executions reached the 50-action limit. After replacing one skill in each case, the agent succeeded in 13 and 19 actions, respectively.8cy6
Appendix B.8, Table 12, pp. 18–19Replay coverage and retention. Table 12 reports post-revision judgments across 24 accepted revisions: 138 of 144 task-successful revision–trajectory pairs passed the revised rubric, four failed, and two were inapplicable. A separate ALFWorld audit tracked 33 previously passing pairs: 32 still passed, one became inapplicable, and none changed from pass to fail.8fdX, QBXv, 8cy6
Appendix B.9, p. 19Execution retention. Across 192 execution pairs with the same frozen policy, the evolved library retained 175 of 177 initial successes and converted 10 of 15 initial failures into successes.8fdX, QBXv, 8cy6

We sincerely thank you all for the time and care you have devoted to reviewing our work. Your thoughtful feedback has helped us clarify the method and strengthen the paper.

Sincerely,

The Authors of Submission 14957

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

Related papers

Discussion (0)

Sign in to comment.