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

Jie Ma; Zhipeng Qian; Yufei Ma; Zihan Liang; Jiayi Ji; Qingpeng Cai; Ben Chen; Peng Jiang

cs.CVOct 8, 2026

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

Authors: Jie Ma, Zhipeng Qian, Yufei Ma, Zihan Liang, Jiayi Ji, Qingpeng Cai, Ben Chen, Peng Jiang, +1 more

Organizations: Xiamen University · Kuaishou Technology · Nankai University

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 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.

Figures & tables

Figure 1: Comparison of self-evolving skill library paradigms. (a) Existing methods rely on growth-oriented evolution. Failures prompt the addition of new skills, but previously stored skills are rarely questioned, which often leads to boundary conflicts. (b) Skill-V treats skills as versioned, falsifiable contracts. It adds missing skills from failures (Path 1) and revises existing operational boundaries when rubric judgments disagree with task outcomes (Path 2). Through validated revisions, the library expands its coverage while actively maintaining the reliability of stored knowledge.

Figure 2: Overview of Skill-V. Skill-V first retrieves and filters task-relevant skills to guide policy interaction, then uses the resulting trajectories and environment outcomes to evolve the skill library. Failures expose missing strategies and lead to the addition of new semantic skills, whereas rubric–outcome disagreements provide counterexamples for revising existing executable skill contracts. Contract revisions are committed only after deterministic validation, and the updated library is reused in future interactions.

Closed-source LLMs
	ALFWorld							WebShop
Method	Pick	Look	Clean	Heat	Cool	Pick2	All	Score	Succ.
GPT-4o	75.3	60.8	31.2	56.7	21.6	49.8	48.0	31.8	23.7
Gemini-2.5-Pro	92.8	63.3	62.1	69.0	26.6	58.7	60.3	42.5	35.9
Training-free agents with Qwen2.5-7B-Instruct
Qwen2.5	33.4	21.6	19.3	6.9	2.8	3.2	14.8	26.4	7.8

Table 1: Main results on ALFWorld and WebShop. ALFWorld reports success rates (%) for each task type and overall. WebShop reports the average task score and success rate (%). The best result is shown in bold and the second best is underlined .

\dagger

Skill1 expands its ALFWorld skill library to its capacity of

5{,}000

skills; stored units may not be directly comparable across methods.

Table 4

Variant	Success (%) $\uparrow$	Committed	Non- regressive	Pass full gate	Disc	FNR	BAcc
	Task perf.	Revisions	Retrospective audit		Metric regressions $\downarrow$
w/o Evidence Gate	75.00	46	37 (80.40%)	27 (58.70%)	9	0	3
Full Skill-V	95.30	10	10 (100.00%)	10 (100.00%)	0	0	0

Table 4: Effect of the Evidence Gate on ALFWorld task performance and revision reliability. Non-regressive revisions do not worsen Disc, FNR, or BAcc on replay evidence. Passing the full gate additionally requires satisfying all remaining acceptance conditions. Regression counts are not mutually exclusive.

Table 6

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Stored content	Role
Semantic specification $\phi_{i}$	skill_id , title , principle , when_to_apply , protected skill_constraints	Defines what capability the skill represents and establishes the semantic boundary that a revision is not allowed to overwrite.
Reusable procedure $p_{i}$	Ordered natural-language steps	Conditions the policy during interaction and may be refined when experience reveals an incomplete strategy.
Executable rubric $\rho_{i}$	Applicability, required steps, forbidden behaviors, preconditions, postconditions, and termination conditions	Translates the intended behavior into conservative criteria that can be checked deterministically against an executed trajectory.
Evidence interface	Criterion identifier, severity, weight, match mode, source, operator, and pattern	Grounds each executable criterion directly in the task, action, observation, or trajectory rather than in the policy’s hidden reasoning.
Evolution record $\nu_{i}$	Skill version, rubric version, traceability, update reason, and evidence case identifiers	Links each committed version to its source skill and to the specific experience that motivated the change.

Appendix

Table 7: Fields of a stored Skill-V contract and their roles in skill use and evolution.

Case 1: Tightening an under-specified boundary gen_004 : Track Counts & Progress
Skill intent, preserved: Track how many goal objects remain and terminate only after the required count is reached.
Counterexample Contract: pass; outcome: failure. Only one target object was collected in a quantity-two task.	$\Longrightarrow$	Rubric change Before: no executable distinct-object count. After: require $\geq 2$ take events on distinct objects with a hard event_count criterion.	$\Longrightarrow$	Replay validation Old Revised Disc $\downarrow$ 0.400 0.100 FNR $\downarrow$ 1.000 0.000 BAcc $\uparrow$ 0.500 0.917 Support: $10\rightarrow 10$ Accepted Skill v1 $\rightarrow$ v2; Rubric v1 $\rightarrow$ v2.
Case 2: Relaxing an over-restrictive boundary gen_008 : Avoid Redundant Rechecks
Skill intent, preserved: Record inspected locations and avoid revisiting them unless new evidence suggests a state change.
Counterexample Contract: fail; outcome: success. A successful trajectory contained a brief pair of identical consecutive actions.	$\Longrightarrow$	Rubric change Before: flag two identical consecutive actions as a loop. After: preserve loop detection but trigger it after three identical consecutive actions.	$\Longrightarrow$	Replay validation Old Revised Disc $\downarrow$ 0.500 0.333 FNR $\downarrow$ 1.000 0.667 BAcc $\uparrow$ 0.500 0.667 Support: $12\rightarrow 12$ Accepted Skill v1 $\rightarrow$ v2; Rubric v2 $\rightarrow$ v3.

Appendix

Table 8: How outcome evidence drives contract evolution in the main ALFWorld experiment at step 150. Skill-V tightens an under-specified boundary that admits failed behavior and relaxes an over-restrictive boundary that rejects successful behavior. Each candidate is evaluated against its previous version on identical replay support. The second revision is accepted as a non-regressive improvement, although it has not yet reached target thresholds.

Rubric	Outcome	Agreement	Interpretation
Pass	Success	Yes	Evidence supporting the current operational boundary.
Fail	Failure	Yes	Consistent negative evidence.
Pass	Failure	No	The contract may be too permissive or under-specified.
Fail	Success	No	The contract may be too restrictive.

Appendix

Table 9: Interpretation of rubric-outcome evidence.

Configuration	ALFWorld	WebShop
Training batch size	16	32
Validation batch size	64	128
Rollouts per instruction	8	8
Maximum environment steps	50	15
Maximum prompt length	8192	6000
Maximum response length	512	768

Appendix

Table 10: Domain-specific configuration for the reported step-150 results. Update limits refer to per-library-update event budgets rather than global limits.

Domain	Initial	Step 150	Net Growth	Accepted Revisions
ALFWorld	55	205	150	10
WebShop	66	153	87	14

Appendix

Table 11: Skill-library evolution through step 150. Total entries include common mistakes, which are stored separately from retrievable contracts.

Metric	Value
Validation success (%)	95.3
Validation task score	8.130
Training-episode success (%)	86.7
Validation rubric score	0.894
Rubric-outcome disagreement	0.033
Hard-criterion violation rate	0.062

Appendix

Table 12: Aggregate ALFWorld diagnostics at the selected step-150 checkpoint.

Explore similar work

Oct 7, 2026cs.AI

SkillSandbox: Skill Verification via Dynamic Scenario Synthesis

Self-evolving agents distill task-solving experience into skills for future reuse, but these skills can encode incorrect procedures or non-transferable knowledge. It is therefore critical to verify each skill's reusability: whether its guidance remains useful beyond the experience from which it was distilled. Such verification requires observing how a skill affects execution in new tasks, yet existing tasks may not expose the situations where the target skill can actually be exercised. To construct such situations, we propose SkillSandbox, a framework that dynamically synthesizes a task and its environment for each skill that are skill-relevant yet novel. A Proposer specifies the conditions to preserve and the source-specific details to vary, a Builder constructs an executable scenario, and a Verifier compares executions with and without the skill. The Verifier assesses executability, utility, and efficiency to assign a Keep or Reject verdict, determining whether the skill enters the library. Across ALFWorld and WebShop with three models, SkillSandbox consistently yields the strongest downstream performance and improved execution efficiency. Further analyses examine whether these gains reflect accurate assessment of skill reusability and identify which components of SkillSandbox contribute to them.

Serin Kim, Kwangwook Seo, Dokyung Song +2

Yonsei University

Agent Skill Learning Self-Evolving Agents

Ask PDF

Jun 12, 2026cs.AI

SkillAudit: Ground-Truth-Free Skill Evolution via Paired Trajectory Auditing

Agent skills are structured procedural packages that guide frozen LLM agents in specialized workflows. Skills rarely remain sufficient after deployment: edge cases, API changes, and deployment constraints become visible only through use, making skill evolution a practical necessity. Existing methods depend on privileged feedback such as held-out validation scores, hidden test outcomes, or environment rewards -- signals often unavailable when a practitioner has only a task description and workspace data. We introduce SkillAudit, a framework for evolving agent skills without ground-truth feedback. The key idea is paired trajectory auditing: at each iteration, the same task is executed with and without the candidate skill, isolating how the skill changes agent behavior without external labels. To turn behavioral differences into edit guidance, SkillAudit uses Process-Aligned Contrastive Evaluation (PACE), a cluster of evaluators that maps trajectory divergences to diagnostic signals linked to specific passages in the skill document. A structural verifier, compiled once from the task specification and then fixed, checks task constraints and rolls back harmful updates. SkillAudit routes edits through two pipelines: Refine removes noisy or irrelevant guidance from broadly useful skills, while Repair replaces passages that conflict with the task. Across 89 containerized tasks spanning 8 professional domains, SkillAudit achieves 73.9% average task reward, outperforming an agent without skills (40.9%) and the static expert skill (56.7%). These gains are obtained without accessing hidden tests, reference solutions, or external scoring functions during evolution.

Haowen Gao, Haoran Chen, Can Wang +5

State Key Laboratory of AI Safety, Institute of Computing Technology, CAS, Beijing, China · Tongyi Lab, Alibaba Group, Beijing, China

Algorithmic Auditing LLM Agent Self-Improvement LLM Agent Skill Learning Reference-Free Evaluation

Ask PDF

Aug 6, 2026cs.AI

SkillHEX: Improving Agent Skills via Hypothesis-Driven Autonomous Exploration and Exploitation

Although agent skills equip LLMs with reusable procedural knowledge, manual maintenance suffers from high costs, unscalability, and misalignment. Real-world deployments thus require autonomous, on-demand skill evolution at test time, constrained by limited interaction budgets and a lack of training or validation sets. This setting introduces a severe sparse reward challenge, where outcomes conflate multiple latent failure causes. Under such ambiguity, existing methods that greedily refine a single incumbent skill are particularly vulnerable to an exploitation trap, allowing early misdiagnoses to exhaust limited trials along unproductive trajectories. To address this, we introduce SkillHEX, a closed-loop framework coupling hypothesis-driven self-verification with evidence-guided tree search. SkillHEX translates falsifiable failure hypotheses into executable tests, producing diagnostic evidence as dense reward without additional environment attempts. This evidence guides a search over persistent skill-revision branches, dynamically balancing the exploitation of supported edits with the exploration of plausible alternatives. Evaluated on 87 tasks from SkillsBench, SkillHEX outperforms existing self-evolving methods and achieves an average pass rate of 55.9% and 57.9% using GPT-5.3-Codex and Claude Opus 4.7 under a five-iteration budget, respectively.

Yuru Feng, Yaoqi Chen, Beidi Zhao +7

Microsoft · University of California, San Diego · University of Science and Technology of China +1

Test-Time Adaptation LLM Agent Self-Improvement LLM Agent Skill Learning LLM Agent Verification LLM Agent Reliability

Ask PDF