VCE-Skill: Enhancing Skill Self-Evolution with Version-Change Experience

Jianming Chen; Xuanbin Ye; Yawen Wang; Junjie Wang; Qing Wang; Fanjiang XU

cs.MAAug 17, 2026

VCE-Skill: Enhancing Skill Self-Evolution with Version-Change Experience

Authors: Jianming Chen, Xuanbin Ye, Yawen Wang, Junjie Wang, Qing Wang, Fanjiang XU

Organizations: Institute of Software Chinese Academy of Sciences, Beijing, China · Science & Technology on Integrated Information System Laboratory, Beijing, China · State Key Laboratory of Complex System Modeling and Simulation Technology, Beijing, China · University of Chinese Academy of Sciences, Beijing, China ·

Abstract

Agents increasingly rely on reusable skills to encode task knowledge, tool-use procedures, and validation rules. Existing skill self-evolution methods primarily revise skills using execution trajectories collected from current tasks, leaving the evolution knowledge accumulated in public skill version histories largely untapped. Our pilot study reveals a clear complementarity between the two sources: public skill changes provide reusable evolution priors, whereas trajectories provide evidence grounded in the current task. Motivated by this, we propose VCE-Skill, which distills noisy and implementation-specific public skill changes into reusable, structured version-change experience and adaptively fuses it with trajectory-derived proposals from the base evolver, thereby exploiting external experience while retaining task-specific evidence. Extensive experiments demonstrate that VCE-Skill improves skill self-evolution, increasing mean scores by 3.20--4.98 points; transfer experiments further show that the resulting skills achieve stronger cross-model transfer performance. Our work highlights public skill version changes as a previously underexplored yet effective source of prior knowledge and advances trajectory-driven skill self-evolution.

Figures & tables

Figure 1: Public and trajectory-derived skill changes.

Figure 2: Overview of VCE-Skill. The left stage distills the version-change experience from public skill versions. The right stage enhances the skill self-evolution by allocating adaptive attention between external experience and trajectory-derived proposals.

Figure 3: Illustration of experience distillation.

Skill	SearchQA	OfficeQA	ALFWorld	Spreadsheet	BFCL-v4	Avg.
No Skill	68.42	33.79	52.87	26.93	54.75	47.35
LLM Skill	67.29	33.41	54.74	29.48	55.12	48.01
EvoSkill	67.08	34.13	57.46	39.45	58.45	51.31
EvoSkill+VCE	70.46 (+3.38)	41.79 (+7.66)	59.13 (+1.67)	45.13 (+5.68)	60.71 (+2.26)	55.44 (+4.13)
SkillClaw	69.41	36.12	59.01	42.46	61.23	53.65
SkillClaw+VCE	72.48 (+3.07)	41.48 (+5.36)	62.48 (+3.47)	46.73 (+4.27)	63.77 (+2.54)	57.39 (+3.74)

Table 1: Performance comparison on different models and benchmarks. For VCE-enhanced variants, red (green) parenthesized values indicate improvements (declines) over their corresponding baseline. Bold indicates the best result.

Figure 4: The performance of different variants of VCE-Skill.

Target Model	Source Skill	SearchQA	OfficeQA	ALFWorld	Spreadsheet	BFCL-v4	Avg.
	SkillOpt	78.01	47.47	79.16	52.11	64.10	64.17
GPT-5.2	SkillOpt+VCE	83.24	55.34	86.05	57.89	68.54	70.21 (+6.04)
	SkillOpt	71.16	44.94	69.60	38.16	62.10	57.19
DeepSeek-v3.2	SkillOpt+VCE	78.12	49.53	75.11	43.94	65.48	62.44 (+5.25)
	SkillOpt	81.10	53.03	83.69	56.01	66.42	68.05
Claude Sonnet 5	SkillOpt+VCE	87.37	58.14	86.80	61.73	71.97	73.20 (+5.15)

Table 2: Cross-model transfer performance of skills.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Dimension	Annotation target	Decision rule
Component	Affected skill artifacts	Assign all affected artifact types, including instructions, scripts, references, configurations, and templates.
Intent	Objective of the update	Assign the primary outcome pursued by the change.
Pattern	Modification strategy	Abstract the operational mechanism into a repository-independent family.

Appendix

Table 1: Annotation dimensions and selected taxonomy labels used in the representative skill-change units.

Component	Instruction + Script
Operation	Modify + Modify
Intent	INTENT_10 : Compatibility and interoperability
Pattern Family	FAMILY_08 : Executable workflow automation

Appendix

Figure 1: Representative skill-change units from public version histories and trajectory-driven evolution. The public update coordinates instruction and executable-code changes to preserve compatibility with an external service. The trajectory-derived update adds a local extraction rule in response to an observed SearchQA failure.

Figure 2: Skill counts, substantive version updates, and affected-component shares across 12 task domains in the broader public-history survey. Component shares are non-exclusive because one update may affect several artifacts.

Figure 3: Trajectory utility and validation-score change across skill-evolution iterations.

Failure-mode coverage	$\Delta$ Train $\uparrow$	$\Delta$ Test $\uparrow$	Gap $\downarrow$	P2F $\downarrow$
Low (1–3 modes)	4.82	2.54	2.28	12.7
Medium (4–8 modes)	4.17	2.81	1.36	9.7
High (9–10 modes)	4.23	3.19	1.04	5.9

Appendix

Table 2: Training-test gaps and pass-to-fail regressions under different failure-mode coverage.

Symbol	Meaning
$u_{j}$	update event $\langle d_{j},\{c_{j}\},o_{j},i_{j}\rangle$ containing a raw diff segment, affected components, semantic operation, and intent
$\mathcal{B}$	version-change experience bank containing update patterns, evolution insights, and domain insights
$S_{t}$	target skill retained by the base evolver after online iteration $t$
$\tau_{t}$	task-execution trajectory produced with $S_{t}$
$z_{t}$	selection context containing target task, current skill, and available prior feedback
$\mathcal{X}_{t}$	at most five experience entries selected from $\mathcal{B}$

Appendix

Table 3: Core notation used in the implementation.

Figure 13

Figure 14

Figure 15

Figure 16

Figure 13: Ablation results averaged over all five benchmarks. Each panel uses EvoSkill, SkillClaw, or SkillOpt as the base self-evolver. w/o Dist replaces distilled experience with raw version changes, and w/o Atte assigns equal weights to external experience and self-generated proposals. Higher is better.

Source Model	Target Model	Source Skill	SearchQA	OfficeQA	ALFWorld	Spreadsheet	BFCL-v4	Avg.	Avg. Gain
Qwen3.5-27B	GPT-5.2	SkillOpt	78.01	47.47	79.16	52.11	64.10	64.17	–
	GPT-5.2	SkillOpt+VCE-Skill	83.24	55.34	86.05	57.89	68.54	70.21	6.04
	DeepSeek-v3.2	SkillOpt	71.16	44.94	69.60	38.16	62.10	57.19	–
	DeepSeek-v3.2	SkillOpt+VCE-Skill	78.12	49.53	75.11	43.94	65.48	62.44	5.25
	Claude Sonnet 5	SkillOpt	81.10	53.03	83.69	56.01	66.42	68.05	–
	Claude Sonnet 5	SkillOpt+VCE-Skill	87.37	58.14	86.80	61.73	71.97	73.20	5.15

Appendix

Table 4: Complete cross-model transfer results over all 12 directed source–target model pairs. Each source-trained skill is directly applied to the target model without further evolution. Avg. reports the mean performance across the five benchmarks, and Avg. Gain reports the mean improvement of SkillOpt+VCE-Skill over SkillOpt computed from the unrounded results. Bold indicates the higher result within each source–target pair.

Parameter	Value	SearchQA	OfficeQA	ALFWorld	Spreadsheet	BFCL-v4
$K$	3	76.24	47.54	61.30	51.39	55.98
	4	78.75	49.34	63.38	52.12	57.42
	5	78.86	49.42	63.43	52.14	57.52
	6	78.03	48.82	62.76	51.59	57.03
	7	76.98	48.07	61.90	50.89	56.41
$\eta$	0.01	75.16	46.96	61.42	50.67	56.34

Appendix

Table 5: Hyperparameter sensitivity of SkillOpt+VCE-Skill with Qwen3.5-27B. Each row reports mean performance on five benchmarks while varying one hyperparameter and fixing the remaining settings to their defaults. Bold indicates the best score within each hyperparameter group.

Motivation study
Decision Type	Samples	HAR (%) $\uparrow$	$\kappa$ $\uparrow$
Motivation-study Taxonomy	200	94.0	0.84
Skill-Change Coding	200	100.0	0.84
VCE-Skill method
Change Abstraction	200	92.5	0.82
Experience Generation	200	90.5	0.78

Appendix

Table 6: Human audit results for the LLM-based semantic decisions. HAR is computed from the final adjudicated labels, whereas

\kappa

measures agreement between the two initial annotators before adjudication.

LLM Agent Skill Learning

Large language model agent skill learning studies how agents acquire, represent, reuse, and refine procedures for multi-step tasks.

See all 398 papers

WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution
Aug 27, 2026Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng +3LLM Agent Skill Learning Memory-Augmented Agents
Google Research · Virginia Tech
Skill-V: Verifiable Self-Evolving Skill Library for Interactive Agents
Oct 8, 2026Jie Ma, Zhipeng Qian, Yufei Ma +6LLM Agent Skill Learning LLM Agent Skill Retrieval
Xiamen University · Kuaishou Technology · Nankai University
SkillEvoReg: Regularizing Agent Skill Evolution Against Overfitting
Sep 25, 2026Guanyu Nie, Fangzhou Zhu, Shixiong Kai +3LLM Agent Skill Learning Agent Skill Learning
Huawei Noah’s Ark Lab

Aug 27, 2026J/K move · Enter open · S save

WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins+2

Google Research · Virginia Tech

LLM Agent Skill Learning Memory-Augmented Agents

Open paper AskPDF arXiv

Agent skills package specialized knowledge and workflows into reusable resources that extend AI agent capabilities. Recent work automatically discovers such skills from agent experience, which enables agents to progressively adapt through interaction. However, the insights that guide skill development typically remain scattered across optimization histories, limiting their systematic reuse across iterations. We introduce WikiSkill, a framework that co-evolves agent skills with a persistent knowledge base (wiki). At a high level, WikiSkill separates raw execution experience, accumulated knowledge, and executable skills, while continuously consolidating experience into the wiki, which subsequent skill updates can build on. Across diverse benchmarks and models, WikiSkill consistently outperforms state-of-the-art skill-evolution methods and improves over no-skill baselines in most model-benchmark settings. We find that skill evolution complements model scaling: larger models generally benefit more from evolved skills, while smaller models with skills can outperform substantially larger models without them. We also find that evolved skills transfer effectively across models and model families, and skills evolved by other models can outperform self-evolved skills. Finally, our ablation studies confirm that persistent knowledge accumulation in the wiki is critical for effective skill evolution. These results demonstrate the benefits of systematically accumulating and refining agent experience for developing reusable and transferable skills.

Abstract

LLM Agent Skill Learning

Explore similar work

WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

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

SkillEvoReg: Regularizing Agent Skill Evolution Against Overfitting