SkillCommit: Evolving Agent Skills through Behaviorally Validated Scope Expansion

Yu He; Weikai Yang

cs.AIAug 15, 2026

SkillCommit: Evolving Agent Skills through Behaviorally Validated Scope Expansion

Authors: Yu He, Weikai Yang

Organizations: Nanjing University · The Hong Kong University of Science and Technology (Guangzhou)

Abstract

Large language model (LLM) agents can continually improve without parameter updates by converting historical experience into reusable procedural knowledge. However, existing methods often consolidate experience based on semantic similarity or LLM judgments, which may merge superficially related but behaviorally incompatible strategies and thereby degrade performance. To address the issue, we propose SkillCommit, an online skill evolution framework that continuously transforms experience into a hierarchical library of reusable skills. Rather than generalizing from individual experiences immediately, SkillCommit first preserves each experience as an instance-specific patch, retaining its locally validated behavior. As patches accumulate, SkillCommit abstracts those sharing a common behavioral mechanism into higher-level skills. Specifically, for each incoming skill, embedding-based retrieval first identifies candidate related skills. Cross-instance replay and an LLM-based mechanism check determine whether these skills transfer across cases and share a common underlying mechanism. Candidates that pass both checks are abstracted into a higher-level skill and committed only if it preserves the validated behavior of all constituent skills. Experiments on RuleArena, OpenExempt and KOR-Bench demonstrate that SkillCommit consistently improves agent performance across diverse domains. Moreover, the learned skills transfer across model scales and families, enabling cross-model experience transfer.

Figures & tables

Figure 1: Why semantic similarity alone is insufficient for identifying reusable skills. (a) Similar wording can conceal different behavioral mechanisms. (b) Different wording can describe the same reusable behavioral pattern.

Figure 2: SkillCommit overview. (a) Induce and validate instance-specific patches from execution feedback. (b) Identify behaviorally compatible skills through cross-instance replay and mechanism assessment. (c) Consolidate compatible skills only when the resulting abstraction preserves their validated behavior.

Skill author: DeepSeek-V4-Flash Skill user: DeepSeek-V4-Flash
Method	RuleArena		OpenExempt		KOR-Bench		Avg. $\uparrow$
	Airline $\uparrow$	NBA $\uparrow$	EV $\uparrow$	EC $\uparrow$	Logic $\uparrow$	Cipher $\uparrow$
No Skill	72.00	28.89	76.86	86.33	72.00	73.33	68.24
ExpeL	88.00	53.33	50.17	88.36	80.00	78.67	73.09
ReasoningBank	66.00	31.11	65.54	74.25	68.00	72.00	62.82
ACE	51.00	6.67	87.32	95.65	72.00	70.67	63.89

Table 1: Main results (%). RuleArena and KOR-Bench report strict accuracy; OpenExempt reports macro-F1. Bold/underline mark best/second-best scores per configuration (including ties); darker blue indicates higher scores within each task/configuration. Avg. is the unweighted mean of the six displayed task scores.

Figure 3: Error corrections and regressions on RuleArena. Help denotes corrected errors, Hurt denotes newly introduced errors, and labels indicate net improvements (Help

-

Hurt).

Stage removal
Variant	NBA (Acc. $\uparrow$ )	EC (Macro-F1 $\uparrow$ )
Full	$71.11_{\makebox[31.0pt]{}}$	$96.01_{\makebox[31.0pt]{}}$
No Skill	$28.89_{\makebox[31.0pt][l]{{\color[rgb]{0.2,0.2,1}$ (-59.4%) $}}}$	$86.33_{\makebox[31.0pt][l]{{\color[rgb]{0.2,0.2,1}$ (-10.1%) $}}}$
w/o Patch Induction	$53.33_{\makebox[31.0pt][l]{{\color[rgb]{0.2,0.2,1}$ (-25.0%) $}}}$	$91.41_{\makebox[31.0pt][l]{{\color[rgb]{0.2,0.2,1}$ (-4.8%) $}}}$
w/o Grouping	$42.22_{\makebox[31.0pt][l]{{\color[rgb]{0.2,0.2,1}$ (-40.6%) $}}}$	$88.87_{\makebox[31.0pt][l]{{\color[rgb]{0.2,0.2,1}$ (-7.4%) $}}}$
w/o Consolidation	$48.89_{\makebox[31.0pt][l]{{\color[rgb]{0.2,0.2,1}$ (-31.3%) $}}}$	$89.12_{\makebox[31.0pt][l]{{\color[rgb]{0.2,0.2,1}$ (-7.2%) $}}}$

Table 2: Ablation study of SkillCommit . Subscripts show relative changes from Full.

Figure 4: Skill transfer across model scales and families. Arrows indicate skill author

\rightarrow

frozen skill user.

Figure 5: Illustrative patch evolution in exemption reasoning: (a) local patches; (b) retrieval and replay; (c,e,g) new-skill consolidation; (d,f,h) skill expansion. Rows show full support, partial support, and rejection; brown marks failed checks.

Figure 6: Skill evolution on EC. Bars show version lifespans, dashed links show parentage, and arrows mark final active versions. Crosses mark uncommitted candidates with counts.

S_{i}^{v}[n]

labels family, version, and source count.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Task	Induct.	Test	Metric
RuleArena Airline	100	100	strict accuracy
RuleArena NBA	170	45	strict accuracy
OpenExempt EC	210	105	macro-F1
OpenExempt EV	210	105	macro-F1
KOR-Bench Logic	175	75	strict accuracy
KOR-Bench Cipher	175	75	strict accuracy

Appendix

Table 3: Construction and test cases.

Setting	Author $\rightarrow$ user	Purpose
DS matched	DeepSeek-V4-Flash $\rightarrow$ same	matched model
Qwen matched	Qwen3.6-35B-A3B $\rightarrow$ same	matched model
Cross-scale	Qwen3.6-35B-A3B $\rightarrow$ Qwen3.5-9B	same family
Cross-family	GPT-5.4 $\rightarrow$ Qwen3.5-9B	family + scale

Appendix

Table 4: Skill-author and skill-user configurations. Exact provider model IDs are given in the surrounding text.

Benchmark	Patch author	Solver/replay
RuleArena	8,192	24,576
OpenExempt	8,192	16,384
KOR-Bench	16,384	8,192

Appendix

Table 5: Maximum output tokens in the Qwen/GPT construction pipelines.

Task family	$K$	$L$	$R$	$\gamma$	$m$	$W$
RuleArena	3	3	4	forward	3	2
OpenExempt	3	3	4	$2/3$	2	2
KOR-Bench	3	3	4	$2/3$	2	2

Appendix

Table 6: Construction hyperparameters.

K

is the maximum patch generate–replay attempts,

L

is the maximum pattern-authoring attempts including the initial candidate,

R

is the semantic shortlist size,

\gamma

is the active-skill compatibility cutoff,

m

is the minimum residual group/split size, and

W

is the maximum number of active-skill source witnesses used for reverse replay. RuleArena requires the active skill to pass on the incoming source before mechanism assessment; the other pipelines use the displayed graded threshold.

Method	Memory unit	Construction	Test-time use
ExpeL	Insight rules; successful trials	Reflect over successes and failures	Retrieve similar successful trials
Reasoning Bank	Structured trajectory memories	Self-judge and append each experience	Retrieve nearest prior experience
ACE	Structured global playbook	Generator–Reflector–Curator loop	Load selected frozen playbook
Trace2 Skill	Hierarchical skill document	Local patches $\rightarrow$ hierarchical merge	Preload frozen skill file
SkillCAT	Core skill and capability nodes	Contrastive edits; assessed updates; hierarchical merge	Route to nodes and prerequisites
GSE	Skills and relation graph	Generalize edits; compare candidate and control replays	Route with prerequisites and conflicts

Appendix

Table 7: Standardized summary of the reproduced baseline pipelines.

Construction outcome	Count
Source-validated episode patches	61
Active higher-level skills	7
Residual episode patches	15
Unique episode sources covered by active skills	46
Root-skill commits	4
Split-skill commits	3

Appendix

Table 8: RuleArena NBA online construction outcomes.

Method	Generations	Tokens	When paid
ExpeL	559	5.00M	Offline
ReasoningBank	90	0.82M	Online updates
ACE	1,115	14.52M	Offline
Trace2Skill	319	2.15M	Offline
SkillCommit	952	7.72M	Offline

Appendix

Table 9: Logged memory-construction cost on RuleArena NBA with DeepSeek-V4-Flash. Tokens are logical tokens rather than monetary cost and are rounded to two decimal places.

LLM Agent Self-Improvement

Large language model agent self-improvement studies how agents use experience and feedback to improve their behavior, tools, and workflows.

See all 257 papers

SkillX: Automatically Constructing Skill Knowledge Bases for Agents
Apr 6, 2026Chenxi Wang, Zhuoyun Yu, Xin Xie +8LLM Agent Skill Learning LLM Agents
1Zhejiang University · 2Ant Digital Technologies, Ant Group.
SkillRevise: Improving LLM-Authored Agent Skills via Trace-Conditioned Skill Revision
May 31, 2026Yuxuan Liu, Zhaochen Su, Lingyun Xie +11LLM Agent Self-Improvement LLM Agent Skill Learning
The Hong Kong University of Science and Technology · Harbin Institute of Technology · Harbin Institute of Technology, Shenzhen +2
SkillClaw: Let Skills Evolve Collectively with Agentic Evolver
Apr 9, 2026Ziyu Ma, Shidong Yang, Yuxiang Ji +5LLM Agent Skill Learning LLM Agent Self-Improvement
DreamX Team, Alibaba Group

Apr 6, 2026J/K move · Enter open · S save

SkillX: Automatically Constructing Skill Knowledge Bases for Agents

Chenxi Wang, Zhuoyun Yu, Xin Xie, Wuguannan Yao+7

1Zhejiang University · 2Ant Digital Technologies, Ant Group.

LLM Agent Skill Learning LLM Agents Long-Horizon LLM Agents

Open paper AskPDF arXiv

Learning from experience is critical for building capable large language model (LLM) agents, yet prevailing self-evolving paradigms remain inefficient: agents learn in isolation, repeatedly rediscover similar behaviors from limited experience, resulting in redundant exploration and poor generalization. To address this problem, we propose SkillX, a fully automated framework for constructing a plug-and-play skill knowledge base that can be reused across agents and environments. SkillX operates through a fully automated pipeline built on three synergistic innovations: (i) Multi-Level Skills Design, which distills raw trajectories into three-tiered hierarchy of strategic plans, functional skills, and atomic skills; (ii) Iterative Skills Refinement, which automatically revises skills based on execution feedback to continuously improve library quality; and (iii) Exploratory Skills Expansion, which proactively generates and validates novel skills to expand coverage beyond seed training data. Using a strong backbone agent (GLM-4.6), we automatically build a reusable skill library and evaluate its transferability on challenging long-horizon, user-interactive benchmarks, including AppWorld, BFCL-v3, and

τ^2

-Bench. Experiments show that SkillKB consistently improves task success and execution efficiency when plugged into weaker base agents, highlighting the importance of structured, hierarchical experience representations for generalizable agent learning. Our code will be publicly available soon at https://github.com/zjunlp/SkillX.

Abstract

LLM Agent Self-Improvement

Explore similar work

SkillX: Automatically Constructing Skill Knowledge Bases for Agents

SkillRevise: Improving LLM-Authored Agent Skills via Trace-Conditioned Skill Revision

SkillClaw: Let Skills Evolve Collectively with Agentic Evolver