EASE: Behavior-Adaptive Skill Curation for Self-Evolving Agents

Zhen Xiong; Qiaoyu Tan

cs.AISep 29, 2026

EASE: Behavior-Adaptive Skill Curation for Self-Evolving Agents

Authors: Zhen Xiong, Qiaoyu Tan

Organizations: New York University · New York University Shanghai

Abstract

Agent skills provide a lightweight mechanism for self-evolving agents to accumulate reusable procedural knowledge without updating model parameters. However, existing learned skill curators typically optimize curation without explicitly modeling downstream executor behavior. We show that this can cause systematic cross-executor degradation: curators trained with different executors perform best when paired with their own training executor, indicating that effective skill curation is executor-dependent. We formulate behavior-adaptive skill curation and introduce EASE, a framework that learns a single curator that adapts its decisions to different executor behaviors. EASE maintains an online behavioral profile of recent execution patterns and conditions the curator on this profile, the current trajectory, and retrieved skills to add, modify, or remove skills from an evolving repository. We train the shared curator jointly across multiple frozen executors with reinforcement learning, using retrieval-aware and behavior-aware temporal attribution to focus optimization on curation actions with observable downstream influence. Across ALFWorld, ScienceWorld, and WebShop, with executors ranging from Qwen3-8B/32B and GPT-OSS-120B to unseen Kimi K2.6, DeepSeek V4 Flash, and Gemini 3.5 Flash, EASE outperforms strong skill- and memory-based baselines without per-executor finetuning. EASE also maintains 34.5--41.0% fewer skills, improves skill retrieval by 36.3--38.7% and measured edit utility by 51.8--60.0%, and reduces deployment-time inference tokens by 9.1--14.5%. These results establish behavior-adaptive skill curation as an effective principle for building self-evolving agents.

Figures & tables

Figure 1: Skill curation is executor-dependent. We independently train a skill curator with each executor and evaluate all curator–executor pairings. Matched pairs consistently outperform mismatched ones, suggesting that effective skill curation depends on downstream executor behavior and motivating behavior-adaptive skill curation.

Figure 2: Overview of EASE framework. Top: the online loop for one executor. (1) A frozen executor

E

solves task

x_{t}

with skills retrieved from the repository

\mathcal{S}_{t}

. (2) The resulting trajectory and feedback update the online behavioral profile

P_{t}(E)

, which summarizes execution statistics over recent

W

tasks. (3) Conditioned on the executor descriptor

e_{t}=[P_{t}(E);m(E)]

, the current trajectory, and the retrieved skills, a skill curator

\pi_{\theta}

performs different operations. (4) The updated repository

\mathcal{S}_{t+1}

serves subsequent tasks. Bottom left:

\pi_{\theta}

is trained with GRPO across a pool of frozen executors with temporal-attribution gates. Bottom right: at deployment, the frozen curator adapts to unseen executors without per-executor finetuning.

Training-Free
Method	Qwen3-8B ( $\alpha$ )				Qwen3-32B ( $\beta$ )				GPT-OSS-120B ( $\gamma$ )
Method	ALF	SciW	WS	Avg.	ALF	SciW	WS	Avg.	ALF	SciW	WS	Avg.
No Memory	39.6	24.9	33.3	32.6	46.3	32.1	41.5	40.0	51.9	35.1	43.6	43.5
AWM	41.0	24.3	30.7	32.0	47.1	33.5	41.9	40.8	50.1	36.4	44.8	43.8
ReasoningBank	43.3	25.7	35.4	34.8	49.8	35.1	43.4	42.8	53.7	38.1	46.8	46.2
MemP	42.7	26.4	35.7	34.9	48.2	36.8	45.3	43.4	54.1	37.4	45.3	45.6

Table 1: Main results on ALFWorld (ALF), ScienceWorld (SciW), and WebShop (WS). We report category-macro success rate (%) on ALF and the benchmark score on SciW and WS; Avg. is their arithmetic mean. Upper : the three executors in the curator’s training pool (

\alpha

\beta

\gamma

). Lower : three unseen executors; SkillOS- best reports, per executor and benchmark, the highest score among SkillOS-

\alpha

, SkillOS-

\beta

, and SkillOS-

\gamma

. Bold and underline mark the best and second-best results.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Value
Retrieved skills, $k$	5
Profile window, $W$	16
Rollouts per group, $N$	8
Training tasks per stream, $T$	10
Skill-free references per task and executor, $m$	4
Behavioral-deviation threshold, $\tau$	0.05

Appendix

Table 2: Method and evaluation settings. The profile window is varied only in the profile-dynamics diagnostics. ALFWorld evaluation tasks are drawn from its 134-task valid unseen split.

Setting	Value
Curator backbone	Qwen3-8B
Finetuning method	LoRA
LoRA rank	16
LoRA scaling factor	32
LoRA dropout	0.05
LoRA target modules	q_proj , k_proj , v_proj , o_proj , gate_proj , up_proj , down_proj

Appendix

Table 3: Curator optimization and quality-judge settings. Executor models and the judge remain frozen throughout training; only the curator’s LoRA parameters are optimized.

Family	Example statistics and observability
Action validity	Rejected-action rate and parse failures from action validation.
Failure recovery	Repeated rejected actions; steps to the next accepted action, with unrecovered episodes counted separately.
State tracking	Repeated visits to logged states; precondition violations when exposed by the environment.
Efficiency	Steps per episode and frequency of reaching the interaction budget.
Error types	Subtask failures or reward-component losses when the environment exposes them.
Skill utilization	Retrieval frequency; explicit skill references or invocations when logged; outcomes conditional on retrieval or observed use, reported separately.

Appendix

Table 4: Executor behavioral profile features. Examples are included only when supported by observable signals; retrieval alone is not evidence that a skill was followed.

Configuration	Executors
Configuration	Qwen3-8B	Qwen3-32B	GPT-OSS-120B
Pooled, unconditioned	47.5	60.2	64.5
Static-card conditioning	49.3	64.8	68.2
EASE ( ours )	55.0	69.7	73.1

Appendix

Table 5: Conditioned curator comparisons on ALFWorld. Entries are category-macro SR (%). All configurations enable both temporal-attribution gates; the first two rows correspond to the “No descriptor” and “Static card only” variants in Figure 4 a.

Configuration	Executors
Configuration	Qwen3-8B	Qwen3-32B	GPT-OSS-120B
Updating profile, $W_{\mathrm{ref}}=16$	58.0	72.3	75.2
Profile frozen at $t_{0}$ , $W_{\mathrm{ref}}=16$	54.2	70.1	74.1
Updating profile, $W_{\mathrm{short}}=4$	56.3	71.3	74.6
Updating profile, $W_{\mathrm{long}}=64$	58.3	72.1	75.3
Updating profile, full history	56.7	71.4	74.7

Appendix

Table 6: Profile-dynamics diagnostics. Entries report category-macro SR (%) on the common 70-task suffix following a 64-task warm-up in the 134-task valid unseen stream. Except for the frozen-repository control, all conditions continue editing skills. The short, reference, and long windows contain 4, 16, and 64 completed tasks, respectively.

Executor: Qwen3-8B
Metric	SkillOS-base	SkillOS	EASE
SR (%)	43.2	50.0	55.0
Curator tokens / task	3k	2.5k	2.65k
Executor tokens / task	9k	7.5k	6.9k
Skill-context tokens / task	3.4k	2.8k	1.9k
Profile/card tokens / task	-	-	245

Appendix

Table 7: Performance and inference-token measurements on ALFWorld. SR for SkillOS and EASE matches Table 1 . Token counts and steps are per-task averages; skill-context and descriptor tokens are included in the curator/executor totals. Dashes denote the absence of a profile/card input.

Executor: Qwen3-8B
Metric	SkillOS-base	SkillOS	EASE
Score	25.8	30.6	31.5
Curator tokens / task	5.8k	5k	5.25k
Executor tokens / task	27k	23k	19k
Skill-context tokens / task	7.8k	6.2k	4.2k
Profile/card tokens / task	-	-	250

Appendix

Table 8: Performance and inference-token measurements on ScienceWorld. Scores for SkillOS and EASE match Table 1 . Token counts and steps are per-task averages; skill-context and descriptor tokens are included in the curator/executor totals. Dashes denote the absence of a profile/card input.

Executor: Qwen3-8B
Metric	SkillOS-base	SkillOS	EASE
Score	35.2	40.6	42.8
Curator tokens / task	6k	5.2k	5.45k
Executor tokens / task	28k	25k	21.5k
Skill-context tokens / task	8k	6.5k	4.5k
Profile/card tokens / task	-	-	250

Appendix

Table 9: Performance and inference-token measurements on WebShop. Scores for SkillOS and EASE match Table 1 . Token counts and steps are per-task averages; skill-context and descriptor tokens are included in the curator/executor totals. Dashes denote the absence of a profile/card input.

Executor: Qwen3-8B
Metric	SkillOS-base	SkillOS	EASE
Version retrieval rate (%)	29.5	47.2	64.6
Edit-rollback $\Delta$ SR ( $pp$ )	+0.5	+1.5	+3.0
Repository size (skills)	85	60	35
Executor: Qwen3-32B
Version retrieval rate (%)	31.4	48.6	67.4

Appendix

Table 10: Skill utilization rate comparison on ALFWorld.

Executor: Qwen3-8B
Metric	SkillOS-base	SkillOS	EASE
Version retrieval rate (%)	30.4	48.9	68.3
Edit-rollback $\Delta$ Score (points)	+0.6	+1.1	+1.8
Repository size (skills)	98	70	44
Executor: Qwen3-32B
Version retrieval rate (%)	31.5	50.0	69.6

Appendix

Table 11: Skill utilization rate comparison on ScienceWorld. Retrieval rates are computed over 115/88/60, 108/82/56, and 104/78/54 complete-window versions (SkillOS-base/SkillOS/EASE) for Qwen3-8B, Qwen3-32B, and GPT-OSS-120B, respectively; edit-rollback benefits use 50 sampled edits per method and executor, each paired on four tasks.

Executor: Qwen3-8B
Metric	SkillOS-base	SkillOS	EASE
Version retrieval rate (%)	30.5	47.4	65.5
Edit-rollback $\Delta$ Score (points)	+0.6	+1.0	+1.7
Repository size (skills)	58	41	27
Executor: Qwen3-32B
Version retrieval rate (%)	30.8	49.3	67.3

Appendix

Table 12: Skill utilization rate comparison on WebShop. Retrieval rates are computed over 95/76/55, 91/75/52, and 88/71/47 complete-window versions (SkillOS-base/SkillOS/EASE) for Qwen3-8B, Qwen3-32B, and GPT-OSS-120B, respectively; edit-rollback benefits use 50 sampled edits per method and executor, each paired on four tasks.

Task order	1	2	3	4	5	6
SR	57.5	56.7	55.8	57.5	57.5	54.2

Appendix

Table 13: Task-order analysis over six permutations of the same 120 ALFWorld tasks with Qwen3-8B as the executor. SR is the stream-level success rate (%).

Explore similar work

May 7, 2026cs.AI

SkillOS: Learning Skill Curation for Self-Evolving Agents

LLM-based agents are increasingly deployed to handle streaming tasks, yet they often remain one-off problem solvers that fail to learn from past interactions. Reusable skills distilled from experience provide a natural substrate for self-evolution, where high-quality skill curation serves as the key bottleneck. Existing approaches either rely on manual skill curation, prescribe heuristic skill operations, or train for short-horizon skill operations. However, they still struggle to learn complex long-term curation policies from indirect and delayed feedback. To tackle this challenge, we propose SkillOS, an experience-driven RL training recipe for learning skill curation in self-evolving agents. SkillOS pairs a frozen agent executor that retrieves and applies skills with a trainable skill curator that updates an external SkillRepo from accumulated experience. To provide learning signals for curation, we design composite rewards and train on grouped task streams based on skill-relevant task dependencies, where earlier trajectories update the SkillRepo, and later related tasks evaluate these updates. Across multi-turn agentic tasks and single-turn reasoning tasks, SkillOS consistently outperforms memory-free and strong memory-based baselines in both effectiveness and efficiency, with the learned skill curator generalizing across different executor backbones and task domains. Further analyses show that the learned curator produces more targeted skill use, while the skills in SkillRepo evolve into more richly structured Markdown files that encode higher-level meta-skills over time.

Siru Ouyang, Jun Yan, Yanfei Chen +13

University of Illinois Urbana-Champaign · Google Cloud AI Research · Massachusetts Institute of Technology

Self-Evolving Agents Skills

Ask PDF

May 11, 2026cs.AI

SkillEvolver: Skill Learning as a Meta-Skill

Agent skills today are static artifact: authored once -- by human curation or one-shot generation from parametric knowledge -- and then consumed unchanged, with no mechanism to improve from real use. We propose \textbf{SkillEvolver}, a lightweight, plug-and-play solution for online skill learning, in which a single meta-skill iteratively authors, deploys, and refines domain-specific skills. The learning target of SkillEvolver is the skill's prose and code, not model weights, so that the resulting artifact drops into any agent without retraining; and the meta-skill itself is just another skill, loaded through the same interface by any protocol-compliant CLI-agent. Unlike trace-distillation, the meta-skill refines only after deploying the learnt skill, such that the learning signal comes from failures another agent encounters while using it -- not from exploratory traces alone. Refinement iterations are governed by a fresh-agent overfit audit that catches possible leakage as well as deployed-skill-specific failures, including the silent-bypass mode in which a skill appears valid in content but is never invoked at runtime. On

83

SkillsBench tasks spanning

15^{+}

domains, SkillEvolver reaches

56.8\%

accuracy versus

43.6\%

for curated human skills and

29.9\%

for the no-skill baseline; on three GPU kernel optimization tasks from KernelBench, it also raises mean speedup from

1.16

1.51

on average.

Genrui Zhang, Erle Zhu, Jinfeng Zhou +2

2Beijing Jiaotong University · 1Tsinghua University

Skills

Ask PDF

May 31, 2026cs.AI

SkillSmith: Co-Evolving Skills and Tools for Self-Improving Agent Systems

Recent self-evolving agents have shown that skills can be discovered, refined, and accumulated through execution. However, existing skill-evolution frameworks typically assume a fixed tool layer and evaluate each skill independently, limiting their ability to repair tool-level failures or reason about interactions among skills. We propose SkillSmith, a synergy-aware skill-tool co-evolution framework. SkillSmith introduces a unified proposal space in which reflection produces atomic bundles that jointly modify skills and tools, allowing tools to be wrapped, edited, composed, split, or retired when skill evolution identifies a reusable capability gap. To guide this joint search, SkillSmith maintains an ecological utility model inspired by Lotka-Volterra dynamics, where an interaction matrix estimated from execution traces captures pairwise complementarity and conflict among skills and provides pressure signals for retrieval, mutation prioritization, and retirement. Furthermore, SkillSmith records anti-patterns, including failure signatures, causal attributions, and remedies, to accelerate diagnosis and veto proposals that repeat known mistakes. Experiments on three benchmarks, including WildClawBench, and five Qwen3.5 model scales show that SkillSmith consistently outperforms strong baselines, with gains that amplify as task complexity and multi-skill co-activation increase.

Yangbo Wei, Zhen Huang, Shaoqiang Lu +4

Shanghai Jiao Tong University, Shanghai, China · Eastern Institute of Technology, Ningbo, China · University of Science and Technology of China, Hefei, China +2

Skill Evolution Self-Evolving Agents Self-Improving Agents Skills

Ask PDF