Source-linked AI summary
SkillX: Automatically Constructing Skill Knowledge Bases for Agents
Chenxi Wang, Zhuoyun Yu, Xin Xie, Wuguannan Yao, Runnan Fang, Shuofei Qiao, Kexin Cao, Guozhou Zheng, Xiang Qi, Peng Zhang, Shumin Deng
TL;DR
LLM agents often relearn tasks independently from limited experience, causing redundant exploration and weak transfer. SkillX automatically builds a reusable hierarchical skill knowledge base through extraction, refinement, and exploration-guided expansion, and its library improves weaker agents across multiple benchmarks.
Problem
Agents commonly learn in isolation from limited experience, leading to redundant exploration, weak generalization, and capability-limited skill extraction.
Method
SkillX constructs a plug-and-play knowledge base by organizing trajectories into planning, functional, and atomic skills, then iteratively refining and expanding the library.
Results
SkillX transfers effectively to other models and improves task performance and execution efficiency across challenging interactive agent benchmarks.
Takeaways & Limitations
Structured hierarchical skills provide a reusable experience representation for transferring capabilities across agents and tasks.
Takeaways & Limitations
SkillX is most naturally applicable when skills can be grounded in relatively stable tool environments, making reuse across substantially different tool ecosystems less straightforward.
Abstract
from arXiv · showhide
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 \textbf{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: \textit{(i) Multi-Level Skills Design}, which distills raw trajectories into three-tiered hierarchy of strategic plans, functional skills, and atomic skills; \textit{(ii) Iterative Skills Refinement}, which automatically revises skills based on execution feedback to continuously improve library quality; and \textit{(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.
1. Introduction
SkillX addresses inefficient, isolated experience learning by constructing reusable hierarchical skill libraries through automated extraction, refinement, and expansion. Its plug-and-play library improves weaker agents across challenging interactive benchmarks.
- Most agents approach new tasks largely from scratch, making experience reuse costly, brittle, and inefficient.
- Existing self-evolving agents face isolated learning, weak generalization, and capability bottlenecks that limit scalable transfer.
- SkillX represents experience as reusable planning, functional, and atomic skills rather than monolithic behaviors.
- SkillX automatically extracts, refines, and expands skills to build a plug-and-play knowledge base from agent experience.
- The resulting library transfers to weaker agents across AppWorld, BFCL-v3, and τ 2-Bench, improving performance by around 10% and increasing execution efficiency.
2. Preliminaries
The preliminaries formulate interactive agents as policies that generate trajectories from observations and actions. Skill-conditioned execution augments this process with retrieved skills to improve task completion and reduce redundant exploration.
- An interactive environment is modeled by states, executable actions, and transition dynamics, while the agent selects actions from its interaction context.
- A trajectory records the sequence of observations and executed actions produced during task solving.
- Without external skills, the LLM agent generates trajectories through direct reasoning over the task and interaction history.
- A skill library and retriever select task-relevant skills that condition the agent’s trajectory generation.
- The objective is to design skills and their usage so the expected task success rate improves.
3. SkillX Design and Implementation
SkillX organizes tool-centric experience into three skill levels and builds the library through trajectory extraction, iterative refinement, and exploration-guided expansion. The pipeline filters redundancy and invalid tool usage while progressively improving coverage and reuse.
- 3.1. Multi-Level Skills Design: Atomic skills specify individual tools, functional skills compose tool actions for subtasks, and planning skills organize subtask ordering, dependencies, and branching.
- 3.1. Multi-Level Skills Design: SkillX extracts multi-level skills from rollouts, compressing successful trajectories into reusable representations.
- 3.1. Multi-Level Skills Design: Functional skills are extracted for subtasks using plans and trajectory evidence, while atomic skills capture invocation patterns, parameter configurations, constraints, and failure modes.
- 3.3. Iterative Skills Refinement: Each refinement iteration rolls out the current library, extracts candidate skills, and applies operators that merge and filter them before updating the library.
- 3.3. Iterative Skills Refinement: Skill merging combines semantically similar skills and can decompose overly complex merged skills into modular reusable units.
- 3.3. Iterative Skills Refinement: Strict general and tool-specific filters remove non-portable skills and validate references against environment tool schemas.
- 3.4. Exploratory Skills Expansion: Experience-guided exploration broadens skill coverage beyond scarce seed demonstrations by exercising a wider range of tools.
4. SkillX Usage
SkillX improves retrieval relevance by rewriting retrieved planning skills into a task-specific pseudo-plan, then using its steps to retrieve and filter functional and atomic skills.
- Planning Skills Retrieval and Pseudo-Plan Rewriting: SkillX retrieves planning skills for similar tasks and rewrites them into a task-specific pseudo-plan.The pseudo-plan adapts retrieved experience to user profiles, contextual constraints, and other execution factors.
- Planning Skills Retrieval and Pseudo-Plan Rewriting: The pseudo-plan is used as an intermediate retrieval query rather than injected into the final system prompt.This design reduces the risk that hallucinated or speculative pseudo-plan content affects agent behavior.
- Functional and Atomic Skills Retrieve: Each pseudo-plan step retrieves relevant functional and atomic skills, after which duplicate skills are removed across steps.Step-level retrieval decomposes the task into targeted skill queries.
- Functional and Atomic Skills Retrieve: The LLM self-filters retrieved candidates to retain only skills applicable to the query and pseudo-plan.The resulting set is the final skill set used to solve the query.
5. Experiment
SkillX is evaluated across long-horizon, user-interactive benchmarks using multiple agent models and memory baselines. Results show that its hierarchical skills improve transfer, task performance, and execution efficiency, while refinement and experience-guided expansion provide additional gains.
- Experimental Settings: Evaluation covers BFCL-v3, AppWorld, and τ 2-bench using Qwen3-32B, Kimi-K2-Instruct-0905, and GLM-4.6 alongside four memory baselines.The baselines are No-memory, A-Mem, AWM, and ExpeL.
- Main Results: SkillX gives Qwen3-32B roughly 10 points across multiple benchmarks while improving execution efficiency; K2 gains mainly on AppWorld.The authors infer that K2 leverages the additional contextual information less effectively on tool-call-intensive benchmarks.
- Main Results: When extraction and execution models are aligned, SkillX consistently outperforms baseline experience representations, including trajectory retrieval and modular workflows.The comparison includes methods marked with ∗ in Table 1.
- Main Results: Experience distilled from GLM-4.6 transfers less effectively through AWM and ExpeL than through SkillX, indicating that representation form is critical for cross-model transfer.The result supports SkillX’s advantage when experience is transferred to weaker base models.
- Analysis: Planning skills reduce execution steps, functional skills contribute most to performance gains, and atomic skills clarify key APIs across diverse models.For weaker models, combining planning with functional skills is especially useful, whereas adding functional and atomic skills can hurt Qwen3-32B through over-imitation.
- Analysis: Iterative refinement improves training and test performance, while experience-guided expansion produces more novel skills and improves performance, especially with scarce training data.Text-only refinement can overfit limited training data, making the number of update rounds important.
- Analysis: Experience-based learning improves execution efficiency, although SkillX does not minimize either execution steps or input tokens individually.The overall performance remains best despite these per-axis trade-offs.
6. Further Analysis
SkillX retains transferability across stronger base models, while ablations show that refinement and expansion can add model- and component-dependent gains. However, τ 2-Bench is excluded from these ablations because its setting makes tool-schema-centered studies less suitable.
- SkillX provides consistent performance gains on stronger base models, whether skills are extracted by those models or constructed using GLM-4.6.
- Ablation results indicate that iterative refinement and skill expansion can offer further improvements depending on the model and component combination.
- τ 2-Bench excludes ablations of skill iteration and expansion because its tool schemas are simple and its training set covers many task patterns.
- For user-centric benchmarks such as τ 2-Bench, whether tool-schema-based experience learning is the most appropriate formulation remains open.
- Qualitative cases show that SkillX helps agents avoid incorrect API sequences, missing prerequisite checks, and failures to handle conversational topic shifts.
7. Related Work
Related work encodes agent experience through cases, workflows, trajectories, and knowledge-base construction. SkillX addresses the unresolved need for a directly reusable skill knowledge base, especially under scarce task data.
- Existing experience-encoding approaches include case-based experiences, workflows, and trajectories stored or reused to guide future agent behavior.
- Experience knowledge-base construction commonly combines static construction from training tasks or curated information with dynamic updating after new task executions.
- Pre-building a strong static ExperienceKB remains necessary in complex agent settings despite the importance of dynamic continual updating.
- SkillX combines task synthesis with skill construction to extend skills under task scarcity and provides a directly reusable skill knowledge base with an automated construction pipeline.
8. Conclusion
SkillX builds a plug-and-play library from hierarchical skills, iterative refinement, and exploratory expansion, and transfers those skills across models. Its direct reuse is less straightforward across different tool ecosystems, while dialogue settings without function calls are outside its primary focus.
- SkillX constructs planning, functional, and atomic skills to support efficient experience transfer across different agent models.
- Its pipeline extracts skills, refines them using execution feedback with merging and filtering, and expands coverage beyond the seed training set.
- SkillX transfers effectively to other models, and the optimized skill library is intended for further community exploration.
- Direct reuse across substantially different domains or tool ecosystems is less straightforward because extracted skills are tied to specific tool schemas.
- The study focuses mainly on tool-using environments rather than more interactive scenarios, particularly dialogue without function calls.
A.1. Benchmark Details
The evaluated benchmarks cover function calling, realistic application interaction, and conversational tool use, while the related methods represent alternative ways to store, derive, and reuse agent experience.
- BFCL-v3 evaluates multi-turn function calling and multi-step reasoning across more than 1,800 test instances and multiple programming languages.
- AppWorld evaluates function-calling agents in realistic application environments spanning nine applications, 457 API endpoints, and about 100 virtual users.
- τ 2-Bench evaluates conversational tool use through multi-turn dialogues requiring context tracking, API selection, and domain-specific business rules.
- A-Mem manages long-term agent knowledge by selectively retaining, retrieving, and revising accumulated experiences according to task objectives and outcomes.
- AWM discovers reusable workflow patterns from completed trajectories and retrieves them through lightweight lexical matching.
- ExpeL derives experiential knowledge by contrasting successful and unsuccessful trajectories for the same task.
A.3. Implementation Details
SkillX extracts and retrieves skills through automated sampling, clustering, vector search, and adaptive relevance filtering.
- Four trajectories are independently sampled per training task, with environment feedback longer than 1500 tokens summarized before skill clustering.
- DBSCAN clusters extracted skills using a cosine similarity threshold of 0.9, then truncates each cluster’s skill set.
- SkillX uses FAISS with an HNSW index and cosine similarity to build a semantic vector store for skill retrieval.
- Retrieval first selects the Top-100 nearest skills, then retains candidates meeting a 0.45 cosine-similarity floor and falling within 0.08 of the query’s best match.
B. Case Study For SkillX
SkillX case studies show reusable skills supporting complex multi-step interactions across AppWorld, BFCL, and τ 2-bench, while accompanying prompts govern extraction and quality control.
- Across AppWorld, BFCL, and τ 2-bench, skill libraries help agents avoid incorrect API sequences, missing prerequisites, and failures to handle topic shifts.The cases report successful complex multi-step execution where the baseline fails, reducing trial and error to first-attempt success.
- In AppWorld, SkillX handles playlist-retrieval pagination and Spotify–Phone integration, whereas the baseline fails on API sequencing and cross-app completion.
- In BFCL, SkillX follows the lock-doors, press-brake, and start-engine prerequisites and authenticates before posting tweets.
- In τ 2-bench, SkillX handles topic shifts, retrieves reservations without reservation numbers, verifies delays, and completes flight-compensation workflows.
- The appendix provides prompts for filtering, feedback summarization, tool-invocation validation, plan extraction, skill merging, and atomic or functional skill extraction.
- Skill-extraction prompts review successful trajectories against the existing library and choose whether to modify, add, or keep skills.They require reusable, domain-specific skills while avoiding duplication and restricting extraction to the specified tool or step.