Skip to content
Back to search

SkillX

SkillX: Automatically Constructing Skill Knowledge Bases for Agents

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/zjunlp--skillx/badge.svg)](/listings/zjunlp--skillx)

HTML

<a href="/listings/zjunlp--skillx"><img src="/api/listings/zjunlp--skillx/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

README.md · README · 265 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. <div align="center">
  2. <h1 align="center"> 👉 SkillX 👈 </h1>
  3. <b>SkillX: Automatically Constructing Skill Knowledge Bases for Agents</b>
  4. [![Awesome](https://awesome.re/badge.svg)](https://github.com/zjunlp/SKillX)
  5. [![On StackMap](https://img.shields.io/endpoint?url=https%3A%2F%2Fstackmap.shipwithai.xyz%2Fapi%2Fbadge%2Fskillx.json)](https://stackmap.shipwithai.xyz/repos/zjunlp/skillx?utm_source=badge)
  6. [![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](https://opensource.org/licenses/MIT)
  7. ![](https://img.shields.io/github/last-commit/zjunlp/SKillX?color=green)
  8. <!-- <p align="center">
  9. <a href="https://arxiv.org/abs/2502.15589">📄arXiv</a> •
  10. <a href="https://x.com/zxlzr/status/1894729164609208338">𝕏 Blog</a> •
  11. <a href="https://huggingface.co/collections/zjunlp/SKillX-67f9faaaa518f2e00b17386b">🤗 Huggingface</a>
  12. </p> -->
  13. </div>
  14. ## Table of Contents
  15. - 👀[Overview](#overview)
  16. - 🔧[Installation](#installation)
  17. - 🏃[Quick Start](#quick-start)
  18. - 🎁[Acknowledgement](#acknowledgement)
  19. - 🚩[Citation](#citation)
  20. ## 📖 Overview
  21. **SkillX** is a fully automated framework that constructs a **reusable, plug-and-play skill knowledge base** for LLM agents from experience.
  22. Instead of storing raw trajectories, workflows, or loosely structured reflections, SkillX distills agent experience into a **three-level skill hierarchy**:
  23. - **Planning Skills** for high-level task organization
  24. - **Functional Skills** for reusable tool-based subroutines
  25. - **Atomic Skills** for execution-oriented tool usage patterns
  26. Built with a strong backbone agent, SkillX produces a transferable skill library that can be directly plugged into weaker base agents and new environments. Across challenging long-horizon, user-interactive benchmarks such as **AppWorld**, **BFCL-v3**, and **τ2-Bench**, SkillX consistently improves both **task success** and **execution efficiency**.
  27. <div align="center" style="width: 100%;">
  28. <img src="assets/overview.png" alt="Case GIF" style="width: 100%; max-width: 100%; height: auto;">
  29. </div>
  30. ---
  31. ## Data Formats
  32. ### Trajectory Input (JSONL)
  33. SkillX expects trajectories in the following schema:
  34. ```json
  35. {
  36. "trajectory_id": "traj_001",
  37. "task_id": "task_001",
  38. "user_task": "How many songs are in my Spotify library?",
  39. "task_history": [
  40. {"role": "system", "content": "You are a helpful assistant..."},
  41. {"role": "assistant", "content": "I'll help you count..."},
  42. {"role": "user", "content": "Output:\n```\n{\"songs\": 150}\n```"}
  43. ],
  44. "reward": 1.0,
  45. "metadata": {}
  46. }
  47. ```
  48. ## 🤖 Key Features
  49. ### Hierarchical Multi-Level Skill Design
  50. SkillX transforms raw trajectories into a structured three-tier skill space:
  51. - **Planning Skills** capture high-level decomposition and ordering
  52. - **Functional Skills** represent reusable multi-step tool subroutines
  53. - **Atomic Skills** encode practical tool usage constraints and patterns
  54. ### Fully Automated Skill KB Construction
  55. SkillX provides an end-to-end automated pipeline that:
  56. - rolls out agents on training tasks,
  57. - extracts reusable skills from successful trajectories,
  58. - consolidates and filters low-quality skills,
  59. - and builds a reusable **plug-and-play skill knowledge base**.
  60. ### Iterative Skill Refinement
  61. SkillX continuously improves the skill library through:
  62. - **skill merging** for consolidating redundant behaviors,
  63. - **quality filtering** for removing brittle or hallucinated skills,
  64. - and **iterative updates** that add, modify, or keep skills based on execution feedback.
  65. ### Exploratory Skill Expansion
  66. Beyond seed demonstrations, SkillX proactively discovers new skills by:
  67. - identifying under-used and failure-prone tools,
  68. - guiding environment exploration,
  69. - synthesizing new tasks from exploratory trajectories,
  70. - and expanding skill coverage beyond the original training distribution.
  71. ### Plug-and-Play Transfer Across Agents
  72. The resulting skill library can be directly injected into different base agents, enabling **strong-to-weak transfer** without retraining the underlying model.
  73. ### Better Performance and Efficiency
  74. SkillX consistently improves:
  75. - **task success rate** on challenging benchmarks,
  76. - **execution efficiency** by reducing unnecessary exploration and tool misuse,
  77. - and **generalization** through structured, reusable experience abstraction.
  78. ---
  79. ## 📊 Highlights
  80. - **~10% absolute improvement** for weaker base agents on multiple benchmarks
  81. - Consistent gains on **AppWorld**, **BFCL-v3**, and **τ2-Bench**
  82. - Stronger transferability than trajectory-based, workflow-based, and memory-based baselines
  83. - Improved **execution efficiency** with fewer redundant steps
  84. - Effective even when the skill library is built by a stronger model and used by weaker ones
  85. ---
  86. ## 🧠 Why SkillX?
  87. Existing experience-learning methods often suffer from:
  88. - **Isolated learning**: agents repeatedly rediscover similar behaviors
  89. - **Weak transferability**: raw trajectories and reflections often do not generalize well
  90. - **Capability bottlenecks**: self-extracted experience is limited by the agent’s own strength
  91. SkillX addresses these issues by building a **structured skill knowledge base** that is:
  92. - **reusable across tasks**
  93. - **transferable across agents**
  94. - **lightweight to retrieve**
  95. - **easy to inject into prompts**
  96. - **more robust than long-context progressive skill formats**
  97. ---
  98. ## 🏗️ Method Overview
  99. SkillX consists of three core components:
  100. ### 1. Multi-Level Skills Extraction
  101. From successful trajectories, SkillX automatically extracts:
  102. - **Planning skills**: concise, reusable task plans
  103. - **Functional skills**: reusable tool-composition procedures
  104. - **Atomic skills**: tool-specific usage guidance, constraints, and failure notes
  105. ### 2. Iterative Skills Refinement
  106. SkillX improves library quality through:
  107. - **Skills Merge**: cluster and consolidate similar skills
  108. - **Skills Filter**: remove non-portable, hallucinated, or invalid skills
  109. - **Skills Update**: add, modify, or keep skills across iterations
  110. ### 3. Exploratory Skills Expansion
  111. SkillX expands beyond observed demonstrations by:
  112. - guiding exploration toward under-covered tools and failure modes,
  113. - synthesizing new tasks from exploration,
  114. - and rerunning extraction + refinement to grow the skill library.
  115. ---
  116. ## 📈 Main Results
  117. SkillX improves agentic performance across multiple LLM backbones and benchmarks.
  118. ### Representative gains
  119. - On **Qwen3-32B**, SkillX brings **around 10-point improvements** on several benchmarks
  120. - On **Kimi-K2-Instruct-0905**, SkillX yields clear gains especially on **AppWorld**
  121. - On **GLM-4.6**, SkillX still improves performance and execution efficiency despite the model already being strong
  122. ### Benchmarks
  123. - **AppWorld**
  124. - **BFCL-v3**
  125. - **τ2-Bench**
  126. ### Key takeaway
  127. SkillX outperforms strong experience-learning baselines such as:
  128. - **A-Mem**
  129. - **AWM**
  130. - **ExpeL**
  131. - **No-memory**
  132. This shows that **how experience is represented** matters as much as, or more than, where it comes from.
  133. ---
  134. ## 🔍 What Makes SkillX Different?
  135. Compared with prior experience formats:
  136. - **Raw trajectories** are verbose and difficult to transfer
  137. - **Insights/reflections** are often too abstract
  138. - **Workflows** may miss low-level tool constraints
  139. - **Claude-style skills** rely on long-context progressive disclosure and complex environment support
  140. In contrast, SkillX offers:
  141. - **hierarchical, itemized, reusable skills**
  142. - **one-time prompt injection**
  143. - **lightweight retrieval**
  144. - **strong transfer across agents and environments**
  145. ---
  146. ## 🚀 Use Cases
  147. SkillX is especially useful for:
  148. - **tool-using LLM agents**
  149. - **long-horizon task execution**
  150. - **interactive application environments**
  151. - **cross-agent knowledge transfer**
  152. - **building reusable agent skill libraries from experience**
  153. ---
  154. ## 🧪 Benchmarks Used
  155. ### AppWorld
  156. A realistic ecosystem of apps and APIs for long-horizon agent execution.
  157. ### BFCL-v3
  158. A challenging benchmark for multi-turn function calling and tool use.
  159. ### τ2-Bench
  160. A user-interactive benchmark focused on conversational tool-using agents.
  161. ---
  162. ## 📦 Planned Release
  163. We will publicly release:
  164. - the **SkillX codebase**
  165. - the **automatically constructed skill knowledge base**
  166. - and supporting resources for skill extraction, refinement, and retrieval
  167. ---
  168. ## 🙏 Acknowledgement
  169. We deeply appreciate the invaluable effort contributed by our dedicated team of developers, supportive users, and esteemed industry partners: [Ant Digital Technologies, Ant Group](https://intl.antdigital.com/en).
  170. This repository builds upon code from [ReMe](https://github.com/agentscope-ai/ReMe) and [AgentEvolver](https://github.com/modelscope/AgentEvolver). The baseline implementations are adapted from [A-MEM](https://github.com/agiresearch/a-mem), [AWM](https://github.com/zorazrw/agent-workflow-memory), and [Expel](https://github.com/LeapLabTHU/ExpeL). We sincerely thank all contributors for their outstanding work!
  171. ## 📚 Citation
  172. If you find this work helpful, please consider citing:
  173. ```bibtex
  174. @article{wang2026skillx,
  175. author = {Chenxi Wang and
  176. Zhuoyun Yu and
  177. Xin Xie and
  178. Wuguannan Yao and
  179. Runnan Fang and
  180. Shuofei Qiao and
  181. Kexin Cao and
  182. Guozhou Zheng and
  183. Xiang Qi and
  184. Peng Zhang and
  185. Shumin Deng},
  186. title = {SkillX: Automatically Constructing Skill Knowledge Bases for Agents},
  187. year = {2026},
  188. eprint = {2604.04804},
  189. archivePrefix = {arXiv},
  190. primaryClass = {cs.CL},
  191. url = {https://arxiv.org/abs/2604.04804}
  192. }
  193. ```
  194. ---

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.