A skill library built by GLM-4.6 boosts GPT-4.1's task success by 10 points

Notes on SkillX: Automatically Constructing Skill Knowledge Bases for Agents (arXiv:2604.04804) — Chenxi Wang, Zhuoyun Yu, Xinghong Xie, Wuguannan Yao, Runnan Fang, Shuofei Qiao, Kexin Cao, Guozhou Zheng, Xiang Qi, Peng Zhang, Shumin Deng · April 2026

Note published July 25, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

SkillX turns an agent's raw trajectories into a three-tier skill knowledge base: strategic plans that sequence a task end to end, functional skills that package a tool-based subroutine, and atomic skills that encode a single tool call's constraints. Two automated passes then work the library. An iterative skill refinement loop merges near-duplicate skills and screens survivors through a general quality check plus a tool-schema check. An exploratory skill expansion step targets tools the agent under-uses or keeps failing on, generating and validating new skills to close the gap. Everything runs off a single backbone, GLM-4.6, with each training task rolled out four times, on small seed sets of 50 to 90 training tasks per benchmark, and no human editing in the loop.

Dropped into other agents, the resulting library lifts task success and trims execution steps, and the smallest models gain the most: Qwen3-32B's BFCL-v3 average success rises from 53.67% to 63.67%, beating prior extracted-experience baselines A-MEM, Agent Workflow Memory, and ExpeL on the same setup. Transfer isn't limited to same-family models either — GPT-4.1 running on a library built entirely by GLM-4.6 gains roughly 10 points on BFCL-v3 — but the gains shrink as the receiving model gets stronger: GLM-4.6's own AppWorld gain is about 8% relative, versus roughly 27% for Qwen3-32B. More refinement isn't automatically better: on GLM-4.6's own BFCL-v3 score, a third refinement pass edges the average down slightly from a second pass's 79.50 to 78.83.