SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks
Published in arXiv preprint arXiv:2602.12670, 2026
Xiangyi Li, Yimin Liu, Wenbo Chen, et al. (incl. Changkun Guan). "SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks." arXiv preprint arXiv:2602.12670, 2026.
SkillsBench is a large community benchmark measuring how structured procedural knowledge packages (“skills”) improve the performance of large language model agents. The benchmark spans 87 tasks across 8 domains, and evaluates when and how much curated skills help agents complete diverse real-world tasks.
I contributed as one of the benchmark’s many co-authors.