Imagen de Portada

SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks

Fuente: arXiv

Guardado en:

Detalles Bibliográficos
Autores principales:	Li, Xiangyi, Chen, Wenbo, Liu, Yimin, Zheng, Shenghan, Chen, Xiaokun, He, Yifeng, Li, Yubo, You, Bingran, Shen, Haotian, Sun, Jiankai, Wang, Shuyi, Li, Binxu, Zeng, Qunhong, Wang, Di, Zhao, Xuandong, Wang, Yuanli, Chaim, Roey Ben, Di, Zonglin, Gao, Yipeng, He, Junwei, He, Yizhuo, Jing, Liqiang, Kong, Luyang, Lan, Xin, Li, Jiachen, Li, Songlin, Li, Yijiang, Lin, Yueqian, Liu, Xinyi, Liu, Xuanqing, Lyu, Haoran, Ma, Ze, Wang, Bowei, Wang, Runhui, Wang, Tianyu, Ye, Wengao, Zhang, Yue, Xing, Hanwen, Xue, Yiqi, Dillmann, Steven, Lee, Han-chung
Formato:	Preprint
Publicado:	2026
Materias:	Artificial Intelligence
Acceso en línea:	Acceder al recurso
Etiquetas:	Agregar Etiqueta Sin Etiquetas, Sea el primero en etiquetar este registro!

_version_	1866914389794226176
author	Li, Xiangyi Chen, Wenbo Liu, Yimin Zheng, Shenghan Chen, Xiaokun He, Yifeng Li, Yubo You, Bingran Shen, Haotian Sun, Jiankai Wang, Shuyi Li, Binxu Zeng, Qunhong Wang, Di Zhao, Xuandong Wang, Yuanli Chaim, Roey Ben Di, Zonglin Gao, Yipeng He, Junwei He, Yizhuo Jing, Liqiang Kong, Luyang Lan, Xin Li, Jiachen Li, Songlin Li, Yijiang Lin, Yueqian Liu, Xinyi Liu, Xuanqing Lyu, Haoran Ma, Ze Wang, Bowei Wang, Runhui Wang, Tianyu Ye, Wengao Zhang, Yue Xing, Hanwen Xue, Yiqi Dillmann, Steven Lee, Han-chung
author_facet	Li, Xiangyi Chen, Wenbo Liu, Yimin Zheng, Shenghan Chen, Xiaokun He, Yifeng Li, Yubo You, Bingran Shen, Haotian Sun, Jiankai Wang, Shuyi Li, Binxu Zeng, Qunhong Wang, Di Zhao, Xuandong Wang, Yuanli Chaim, Roey Ben Di, Zonglin Gao, Yipeng He, Junwei He, Yizhuo Jing, Liqiang Kong, Luyang Lan, Xin Li, Jiachen Li, Songlin Li, Yijiang Lin, Yueqian Liu, Xinyi Liu, Xuanqing Lyu, Haoran Ma, Ze Wang, Bowei Wang, Runhui Wang, Tianyu Ye, Wengao Zhang, Yue Xing, Hanwen Xue, Yiqi Dillmann, Steven Lee, Han-chung
contents	Agent Skills are structured packages of procedural knowledge that augment LLM agents at inference time. Despite rapid adoption, there is no standard way to measure whether they actually help. We present SkillsBench, a benchmark of 86 tasks across 11 domains paired with curated Skills and deterministic verifiers. Each task is evaluated under three conditions: no Skills, curated Skills, and self-generated Skills. We test 7 agent-model configurations over 7,308 trajectories. Curated Skills raise average pass rate by 16.2 percentage points(pp), but effects vary widely by domain (+4.5pp for Software Engineering to +51.9pp for Healthcare) and 16 of 84 tasks show negative deltas. Self-generated Skills provide no benefit on average, showing that models cannot reliably author the procedural knowledge they benefit from consuming. Focused Skills with 2--3 modules outperform comprehensive documentation, and smaller models with Skills can match larger models without them.
format	Preprint
id	arxiv_https___arxiv_org_abs_2602_12670
institution	arXiv
publishDate	2026
record_format	arxiv
spellingShingle	SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks Li, Xiangyi Chen, Wenbo Liu, Yimin Zheng, Shenghan Chen, Xiaokun He, Yifeng Li, Yubo You, Bingran Shen, Haotian Sun, Jiankai Wang, Shuyi Li, Binxu Zeng, Qunhong Wang, Di Zhao, Xuandong Wang, Yuanli Chaim, Roey Ben Di, Zonglin Gao, Yipeng He, Junwei He, Yizhuo Jing, Liqiang Kong, Luyang Lan, Xin Li, Jiachen Li, Songlin Li, Yijiang Lin, Yueqian Liu, Xinyi Liu, Xuanqing Lyu, Haoran Ma, Ze Wang, Bowei Wang, Runhui Wang, Tianyu Ye, Wengao Zhang, Yue Xing, Hanwen Xue, Yiqi Dillmann, Steven Lee, Han-chung Artificial Intelligence Agent Skills are structured packages of procedural knowledge that augment LLM agents at inference time. Despite rapid adoption, there is no standard way to measure whether they actually help. We present SkillsBench, a benchmark of 86 tasks across 11 domains paired with curated Skills and deterministic verifiers. Each task is evaluated under three conditions: no Skills, curated Skills, and self-generated Skills. We test 7 agent-model configurations over 7,308 trajectories. Curated Skills raise average pass rate by 16.2 percentage points(pp), but effects vary widely by domain (+4.5pp for Software Engineering to +51.9pp for Healthcare) and 16 of 84 tasks show negative deltas. Self-generated Skills provide no benefit on average, showing that models cannot reliably author the procedural knowledge they benefit from consuming. Focused Skills with 2--3 modules outperform comprehensive documentation, and smaller models with Skills can match larger models without them.
title	SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks
topic	Artificial Intelligence
url	https://arxiv.org/abs/2602.12670