SKT: Skill-Use Training at Scale via Verified Synthetic Data Generation
2026-08-03 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors focus on helping language-model agents better use 'skills,' which are like reusable actions or procedures. They created SKT, a system that makes sure tasks involving these skills are correct and that the agents use them properly. Using SKT, they made lots of verified tasks and examples, which helped train models to improve skill use. They also built SkillEval, a test set for checking how well models use skills. Their experiments show that training with this verified data helps models perform better and works across different setups and skill sets.
language-model agentsagent skillsdata synthesistask verificationsupervised fine-tuningskill-use performancebenchmarkingtrajectorymulti-skill configurationsskill evaluation
Authors
Zelin Tan, Yiqun Zhang, Hao Li, Zhiyao Cui, Hejia Geng, Shao Zhang, Hangfan Zhang, Yang Chen, Xiaosong Wang, Lilong Wang, Zhenfei Yin, Shuyue Hu, Chen Zhang, Lei Bai
Abstract
Agent skills have become an important mechanism for equipping language-model agents with reusable procedural knowledge. However, providing skills alone does not guarantee that current models can effectively identify, apply, and coordinate them. To improve skill-use capabilities, we introduce SKT, a verified data synthesis pipeline that constructs skill-grounded tasks and executable trajectories from large collections of agent skills. SKT selects suitable single-skill and multi-skill configurations, synthesizes tasks through rule-based and agent-based verification with feedback-guided repair, and retains only successful trajectories that substantially use every required skill. Using 2,000 public skills, SKT produces 4,000 task packages and 27,164 verified trajectories. Based on the same pipeline and a disjoint test pool, we further construct SkillEval, a held-out executable benchmark for evaluating skill use. Experiments across diverse models, benchmarks, and agent harnesses show that supervised fine-tuning on SKT-generated trajectories consistently improves skill-use performance. Verification ablations, cross-harness evaluation, and scaling experiments further demonstrate that these gains depend on high-quality supervision, extend beyond a single agent interface, and increase with broader skill coverage. Together, these results establish verified data synthesis as an effective and scalable approach for skill-use training.