SKT uses verified synthetic trajectories to train agents to use skills better

Shanghai-AI-Laboratory · hf · 2026-08-04

SKT is a verified synthetic data pipeline for training language-model agents to use skills more effectively.
It builds skill-grounded tasks and executable trajectories from a large collection of agent skills, using rule-based and agent-based verification plus feedback-guided repair.
From 2,000 public skills, the system produced 4,000 task packages and 27,164 verified trajectories.
The authors also release SkillEval, a held-out executable benchmark for skill-use evaluation.
Across models, benchmarks, and agent harnesses, supervised fine-tuning on SKT trajectories consistently improves skill-use performance, and the gains scale with higher-quality supervision and broader skill coverage.

More from coding & agent