SKT: Verified Synthetic Data Pipeline Enhances Agent Skill Use
A recent research paper presents SKT, a validated data synthesis framework aimed at enhancing language-model agents' capabilities in recognizing, utilizing, and coordinating skills. This framework generates skill-based tasks and executable paths from extensive databases of agent skills, employing both rule-based and agent-based verification methods along with feedback-driven corrections to ensure only successful trajectories that effectively engage all necessary skills are retained. By leveraging 2,000 public skills, SKT generated 4,000 task packages and 27,164 verified trajectories. Additionally, the same framework facilitated the creation of SkillEval, a separate executable benchmark for assessing skill application. The paper can be found on arXiv with the identifier 2608.02287.
Key facts
- SKT is a verified data synthesis pipeline for improving skill use in language-model agents.
- It constructs skill-grounded tasks and executable trajectories from agent skills.
- It selects single-skill and multi-skill configurations.
- It uses rule-based and agent-based verification with feedback-guided repair.
- It retains only successful trajectories that substantially use every required skill.
- Using 2,000 public skills, SKT produced 4,000 task packages and 27,164 verified trajectories.
- SkillEval is a held-out executable benchmark for evaluating skill use.
- The paper is available on arXiv under identifier 2608.02287.
Entities
Institutions
- arXiv