SKILL-KD introduces contrastive skill distillation for LLM agents. It compares a student’s failed trajectory with a teacher’s trajectory on the same task, converts the actionable discrepancy into a textual skill patch, and validates the patch by rerunning the student. If failure persists, the patch is refined iteratively. The framework also maintains trace-linked edit histories and applies Drift-Aware Skill Consolidation to decide whether a patch should add, modify, delete, or skip a rule. The abstract reports consistent gains over fixed-model adaptation baselines across five agent benchmarks and two student settings.
No heat snapshots are available in the last 24 hours.