SKILL-KD: Contrastive Skill Distillation for LLM Agents

2026-07-30

AI Summary

Existing skill distillation methods struggle to effectively transfer teacher expertise when student agents’ failure trajectories lack informative behavioral cues. This work proposes a contrastive skill distillation framework that compares teacher and student trajectories on the same task to generate editable, textual skill patches. The approach incorporates trajectory replay evaluation, drift-aware skill consolidation, and edit history management to enable iterative refinement and stable knowledge transfer. By explicitly modeling skills as knowledge mediators, the method significantly outperforms fixed-model fine-tuning baselines across five agent benchmarks and two student settings, substantially improving task performance even for frozen student agents.

Abstract

Skill-based prompting has become a practical mechanism for improving large language model (LLM) agents, yet existing skill acquisition methods often treat skills as experience summaries, memory entries, or direct summaries of successful demonstrations. This creates a mismatch for weaker student agents: when a student fails because it lacks task knowledge or operational strategy, its failed trajectory may not contain enough evidence to infer the missing behavior, while the teacher trajectory may be too implicit to be internalized as reusable guidance. We propose SKILL-KD, a contrastive skill distillation framework that treats skills as an explicit distillation medium between agents of different capabilities. Given a student failure and the teacher trajectory on the same task, SKILL-KD distills their actionable discrepancy into a textual skill patch, evaluates the patch by re-running the student, and iteratively refines the patch when the student still fails. To prevent repeated local updates from causing skill drift, SKILL-KD further maintains trace-linked edit histories and performs Drift-Aware Skill Consolidation, deciding whether each patch should add a new rule, delete or modify an existing rule, or be skipped. Across five agent benchmarks and two student settings, SKILL-KD consistently improves frozen student agents over fixed-model adaptation baselines.

Problem

• skill distillation
• LLM agents
• trajectory mismatch
• skill representation
• student-teacher learning

Innovation

• contrastive skill distillation
• skill patch
• drift-aware consolidation
• LLM agents
• trajectory comparison