Skill-SD: Skill-Conditioned Self-Distillation Turns Agent Trajectories into Compact Natural Language Skills for Training-Free Multi-Turn Improvement

arXiv·medium signalXBlueskyLinkedInCopy link

arXiv:2604.10674 (April 14, 2026) introduces Skill-SD, a framework where an LLM agent's completed trajectories are automatically summarized into compact natural language skills describing successful behaviors, mistakes, and workflows. These skills then serve as dynamic training-only supervision for the agent's next attempts — no gradient updates required. Practical implication for builders: you can implement a version of this pattern with your existing agents by logging trajectories, summarizing them into skill documents, and injecting relevant skills into future system prompts based on task similarity.