SkillRL: Evolving LLM Agents via Distilled Skills скачать в хорошем качестве

SkillRL: Evolving LLM Agents via Distilled Skills

In this AI Research Roundup episode, Alex discusses the paper: 'SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning' SkillRL addresses the inability of current LLM agents to learn efficiently from past experiences due to noisy raw data. The framework introduces experience-based skill distillation, converting trajectories into strategic demonstrations and failure lessons. This process achieves a significant 10-20x token compression compared to storing raw interaction histories. Distilled knowledge is then organized into a Hierarchical SkillBank that uses adaptive retrieval to improve generalization across tasks. Ultimately, SkillRL enables agents to evolve by recursively refining their behavioral patterns from both successes and failures. Paper URL: https://arxiv.org/abs/2602.08234 #AI #MachineLearning #DeepLearning #LLMAgents #ReinforcementLearning #SkillLearning #KnowledgeDistillation Resources: GitHub: https://github.com/aiming-lab/SkillRL