Skill-RAG: Failure-State-Aware Retrieval Augmentation via Hidden-State Probing and Skill Routing
Researchers have introduced Skill-RAG, a novel framework designed to improve retrieval-augmented generation (RAG) systems. This new approach addresses persistent retrieval failures by diagnosing the root cause of query-evidence misalignment rather than simply retrying. Skill-RAG employs a hidden-state prober and a skill router that selects from four distinct retrieval skills to correct these misalignments before generating a response. Experiments demonstrate significant accuracy improvements, particularly on challenging and out-of-distribution datasets. AI
IMPACT Enhances LLM knowledge grounding by addressing specific retrieval failure modes, potentially improving accuracy on complex queries.