DynaKRAG: A Unified Framework for Learnable Evidence Control in Multi-Hop Retrieval-Augmented Generation
The episode discusses DynaKRAG, a unified framework for learnable evidence control in multi-hop retrieval-augmented generation. The hosts explain how DynaKRAG frames evidence acquisition as a state-conditioned control problem, allowing AI systems to learn the best sequence of actions dynamically based on their current evidence state. The paper shows strong performance and efficiency gains.