DynaKRAG: A Unified Framework for Learnable Evidence Control in Multi-Hop Retrieval-Augmented Generation
A framework for learnable evidence control in multi-hop retrieval-augmented generation.
The paper introduces DynaKRAG, a unified framework that formulates multi-hop evidence acquisition as state-conditioned control over atomic evidence operations. It uses a learned controller to select the next operation and updates the evidence state accordingly. The authors evaluate DynaKRAG on several benchmarks and demonstrate its effectiveness in coordinating retrieval, diagnosis, and gap-directed acquisition.
Based on: DynaKRAG: A Unified Framework for Learnable Evidence Control in Multi-Hop Retrieval-Augmented Generation · arXiv