Tag
DynaKRAG introduces a unified framework for multi-hop retrieval-augmented generation that learns a state-conditioned policy to select evidence operations, outperforming baselines on HotpotQA, 2Wiki, and MuSiQue.
This paper studies retrieval-augmented generation as an in-context optimization process, showing that linear self-attention can implement gradient descent on a unified RAG objective. It proposes a lightweight method for frozen RAG LLMs that predicts context-conditioned updates, improving performance across multiple QA benchmarks.