Dynamic and Parametric Retrieval-Augmented Generation (arxiv.org)

Researchers on arXiv presented a tutorial-style paper framing two fast-growing directions in Retrieval-Augmented Generation (RAG): Dynamic RAG and Parametric RAG. Instead of the traditional static retrieve-then-generate pipeline and brittle in-context knowledge injection, Dynamic RAG lets an LLM decide during generation when and what to fetch, enabling adaptive, multihop information access that tracks the model’s evolving needs. Parametric RAG rethinks how retrieved knowledge is integrated — moving beyond concatenating documents in the input to injecting or baking knowledge into model parameters or parameterized conditioning mechanisms for greater efficiency and lasting effect. The tutorial synthesizes recent advances, theoretical foundations, and practical insights, highlighting how these shifts address key limitations of static RAG: improved handling of complex, multi-step queries; reduced dependence on context-window size; and more efficient reuse of retrieved facts across interactions. Technically, Dynamic RAG spans strategies for online retrieval control and belief-state-aware querying, while Parametric RAG covers approaches that embed external knowledge at the parameter level (e.g., via fine-tuning, adapters, or parameter-efficient tuning) or hybrid schemes that combine stateful parameter updates with on-demand retrieval. Together these directions promise more robust, scalable knowledge grounding for LLMs and chart concrete research paths for better reasoning, efficiency, and long-term knowledge maintenance.