DyPRAG: Dynamic Parametric Retrieval-Augmented Generation for On-the-Fly Knowledge Adaptation

ACL ARR 2026 January Submission715 Authors

Keywords: Retrieval-Augmented Generation, Inference-time Knowledge Adaptation, Parametric Knowledge in LLMs

Abstract: Once Large Language Models (LLMs) complete their training, the intrinsic parametric knowledge encoded within the model becomes fixed. Retrieval-Augmented Generation (RAG) alleviates this by supplying external documents as context, but a fundamental tension remains: the model's parameters are unchanged, often leading to conflicts where the model's outdated internal parameters struggle to synergize with the fresh retrieved information. In this paper, we propose Dynamic Parametric RAG (DyPRAG), a novel framework that leverages a lightweight hypernetwork, termed parameter translator, to efficiently convert symbolic documents into parametric knowledge at inference-time. Specifically, the parameter translator maps documents directly into Low-Rank Adaptation (LoRA) weights for the Feed-Forward Networks (FFNs) of the LLM, enabling on-the-fly knowledge adaptation. Extensive experiments across diverse datasets demonstrate the superior effectiveness and generalization capability of DyPRAG. Crucially, our analysis confirms that DyPRAG harmonizes the model's internal and external knowledge sources, leading to a measurable reduction in knowledge conflicts and a more effective synthesis of information. Our code is available at https://anonymous.4open.science/r/DyPRAG-715.

Paper Type: Long

Research Area: Retrieval-Augmented Language Models

Research Area Keywords: retrieval-augmented generation,knowledge-augmented methods

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches low compute settings-efficiency, Publicly available software and/or pre-trained models

Languages Studied: English

Submission Number: 715