RING: Retrieval-Internalized Generation for Continual Large-Scale Knowledge Injection

2026-08-02

AI Summary

This work addresses the latency and engineering overhead introduced by retrieval-augmented generation (RAG) during inference by proposing the RING paradigm, which internalizes external knowledge into model parameters through a mixture-of-experts memory architecture and, for the first time, enables end-to-end learning of parameterized retrieval strategies. Key innovations include a Dual Causal Attention mechanism, a three-stage training pipeline incorporating hierarchical-reward reinforcement learning, and News-2025—a novel evaluation benchmark designed to prevent test-set leakage for assessing generalization to unseen knowledge. Experimental results demonstrate that RING matches or surpasses conventional RAG and existing parameterized knowledge injection methods in both accuracy and efficiency, thereby validating the effectiveness and scalability of internalized retrieval.

Abstract

Retrieval-augmented generation (RAG) improves factuality but adds latency and engineering overhead at serving time. We propose RING (Retrieval-Internalized Generation), a holistic paradigm spanning both architecture and training that injects large-scale external knowledge into a \textit{Mixture-of-Memory Experts} and learns parametric search over this internal memory via reinforcement learning, removing the external retriever entirely. Training proceeds in three stages: continued pre-training injects new corpora into a Knowledge Expert via our novel \textit{Dual Causal Attention}; supervised fine-tuning teaches a ``search-then-answer'' pattern; and reinforcement learning with hierarchical rewards optimizes the routing-and-search policy over the parametric memory. Unlike prior parametric injection methods that pair internal memory with a fixed or rule-based retriever, RING {learns} its retrieval policy directly from task signals. We further frame RING theoretically as a search-free approximation to the classical RAG objective. To evaluate large-scale injection of genuinely {new} knowledge without test-time leakage, we further construct News-2025, a benchmark built from news strictly post-dating the base LLM's pretraining cutoff. RING matches or surpasses both search-based RAG and parametric injection baselines in accuracy and efficiency.

Problem

• Retrieval-Augmented Generation
• Knowledge Injection
• Parametric Memory
• Continual Learning
• Large Language Models

Innovation

• Retrieval-Internalized Generation
• Mixture-of-Memory Experts
• Parametric Search
• Dual Causal Attention
• Reinforcement Learning