IDEAL-RAG: Instruction-driven Dual-standpoint Elicitation and Alignment Linking for Retrieval Augmented Generation

ICLR 2026 Conference Submission17492 Authors

Keywords: Large Language Models, Retrieval-Augmented Generation, In-Context Learning, Hallucination Reduction, Prompt Engineering

TL;DR: IDEAL-RAG strengthens RAG by explicitly balancing what an LLM knows with what it retrieves, enabling more reliable reasoning under noisy contexts.

Abstract: Retrieval-augmented generation (RAG) equips large language models (LLMs) with external evidence, yet even minor retrieval noise or adversarial edits can override parametric knowledge and trigger hallucinations. Prior work mainly denoises contexts; far fewer methods explicitly balance internal memory with retrieved text. We present IDEAL-RAG, a three-stage, instruction-driven framework that (i) elicits latent knowledge, (ii) forms independent standpoints from internal memory and retrieved passages, and (iii) cross-checks them to produce a traceable rationale—without modifying retrievers or requiring additional labels. Across standard open-domain QA settings, IDEAL-RAG matches strong baselines on clean retrieval and, under adversarial counterfactual contexts, improves exact-match by up to +22.8\% while roughly halving accuracy loss. Mechanistic analyses explain the gains: a Counterfactual Sensitivity Score (CSS) shows smaller confidence swings, and a layer-wise Parametric Knowledge Score (PKS) reveals steadier reliance on internal memory; ablations further identify parametric-knowledge elicitation as the primary driver of robustness. These results indicate that deliberate negotiation between what an LLM knows and what it reads yields more dependable RAG systems.

Primary Area: generative models

Submission Number: 17492