🤖 AI Summary
In retrieval-augmented generation (RAG), conflicts between large language models’ (LLMs’) internal knowledge and externally retrieved information induce unreliable responses, yet the underlying uncertainty dynamics remain poorly understood. Method: This paper models such conflicts through an information-theoretic lens, revealing—for the first time—the anomalous drop in LLM confidence under knowledge ambiguity. Building on this insight, we propose Swin-VIB: a cascaded framework grounded in the variational information bottleneck (VIB) that adaptively filters and injects retrieved content while jointly optimizing LLM preference modeling and response generation. Contribution/Results: Evaluated across single-choice QA, open-ended QA, and standard RAG benchmarks, Swin-VIB significantly improves response reliability—achieving ≥7.54% absolute accuracy gain over the strongest baseline in single-choice QA. Our work establishes a novel, conflict-aware paradigm for trustworthy RAG.
📝 Abstract
The proliferation of large language models (LLMs) has significantly advanced information retrieval systems, particularly in response generation (RG). Unfortunately, LLMs often face knowledge conflicts between internal memory and retrievaled external information, arising from misinformation, biases, or outdated knowledge. These conflicts undermine response reliability and introduce uncertainty in decision-making. In this work, we analyze how LLMs navigate knowledge conflicts from an information-theoretic perspective and reveal that when conflicting and supplementary information exhibit significant differences, LLMs confidently resolve their preferences. However, when the distinction is ambiguous, LLMs experience heightened uncertainty. Based on this insight, we propose Swin-VIB, a novel framework that integrates a pipeline of variational information bottleneck models into adaptive augmentation of retrieved information and guiding LLM preference in response generation. Extensive experiments on single-choice, open-ended question-answering (QA), and retrieval augmented generation (RAG) validate our theoretical findings and demonstrate the efficacy of Swin-VIB. Notably, our method improves single-choice task accuracy by at least 7.54% over competitive baselines.