When Context Misleads: Intent-Guided Decoding for Robust Retrieval-Augmented Generation

2026-08-17

AI Summary

This study addresses generation biases in Retrieval-Augmented Generation (RAG) arising from misleading retrieval and static trust strategies by proposing an intent-guided decoding framework. The method employs a dynamic arbitration mechanism to adaptively balance retrieved context with parametric memory, while integrating answer-level filtering and token-level correction to optimize generation trajectories for both factuality and faithfulness. Experimental results demonstrate that this framework achieves performance gains of up to 65.4 percentage points on factual conflict benchmarks while maintaining or enhancing strict context adherence. Consequently, it significantly mitigates knowledge conflicts and over-reliance issues inherent in current RAG systems, offering a robust solution for reconciling external evidence with internal model knowledge during decoding.

Abstract

Retrieval-augmented generation (RAG) improves large language models by grounding generation in external evidence, but it also introduces a source trust problem: retrieved context may be useful, irrelevant, or even misleading. Existing RAG systems often apply a fixed trust policy toward retrieved evidence, which can either over-trust incorrect context or underuse context when the user explicitly asks for context-following behavior. Therefore, we propose Intent-Guided Decoding (IGD), a framework that arbitrates between retrieved context and parametric memory according to user intent. IGD uses answer-level filtering and token-level correction to steer the final decoding trajectory between retrieved context and parametric memory. We evaluate IGD on three faithful QA benchmarks and three factual-conflict benchmarks across five LLMs, IGD substantially improves factual recovery, achieving gains of up to 65.4 percentage points on factual-conflict benchmarks over Direct RAG, while preserving or improving strict context-following behavior, this findings highlight the importance of balancing factuality and faithfulness in RAG.

Problem

• Retrieval-Augmented Generation
• Context Trust
• Misleading Context
• Factuality
• Faithfulness

Innovation

• Intent-Guided Decoding
• Retrieval-Augmented Generation
• Token-level Correction
• Factual Recovery
• Context Faithfulness