Bridging External and Parametric Knowledge: Mitigating Hallucination of LLMs with Shared-Private Semantic Synergy in Dual-Stream Knowledge

TL;DR

The paper tackles hallucination in retrieval-augmented generation by addressing conflicts between external knowledge and LLM parametric knowledge. It introduces DSSP-RAG, a framework combining unsupervised hallucination detection based on sampling-subspace stability, Energy Quotient (EQ) based external-knowledge filtering, and a mixed-attention mechanism for dual-stream knowledge augmentation that yields shared and private semantics. Through extensive experiments on five QA benchmarks and multiple backbones, DSSP-RAG consistently outperforms strong baselines, with notable gains on multi-hop tasks and improved efficiency, demonstrating robust knowledge integration and conflict resolution. The work advances reliable RAG systems by providing an unsupervised, modular approach to regulate external knowledge use, reduce noise, and enhance reasoning reliability in real-world deployments.

Abstract

Retrieval-augmented generation (RAG) aims to mitigate the hallucination of Large Language Models (LLMs) by retrieving and incorporating relevant external knowledge into the generation process. However, the external knowledge may contain noise and conflict with the parametric knowledge of LLMs, leading to degraded performance. Current LLMs lack inherent mechanisms for resolving such conflicts. To fill this gap, we propose a Dual-Stream Knowledge-Augmented Framework for Shared-Private Semantic Synergy (DSSP-RAG). Central to it is the refinement of the traditional self-attention into a mixed-attention that distinguishes shared and private semantics for a controlled knowledge integration. An unsupervised hallucination detection method that captures the LLMs' intrinsic cognitive uncertainty ensures that external knowledge is introduced only when necessary. To reduce noise in external knowledge, an Energy Quotient (EQ), defined by attention difference matrices between task-aligned and task-misaligned layers, is proposed. Extensive experiments show that DSSP-RAG achieves a superior performance over strong baselines.

Figures (14)

• Figure 1: Comparison of Inference Results Using External vs. Parameterized Knowledge Across Methods
• Figure 2: Dual-Stream Knowledge-Augmented Framework for Shared-Private Semantic Synergy (DSSP-RAG)
• Figure 3: Performance variation with layer-wise pruning
• Figure 4: The analysis of hyperparameters.
• Figure 5: JSD Distances Between Sampling Subspaces of different prompts Under Non-Hallucination and Hallucination Scenarios in LLMs.