InterpDetect: Interpretable Signals for Detecting Hallucinations in Retrieval-Augmented Generation

TL;DR

The paper tackles hallucinations in Retrieval-Augmented Generation by disentangling external context from parametric knowledge. It introduces External Context Score (ECS) and Parametric Knowledge Score (PKS) and trains regression classifiers on span-level signals derived from a 0.6B model to detect hallucinations, with proxy-model evaluation illustrating cross-model generalization to larger systems. The findings show that insufficient external-context usage and over-reliance on parametric knowledge—especially in later FFN layers—drive hallucinations, and that the proposed mechanistic signals provide compact, generalizable predictors that perform comparably to larger baselines in production-like settings. The approach offers a practical, scalable pathway for safer RAG deployment and lays the groundwork for future interventions during generation to steer outputs toward grounded, faithful content.

Abstract

Retrieval-Augmented Generation (RAG) integrates external knowledge to mitigate hallucinations, yet models often generate outputs inconsistent with retrieved content. Accurate hallucination detection requires disentangling the contributions of external context and parametric knowledge, which prior methods typically conflate. We investigate the mechanisms underlying RAG hallucinations and find they arise when later-layer FFN modules disproportionately inject parametric knowledge into the residual stream. To address this, we explore a mechanistic detection approach based on external context scores and parametric knowledge scores. Using Qwen3-0.6b, we compute these scores across layers and attention heads and train regression-based classifiers to predict hallucinations. Our method is evaluated against state-of-the-art LLMs (GPT-5, GPT-4.1) and detection baselines (RAGAS, TruLens, RefChecker). Furthermore, classifiers trained on Qwen3-0.6b signals generalize to GPT-4.1-mini responses, demonstrating the potential of proxy-model evaluation. Our results highlight mechanistic signals as efficient, generalizable predictors for hallucination detection in RAG systems.

Figures (3)

• Figure 1: Pipeline for Span-Level Hallucination Detection
• Figure 2: Relationship Between LLM Utilization of External Context and Hallucination
• Figure 3: Relationship Between LLM Utilization of Parametric Knowledge and Hallucination