FACTUM: Mechanistic Detection of Citation Hallucination in Long-Form RAG

TL;DR

The paper tackles citation hallucination in long-form retrieval-augmented generation by proposing FACTUM, a mechanistic framework that decomposes model processing into Attention and FFN pathways and four signals: Contextual Alignment Score, Beginning-of-Sentence Attention, Parametric Force, and Pathway Alignment. It demonstrates that the complete signature of these signals yields state-of-the-art detection (up to AUC 0.737) and reveals scale-dependent, evolving strategies—from agreement across all components at smaller scales to specialized, orthogonal contributions at larger scales. It also shows that higher FFN activity can be constructive rather than detrimental, challenging the view that parametric memory alone drives hallucination. Overall, FACTUM provides a nuanced, interpretable basis for verifiable, trustworthy RAG systems and offers practical guidance for reducing attribution drift in long-form generation.

Abstract

Retrieval-Augmented Generation (RAG) models are critically undermined by citation hallucinations, a deceptive failure where a model confidently cites a source that fails to support its claim. Existing work often attributes hallucination to a simple over-reliance on the model's parametric knowledge. We challenge this view and introduce FACTUM (Framework for Attesting Citation Trustworthiness via Underlying Mechanisms), a framework of four mechanistic scores measuring the distinct contributions of a model's attention and FFN pathways, and the alignment between them. Our analysis reveals two consistent signatures of correct citation: a significantly stronger contribution from the model's parametric knowledge and greater use of the attention sink for information synthesis. Crucially, we find the signature of a correct citation is not static but evolves with model scale. For example, the signature of a correct citation for the Llama-3.2-3B model is marked by higher pathway alignment, whereas for the Llama-3.1-8B model, it is characterized by lower alignment, where pathways contribute more distinct, orthogonal information. By capturing this complex, evolving signature, FACTUM outperforms state-of-the-art baselines by up to 37.5% in AUC. Our findings reframe citation hallucination as a complex, scale-dependent interplay between internal mechanisms, paving the way for more nuanced and reliable RAG systems.

Figures (2)

• Figure 1: An example demonstrating how citation hallucination makes factually correct information dangerously unreliable. Both scenarios receive the same query and retrieved documents. Left (Correct Citation): The model correctly cites its source '[Source: 1]’, making the claim verifiable. Right (Hallucinated Citation): The model generates the same correct fact but falsely attributes it to '[Source: 2]’, pointing the user to conflicting information and making the claim unverifiable. This isolates citation hallucination as a critical failure of the model’s attribution mechanism.
• Figure 2: An illustration of the FACTUM framework analyzing a model's internal state at a citation token (e.g., 1), inspired by prior work sun2025redeepdetectinghallucinationretrievalaugmented. The framework derives scores from the two primary pathways. From the Attention Pathway, it computes the Contextual Alignment Score (CAS) to measure grounding in the source documents, and the Beginning-of-Sentence Attention Score (BAS) to measure information synthesis. It also measures the FFN Pathway's Parametric Force Score (PFS) to quantify the magnitude of the FFN's update, and the Pathway Alignment Score (PAS) to assess the geometric alignment between the two pathway updates.