Causal Tracing of Object Representations in Large Vision Language Models: Mechanistic Interpretability and Hallucination Mitigation

TL;DR

The paper tackles mechanistic interpretability and hallucination in large vision-language models by introducing Fine-grained Cross-modal Causal Tracing (FCCT), which uses Gaussian perturbations and activation patching to quantify cross-modal causal effects across visual and textual tokens and three core components across all layers, with the Recovery Rate $RR$ as the causal metric. FCCT uncovers that last-token MHSA in middle layers drives cross-modal aggregation, FFNs exhibit a three-stage hierarchical representation, and hidden states undergo a hierarchical semantic shift, informing the design of a training-free inference-time technique called Intermediate Representation Injection (IRI). IRI injects mid-layer cross-modal representations into later layers guided by $RR$, significantly mitigating object hallucinations while preserving inference speed. Across five LVLMs and five benchmarks, IRI achieves state-of-the-art hallucination mitigation and validates FCCT's mechanistic insights, offering practical guidance for robust cross-modal perception in LVLMs.

Abstract

Despite the remarkable advancements of Large Vision-Language Models (LVLMs), the mechanistic interpretability remains underexplored. Existing analyses are insufficiently comprehensive and lack examination covering visual and textual tokens, model components, and the full range of layers. This limitation restricts actionable insights to improve the faithfulness of model output and the development of downstream tasks, such as hallucination mitigation. To address this limitation, we introduce Fine-grained Cross-modal Causal Tracing (FCCT) framework, which systematically quantifies the causal effects on visual object perception. FCCT conducts fine-grained analysis covering the full range of visual and textual tokens, three core model components including multi-head self-attention (MHSA), feed-forward networks (FFNs), and hidden states, across all decoder layers. Our analysis is the first to demonstrate that MHSAs of the last token in middle layers play a critical role in aggregating cross-modal information, while FFNs exhibit a three-stage hierarchical progression for the storage and transfer of visual object representations. Building on these insights, we propose Intermediate Representation Injection (IRI), a training-free inference-time technique that reinforces visual object information flow by precisely intervening on cross-modal representations at specific components and layers, thereby enhancing perception and mitigating hallucination. Consistent improvements across five widely used benchmarks and LVLMs demonstrate IRI achieves state-of-the-art performance, while preserving inference speed and other foundational performance.

Figures (4)

• Figure 1: An overview of our proposed Fine-grained Cross-modal Causal Tracing (FCCT) findings and Intermediate Representation Injection (IRI) method.
• Figure 2: Overview of our proposed Fine-grained Cross-modal Causal Tracing method. Activation patching computes the causal effect of a specific component by running the LVLMs three times: a clean run (step1) with original image, a corrupted run (step2) with image added Gaussian noise, and a patched run (step3) with corrupted input but restoring specific component using the value in the clean run. We use Recovery Rate to quantify the causal effect of each restored component.
• Figure 3: Results and key findings of FCCT framework on LLaVA-1.5-7b and Qwen-VL-Chat. The symbols from to represent the seven token categories defined above: Early Visual Tokens, Object Visual Tokens, Late Visual Tokens, Early Textual Tokens, Textual Object Tokens, Late Textual Tokens, and The Last Token.
• Figure 4: Visualization of normalized attention weights to visual object tokens and corresponding textual object tokens across layers. We report the average result on 3,000 VQAs.