Causal Evidence for Attention Head Imbalance in Modality Conflict Hallucination

TL;DR

This paper investigates the internal mechanisms behind modality-conflict hallucination in multimodal large language models, identifying key attention heads responsible and proposing a causal intervention to reduce hallucinations effectively.

Contribution

It introduces a causal analysis of attention heads in MLLMs, revealing opposing roles in hallucination, and proposes MACI, a targeted intervention that significantly reduces hallucinations during inference.

Findings

Identified attention head groups with opposing causal roles in hallucination.

Distributed heads drive hallucinations; high-importance heads resist them.

MACI reduces hallucination with a favorable trade-off on multiple benchmarks.

Abstract

Modality-conflict hallucination occurs when multimodal large language models (MLLMs) prioritize erroneous textual premises over contradictory visual evidence. To understand why visual evidence fails to prevail during generation, we take a mechanistic perspective and examine which internal components drive or resist this failure. We perform head-level causal analysis using path patching across five open-source MLLMs and identify two groups of attention heads with opposing causal roles: hallucination-driving heads and hallucination-resisting heads. We find a consistent asymmetry: driving effects are more broadly distributed and carry greater aggregate weight, whereas resisting effects concentrate in a small number of high-importance heads. Ablation experiments further confirm that these groups exert opposing effects during generation: distributed driving influence and localized resistance together form an imbalanced routing structure that biases generation toward the erroneous premise. Motivated by this finding, we propose MACI (Modality-conflict-Aware Causal Intervention), a conditional intervention that suppresses causally identified hallucination-driving heads only when conflict is detected. Across five MLLMs, MACI achieves the largest hallucination reduction among compared inference-time baselines on the MMMC benchmark with a favorable hallucination-accuracy trade-off, and transfers zero-shot to the SCI-SemanticConflict test.