Role-Break in Attention Heads: Understanding and Detecting Hallucinations in VLMs

2026-07-31

AI Summary

This work addresses the persistent challenge of hallucinations in vision-language models (VLMs), which often generate text inconsistent with input images and resist generalization across diverse hallucination patterns. The study identifies that such hallucinations stem from localized anomalies at the attention head level—specifically, deviations from each head’s faithful contextual role, termed “Role-Break”—and reveals for the first time a systematic bias structure underlying these deviations. Building on this insight, the authors propose a lightweight, fine-tuning-free linear detector that leverages attention behavior features of dimensionality under 5,000 to achieve an average AUROC of 93.23% across six VLMs and four benchmarks. This approach yields a stable, unified detection signal across models and tasks, and small-scale interventions demonstrate its utility for discriminative correction of hallucinatory outputs.

Abstract

Despite remarkable progress in vision-language generation, Vision-Language Models (VLMs) remain prone to hallucinations, producing content that is inconsistent with or unsupported by the input image. Existing works largely design detection or mitigation methods around one specific hallucination pattern, such as visual-textual imbalance, but real VLM hallucinations arise from a mixture of multiple patterns, so signals bound to a single pattern struggle to remain stable across models and tasks. Under a unified head-level view, we find that hallucination-induced changes manifest as localized deviations from each head's faithful contextual behavior, a phenomenon we term Role-Break. Detailed analysis reveals that these deviations are systematically organized across attention heads, contextual sources, and deviation directions, and that the resulting signal is linearly readable once head identity is preserved. Based on these findings, we build a lightweight linear detector on top of Role-Break that requires no fine-tuning of the VLM, whose feature dimension stays below 5,000 and reaches an average AUROC of 93.23 across six VLMs and four benchmarks. A small-scale intervention experiment further shows that the detected tokens can be directly acted upon in the discriminative setting.

Problem

• hallucination
• Vision-Language Models
• attention heads
• Role-Break
• visual-textual inconsistency

Innovation

• Role-Break
• Vision-Language Models
• Hallucination Detection
• Attention Heads
• Linear Detector