This research identifies the middle layers of Large Vision-Language Models (LVLMs) as the source of object hallucinations and provides novel methods to detect and address this critical issue.

• Reveals that middle layers are where visual hallucinations originate
• Develops an attention-based detector that identifies hallucinations with high accuracy
• Proposes an effective mitigation technique that reduces hallucinations without retraining
• Demonstrates improvements across multiple benchmark datasets

From a security perspective, this work directly enhances model reliability and trustworthiness by reducing false information generation, making AI systems safer for deployment in critical applications.

Original Paper: Devils in Middle Layers of Large Vision-Language Models: Interpreting, Detecting and Mitigating Object Hallucinations via Attention Lens