Finding the Correct Visual Evidence Without Forgetting: Mitigating Hallucination in LVLMs via Inter-Layer Visual Attention Discrepancy

Researchers proposed ILVAD, a training-free, plug-and-play method to reduce hallucinations in large vision-language models.

The method analyzes attention across layers, finds repeatedly activated visual tokens, and reinforces grounded text generation.

By preserving visual evidence and reducing visual forgetting, ILVAD improves visual grounding without retraining.

The approach is designed to work across multiple LVLM architectures and make outputs more reliable.