Training-free hallucination mitigation for LVLMs via entropy-guided visual attention masking.
Official implementation of VACoD: Visual Attention Contrastive Decoding for Hallucination Mitigation in Large Vision-Language Models.
Large Vision-Language Models (LVLMs) often generate hallucinated objects that are not grounded in the input image.
VACoD mitigates hallucinations without retraining by:
- Visual Attention Masking
- Contrastive Decoding
- Entropy-guided Dynamic Layer Selection
- Training-free hallucination mitigation
- Dynamic masking layer selection
- Applicable to multiple LVLM architectures
- CHAIR
- POPE
- MME-H
- MMBench
- MM-Vet
Improved hallucination metrics while preserving general visual understanding performance.
ICML 2026