CHASD: Language Increment-Calibrated Contrastive Decoding against Hallucination in LVLMs
Researchers have developed a new inference-time framework called CHASd to combat hallucinations in Large Vision-Language Models (LVLMs). This method, Contrastive Hallucination-Aware Step-wise Decoding, selectively activates a contrastive decoding branch only when token prediction confidence is low. It uses localized visual perturbations guided by attention to minimize interference with useful visual evidence, improving hallucination metrics on several benchmarks while maintaining efficient inference. AI
IMPACT Reduces object hallucinations in vision-language models, improving reliability for multimodal AI applications.