VLM Hallucination

This field studies when vision-language models (VLMs) invent or misread objects, attributes, relations, and details, often favoring language priors over image evidence. Researchers build targeted benchmarks using counterfactual edits, fine-grained annotations, and medical or document images, and measure both error rates and whether models can detect uncertainty or abstain. Mitigation approaches include attention and causal interventions, contrastive decoding, preference tuning, retrieval, and confidence-based verification. A central challenge is reducing unsupported claims without suppressing correct, useful answers, especially in ambiguous, adversarial, or high-stakes settings.