Why Vision Language Models Ignore What They See

Munawar Hayat discusses visual grounding and object hallucination in vision-language models. Topics include attention-guided alignment, contrastive learning for combined image-and-text retrieval, and evaluating images containing multiple people. The conversation connects identity leakage and attribute blending with the constraints of efficient on-device deployment.