The TWIML AI Podcast

Why Vision Language Models Ignore What They See with Munawar Hayat - #758

In this episode, Qualcomm AI Research scientist Munawar Hayat explains why Vision‑Language Models often ignore visual input, leading to object hallucination, and how his team’s attention‑guided alignment technique improves visual grounding. He also introduces Generalized Contrastive Learning for efficient multi‑modal retrieval without extra inference cost, and the MultiHuman Testbench, which evaluates and mitigates identity leakage and attribute blending in generative models handling multiple people. Throughout, Hayat emphasizes the importance of these advances for on‑device AI performance and reliability.