What is visual question answering (VQA)?

Visual question answering takes an image plus a natural-language question about it and produces an answer — 'How many people are in this photo?' or 'What colour is the car?'. It requires jointly understanding the image and the text and reasoning over both, so it's a standard benchmark for vision-language models. Modern VLMs handle it by encoding the image into visual tokens and letting the language model attend to them while reading the question. It's genuinely useful — reading charts, describing scenes for accessibility, inspecting documents — but watch for failure modes: the model may answer from language priors (guessing a plausible answer without really looking), miscount objects, or hallucinate details not in the image. Grounding and good evaluation matter.