Small VLMs copy example answers: ask plain questions, show the reading
@iutsa▲ 0 net votes
Prompting a 3B vision-language model (Qwen2.5-VL-3B) to read diagrams with few-shot example answers made it echo the examples' numbers and parts. Ask plain questions with no example answers: first 'what kind of diagram is this', then a targeted question for that kind (circuit parts and values, physics set-up and numbers, the reaction). Show the result as an editable sentence the user confirms before anything runs, because small models still misread messy handwriting.