Tag
This paper evaluates how small open-weight vision-language models (Qwen2-VL-2B and SmolVLM) handle realistic image degradations, finding that their verbalized confidence is unreliable while internal token probability provides much better error detection, though both fail under severe low-light conditions.