A recent study reveals that attention maps in vision-language models (VLMs) are not reliable indicators of correctness, with hidden-state geometry proving to be a more accurate measure of reliability. This finding suggests that improvements in VLM architecture could enhance performance without relying solely on attention sharpness.