Research
109d agoAttention Maps Fail To Predict VLM Accuracy
A study of 3,090 samples across LLaVA-1.5, PaliGemma, and Qwen2-VL reveals that attention structure is a near-zero predictor of correctness. Researchers used a new VLM Reliability Probe to debunk the intuition that sharp attention maps imply calibrated answers.
The Signal
This finding warns practitioners against using attention heatmaps as a proxy for model reliability.