Research10h ago
Attention Maps Poorly Predict VLM Correctness
A near-zero correlation of 0.001 exists between attention structure and answer correctness in LLaVA-1.5, PaliGemma, and Qwen2-VL. Researchers used a new VLM Reliability Probe to debunk the intuition that sharp attention maps imply calibrated confidence. This finding warns practitioners against using attention heatmaps as a proxy for model reliability or trust.