Where Reliability Lives in Vision-Language Models: A Mechanistic Study of Attention, Hidden States, and Causal Circuits

This research tests the "Attention-Confidence Assumption" in Vision-Language Models (VLMs), finding that attention structure is a near-zero predictor of correctness. The study uses a unified mechanistic pipeline (VLM Reliability Probe) to analyze attention, generation dynamics, and hidden-state geometry in three VLM families.