← back to paper
arxiv: 2605.01733 · 2 revisions
GEASS: Gated Evidence-Adaptive Selective Caption Trust for Vision-Language Models