Read or Ignore? A Unified Benchmark for Typographic-Attack Robustness and Text Recognition in Vision-Language Models
TL;DR
- RIO-Bench tests when VLMs should read or ignore text: Unifies useful text recognition and resistance to misleading in-image text.
- Strong VLMs struggle to balance both abilities: Evaluates typographic-attack robustness alongside text-reading capability.
- Adaptive defense learns selective text use: A data-driven approach chooses when in-image text should inform the answer.
Type
Publication
ECCV 2026
