SemVink

SemVink: Advancing VLMs’ Semantic Understanding of Optical Illusions via Visual Global Thinking was accepted to the EMNLP 2025 Main Conference and presented as an oral paper.

Vision-language models perform strongly on many semantic tasks, but they often fail at a basic human visual ability: recognizing hidden content in optical illusions and AI-generated images.

We introduce HC-Bench, a benchmark of 112 hidden-content images with hidden texts and objects. On this benchmark, state-of-the-art VLMs struggle even when prompted explicitly.

The key idea behind SemVink (Semantic Visual Thinking) is simple: force the model to reason from a more global visual view. In our experiments, zooming out can substantially improve hidden-content recognition and helps expose a multi-scale weakness in current VLM architectures.

Resources:

• Paper: ACL Anthology
• Project page: johnnyzeppelin.github.io/vlm-semvink
• PDF: semvink_emnlp25.pdf