오랜만에 논문리뷰.
VLM 관련된 논문을 몇 개 읽어볼까 한다. 오랜만에 VLM 자체에 대한 논문들을 보니 재밌다.
ICLR 2026 논문이다.
Abstract
1. Introduction
2. Dissecting Text and Visual Attention Dynamics in VLMs
2.1 RQ1: How does attention transition between text and visual tokens?
2.2 RQ2: Which image regions do different layers attend to?
2.3 RQ3: Do VLMs perceive visual evidence when giving wrong answers?
2.4 RQ4: Why does "Seeing but not believing" happens?
3. Simple Inference - Time Visual Evidence Augmentation
3.1 Visual evidence attribution layer profiling
3.2 Inference - Time visual evidence attribution
3.3 Attention-guided Visual evidence highlighting
4. Experiment results and analysis
5. Related works
6. Conclusions