Visual Access Boundaries in Vision-Language Model Reasoning
Chain-of-Thought (CoT) prompting is widely used as a test-time scaling strategy for Vision-Language Models (VLMs), but it remains unclear what is extended when VLMs generate longer reasoning traces. We ask whether CoT requires continued access to image tokens, or whether it mainly operates over visual information already made available earlier in the forward pass. We introduce Visual Access Sweep, a causal intervention that masks attention from generated-token queries to image-token keys along layer depth and generation time, and define the Visual Access Boundary (VAB) as the minimal access re
Lineage graph
Paper → model → repo connections mined from source citations (Tier-1 exact match).
Why these links exist
Every edge carries a method, confidence, and the source snippet that justified it — so bad links are debuggable.
- FuzzySimilar title/name (fuzzy) · 59%VioletVision-3B →
“Fuzzy title match (0.73): “Visual Access Boundaries in Vision-Language Model Reasoning” ≈ “VioletVision-3B””
- LinkedLinked via arxiv author · 85%Hiroto Osaka →
“Visual Access Boundaries in Vision-Language Model Reasoning”
- LinkedLinked via arxiv author · 85%Shohei Taniguchi →
“Visual Access Boundaries in Vision-Language Model Reasoning”
- LinkedLinked via arxiv author · 85%Gouki Minegishi →
“Visual Access Boundaries in Vision-Language Model Reasoning”
- LinkedLinked via arxiv author · 85%Kai Yamashita →
“Visual Access Boundaries in Vision-Language Model Reasoning”
- LinkedLinked via arxiv author · 85%Masahiro Suzuki →
“Visual Access Boundaries in Vision-Language Model Reasoning”
- LinkedLinked via arxiv author · 85%Yutaka Matsuo →
“Visual Access Boundaries in Vision-Language Model Reasoning”
- FuzzySimilar title/name (fuzzy) · 84%pytorch/vision →
“Fuzzy title match (0.92): “Visual Access Boundaries in Vision-Language Model Reasoning” ≈ “pytorch/vision””