Image editing models fail zero-shot visual planning on abstract mazes and queen puzzles but generalize after finetuning, yet still cannot match human zero-shot efficiency.
Sparkle: Mastering Basic Spatial Capabilities in Vision Language Models Elicits Generalization to Spatial Reasoning
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Probing Visual Planning in Image Editing Models
Image editing models fail zero-shot visual planning on abstract mazes and queen puzzles but generalize after finetuning, yet still cannot match human zero-shot efficiency.