A new automated benchmark of 2D physics questions finds vision-language models are stronger on formulaic tasks than on spatial reasoning, and parameter count does not fully explain performance.
BLIP-2: Bootstrapped language–image pre-training with frozen image encoders and large language models
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
Interpretable Physics Reasoning and Performance Taxonomy in Vision-Language Models
A new automated benchmark of 2D physics questions finds vision-language models are stronger on formulaic tasks than on spatial reasoning, and parameter count does not fully explain performance.