A new controlled benchmark shows that leading vision-language models recognize shapes reliably but struggle with properties like sheen, show position-dependent accuracy, and perform poorly on basic spatial relations.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
CIVET: Systematic Evaluation of Understanding in VLMs
A new controlled benchmark shows that leading vision-language models recognize shapes reliably but struggle with properties like sheen, show position-dependent accuracy, and perform poorly on basic spatial relations.