A format-matched control shows perturbation-based selection gives no reliable accuracy gain over spending the same short, no-CoT budget on the original image.
European Conference on Computer Vision (ECCV) , year =
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1roles
background 1polarities
unclear 1representative citing papers
citing papers explorer
-
It's the Decoding Format, Not the Perturbation: Auditing Consistency-Based Selection for Vision-Language Test-Time Scaling
A format-matched control shows perturbation-based selection gives no reliable accuracy gain over spending the same short, no-CoT budget on the original image.