By probing visual, projection, and response representations, the authors find that most VLM visual knowledge loss for recognition and counting occurs in the language decoder, while spatial understanding is lost in the visual encoder.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Response Wide Shut? Surprising Observations in Basic Vision Language Model Capabilities
By probing visual, projection, and response representations, the authors find that most VLM visual knowledge loss for recognition and counting occurs in the language decoder, while spatial understanding is lost in the visual encoder.