Vision-language models appear to bind an object's image tokens to its textual references through shared binding ID vectors in their activations, and these vectors can be swapped causally to change the model's answers.
Interpreting and Editing Vision- Language Representations to Mitigate Hallucinations
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
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Investigating Mechanisms for In-Context Vision Language Binding
Vision-language models appear to bind an object's image tokens to its textual references through shared binding ID vectors in their activations, and these vectors can be swapped causally to change the model's answers.