VISTA adds a position-weighted, parameter-free L2 alignment loss between text token embeddings and visual hidden states to MLLM training, reporting modest benchmark gains but relying on an invalid information-theoretic proof.
Microsoft coco: Common objects in context
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
VISTA: Enhancing Vision-Text Alignment in MLLMs via Cross-Modal Mutual Information Maximization
VISTA adds a position-weighted, parameter-free L2 alignment loss between text token embeddings and visual hidden states to MLLM training, reporting modest benchmark gains but relying on an invalid information-theoretic proof.