Task-related visual token compression is feasible at the input stage of MLLMs by learning a lightweight CNN mapping from first-layer attention maps to explainability-derived token importance scores.
Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Task-Related Token Compression in Multimodal Large Language Models from an Explainability Perspective
Task-related visual token compression is feasible at the input stage of MLLMs by learning a lightweight CNN mapping from first-layer attention maps to explainability-derived token importance scores.