Index-Preserving Lightweight Token Pruning for Efficient Document Understanding in Vision-Language Models

Researchers have developed a lightweight token pruning framework that reduces computational costs for vision-language models in document understanding tasks by filtering out non-informative background regions before processing. The approach uses a binary patch-level classifier and max-pooling refinement to maintain accuracy while substantially lowering compute demands.

• →New token pruning framework reduces computational burden for vision-language models in document processing
• →Binary patch-level classifier removes non-text areas from document images before VLM processing
• →Max-pooling refinement step recovers fragmented text regions to enhance spatial coherence
• →Experiments show substantial cost reduction while maintaining comparable accuracy on real-world datasets
• →Solution addresses high computational demands that challenge current vision-language model deployment