PAYN Token Compression Strategy for Multimodal Referring Expression Segmentation

The PAYN method optimizes token compression for Multimodal Large Language Models by prioritizing the preservation of visual token position embeddings and local spatial structures. This addresses performance degradation seen in standard compression methods during Referring Expression Segmentation tasks.