Progressive In-Encoder Pruning Runs a Slide-Level Pathology MLLM at 79.84% on TCGA VQA

arXiv·low signalXBlueskyLinkedInCopy link

SLICEChat (arXiv 2609.24894, 21 Sep 2026) attacks the cost of gigapixel whole-slide images, where existing slide-level MLLMs push thousands of patch tokens through multimodal attention and compress only after slide encoding. It puts progressive token pruning inside a hybrid Mamba-Transformer slide encoder, with Mamba layers handling long-range propagation and Transformer layers preserving global interactions as the sequence shortens, and language-supervised region-aware pruning dropping spatially coherent low-utility regions on a controlled keep-rate schedule between stages. On SlideBench VQA it reaches 79.84% on TCGA and 59.09% on BCNB, above prior slide-level pathology MLLMs, with competitive memory and inference latency.