for 3D medical scans face a memory bottleneck. A single CT scan produces tens of thousands of visual tokens, and the resulting cache scales linearly with batch size, making clinical deployment on standard GPUs difficult.
TACT-KV treats the visual key-value cache as a 3D spatial volume rather than a flat sequence. It applies a separable 3D cosine transform to concentrate information and allocates bits adaptively across frequency components.
This achieves the lowest prediction-flip rate on eleven of twelve benchmarks while allowing up to eight times larger batches on the same hardware. The approach also transfers to video tasks, confirming the value of exploiting 3D structure in the cache.