IBM Proposes Spatial Matryoshka Training for Multi-Granularity Document Retrieval

_reachsumit · x · 2026-09-01

IBM introduces ColSNAP (Spatial Nested Average Pooling), a training method for visual document retrieval that generates a nested hierarchy of compression levels. This allows the storage-accuracy tradeoff to be configured at indexing time. Experiments show models trained with ColSNAP maintain near full-resolution performance under substantial compression and transfer effectively across multiple backbones.

Original post →

More from Research

Research channel →