Sparse attention mechanisms restrict the set of keys each query can attend to, but locating those keys often requires scanning the full prefix at every layer. That scan can dominate compute, especially in long-context models.
The new paper introduces MLIPS (Multi-Layer Indexer Pool Sharing), a training-free approach that shares indexer pools across layers. By reusing a common pool, the method aims to eliminate the need for repeated full-prefix scans at each layer, cutting overhead without modifying model weights.
The abstract does not provide experimental results or architectural details, so the reported benefits are limited to the method's design. The authors position MLIPS as a drop-in efficiency improvement for existing sparse-attention models.