Definition
Plain language
An index that uses geometry to skip over irrelevant past tokens during attention.
As stated in the literature
A halfspace-range-searching-based attention sparsification system using bounding-ball pruning across factored subspaces, with completeness guarantees above a threshold and constant-memory index state.
Why it matters: It promises long-context attention with full-attention-quality recall but at a fraction of the compute and memory.
For example, when computing attention for a new token, Louver geometrically rules out distant past tokens that couldn't possibly score high enough to matter, avoiding most dot products.