Sparse attention is designed to visit only a few keys, but the process of finding those keys can be expensive: it may still require scoring every key with every indexer head. A new arXiv paper introduces DoSA (Dots Sparse Attention), a method that makes this computation conditional, so the full scoring step is not always needed.
The abstract describes DoSA as lossless, meaning it does not sacrifice the correctness of the attention output. This is notable because many efficiency tricks trade away some fidelity for speed. DoSA instead aims to reduce the cost of the search itself while preserving the original result.
The paper is brief in its public abstract, so details on how the conditional computation is implemented or benchmarked are not yet available. The core claim, however, is clear: sparse attention can be made cheaper without losing information.