Latent sparse attention speeds up ultra long context processing
Latent Sparse Attention: Towards Subquadratic Ultra-Long Context Attention
Artificial Intelligence
Summary
Handling very long stretches of text is hard because it requires a lot of computer power. The authors introduce a new way called Latent Sparse Attention (LSA) that groups pieces of text into clusters, so the model doesn’t have to look at every word to understand the whole. This method makes processing much faster without losing accuracy, even for texts that are millions of words long. They tested two versions, one that improves existing models and another trained from scratch, and both showed good results.
What this means in practice
- •For ai engineers: Reduce computing costs and speed up training and inference of language models that use extremely long text contexts.
- •For software developers: Build applications that require understanding or generating very long documents efficiently without losing result quality.
Authors
Ales Kubicek, Andrea Caciolai, Liangzu Peng, Sukrit Sumant, Benjamin Hoover, Fraser Greenlee, Ruairidh M. Battleday, Radim Urban, Torsten Hoefler, James C. R. Whittington
Abstract
Agentic workflows are increasingly shifting towards long-horizon tasks, such as open-ended scientific work, where models benefit from long context windows. However, at ultra-long context (above 1M tokens), full attention is prohibitively expensive. A promising approach is sparse attention, in which a dedicated indexer selects a subset of keys for the attention. Yet current methods of indexing either still have quadratic cost in context size or lose accuracy at ultra-long context scale. We address this with Latent Sparse Attention (LSA), an efficient attention method that maintains accuracy at ultra-long context. In LSA, keys are hierarchically clustered in a learned latent space, and queries score the hierarchy of clusters instead of scoring every key, making prefill, decode and training faster at long context. We propose two variants, Distilled-LSA (D-LSA), which can be applied to pre-trained models via distillation, and Native-LSA (N-LSA), which is trained natively on the next-token loss alone. On GLM-5.3-FLASH (QWEN-3.8-27B) at 10M tokens, D-LSA speeds up one attention layer by 22.7x (72.4x) in prefill and 4.6x (42.5x) in decode over its original sparse attention (full attention), while keeping similar accuracy across a variety of tasks. N-LSA matches the accuracy of full attention with a training speedup of 1.6x per layer at 128k tokens. LSA offers practical and efficient attention for models to benefit from ultra-long context.