Search papers, labs, and topics across Lattice.
This paper introduces Hierarchical Landmark Sparse (HiLS) Attention, a novel chunk-wise sparse attention mechanism that learns chunk selection end-to-end under the language-modeling loss, addressing the limitations of existing methods in long-context modeling. By hierarchically factorizing attention and optimizing retrieval scores directly during training, HiLS-Attention achieves performance on par with full attention while enabling extrapolation beyond 64 times the training context length with high retrieval accuracy. The approach not only enhances efficiency but also maintains or improves effectiveness on long-context tasks, demonstrating a significant advancement in scaling large language models.
HiLS-Attention achieves over 64x context length extrapolation with 90% retrieval accuracy, outperforming traditional full attention mechanisms.
Scaling modern large language models (LLMs) to long contexts is limited by the quadratic computation cost, and poor length extrapolation of dense attention. Chunk-wise sparse attention offers a promising alternative, but all existing methods fall short of full attention because of their inaccurate chunk selection. We propose Hierarchical Landmark Sparse (HiLS) Attention, a chunk-wise sparse attention mechanism that learns chunk selection end-to-end under the language-modeling (LM) loss. HiLS factorizes attention hierarchically: each query performs attention independently with each retrieved chunk to extract chunk-specific information, and the resulting outputs are fused according to chunk retrieval scores. By incorporating retrieval scores into the forward attention computation, HiLS optimizes them directly with the LM loss, enabling end-to-end retrieval learning and native sparse training. Experimental results show that HiLS-Attention achieves performance comparable to, and in some cases better than, full attention at in-domain context lengths. Meanwhile, HiLS-Attention extrapolates more than $64\times$ the training context length with 90% retrieval accuracy, far beyond full attention. Moreover, existing full-attention models can be converted to HiLS-Attention with lightweight continued pretraining, preserving in-domain performance while acquiring ultra-long-context extrapolation. Together with its sparse KV access and computation, HiLS-Attention breaks the usual efficiency-performance trade-off, enabling long-context LLMs that are both more efficient and more effective on general long-context tasks than their full-attention counterparts.
