Local Linear Attention: An Optimal Interpolation of Linear and Softmax Attention For Test-Time Regression
This work addresses the need for more expressive attention mechanisms in transformers, offering a novel approach that bridges theoretical insights with practical efficiency, though it appears incremental in advancing existing attention methods.
The paper tackles the problem of improving attention mechanisms in transformers by proposing Local Linear Attention (LLA), derived from nonparametric statistics for test-time regression, and demonstrates that it outperforms strong baselines in tasks like test-time training and in-context learning, with evidence of scalability in large-scale models.
Transformer architectures have achieved remarkable success in various domains. While efficient alternatives to Softmax Attention have been widely studied, the search for more expressive mechanisms grounded in theoretical insight-even at greater computational cost-has been relatively underexplored. In this work, we bridge this gap by proposing Local Linear Attention (LLA), a novel attention mechanism derived from nonparametric statistics through the lens of test-time regression. First, we show that LLA offers theoretical advantages over Linear and Softmax Attention for associative memory via a bias-variance trade-off analysis. Next, we address its computational challenges and propose two memory-efficient primitives to tackle the $Θ(n^2 d)$ and $Θ(n d^2)$ complexity. We then introduce FlashLLA, a hardware-efficient, blockwise algorithm that enables scalable and parallel computation on modern accelerators. In addition, we implement and profile a customized inference kernel that significantly reduces memory overheads. Finally, we empirically validate the advantages and limitations of LLA on test-time regression, in-context regression, associative recall and state tracking tasks. Experiment results demonstrate that LLA effectively adapts to non-stationarity, outperforming strong baselines in test-time training and in-context learning, and exhibiting promising evidence for its scalability and applicability in large-scale models. Code is available at https://github.com/Yifei-Zuo/Flash-LLA.