Training-free Context-adaptive Attention for Efficient Long Context Modeling

Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of natural language processing tasks. These capabilities stem primarily from the self-attention mechanism, which enables modeling of long-range dependencies. However, the quadratic complexity of self-attention with respect to sequence length poses significant computational and memory challenges, especially as sequence length extends to extremes. While various sparse attention and KV cache compression methods have been proposed to improve efficiency, they often suffer from limitations such as reliance on fixed patterns, inability to handle both prefilling and decoding stages, or the requirement for additional training. In this paper, we propose Training-free Context-adaptive Attention (TCA-Attention), a training-free sparse attention mechanism that selectively attends to only the informative tokens for efficient long-context inference. Our method consists of two lightweight phases: i) an offline calibration phase that determines head-specific sparsity budgets via a single forward pass, and ii) an online token selection phase that adaptively retains core context tokens using a lightweight redundancy metric. TCA-Attention provides a unified solution that accelerates both prefilling and decoding while reducing KV cache memory footprint, without requiring parameter updates or architectural changes. Theoretical analysis shows that our approach maintains bounded approximation error. Extensive experiments demonstrate that TCA-Attention achieves a 2.8$\times$ speedup and reduces KV cache by 61% at 128K context length while maintaining performance comparable to full attention across various benchmarks, offering a practical plug-and-play solution for efficient long-context inference.

Contrast and Clustering: Learning Neighborhood Pair Representation for Source-free Domain Adaptation

Domain adaptation has attracted a great deal of attention in the machine learning community, but it requires access to source data, which often raises concerns about data privacy. We are thus motivated to address these issues and propose a simple yet efficient method. This work treats domain adaptation as an unsupervised clustering problem and trains the target model without access to the source data. Specifically, we propose a loss function called contrast and clustering (CaC), where a positive pair term pulls neighbors belonging to the same class together in the feature space to form clusters, while a negative pair term pushes samples of different classes apart. In addition, extended neighbors are taken into account by querying the nearest neighbor indexes in the memory bank to mine for more valuable negative pairs. Extensive experiments on three common benchmarks, VisDA, Office-Home and Office-31, demonstrate that our method achieves state-of-the-art performance. The code will be made publicly available at https://github.com/yukilulu/CaC.