Shiyao Li

Tianyu Fu, Haofeng Huang, Xuefei Ning et al.

Sliding-window attention offers a hardware-efficient solution to the memory and throughput challenges of Large Language Models (LLMs) in long-context scenarios. Existing methods typically employ a single window length across all attention heads and input sizes. However, this uniform approach fails to capture the heterogeneous attention patterns inherent in LLMs, ignoring their distinct accuracy-latency trade-offs. To address this challenge, we propose *Mixture of Attention Spans* (MoA), which automatically tailors distinct sliding-window length configurations to different heads and layers. MoA constructs and navigates a search space of various window lengths and their scaling rules relative to input sizes. It profiles the model, evaluates potential configurations, and pinpoints the optimal length configurations for each head. MoA adapts to varying input sizes, revealing that some attention heads expand their focus to accommodate longer inputs, while other heads consistently concentrate on fixed-length local contexts. Experiments show that MoA increases the effective context length by 3.9x with the same average sliding-window length, boosting retrieval accuracy by 1.5-7.1x over the uniform-window baseline across Vicuna-{7B, 13B} and Llama3-{8B, 70B} models. Moreover, MoA narrows the performance gap with full attention, reducing the maximum relative performance drop from 9%-36% to within 5% across three long-context understanding benchmarks. MoA achieves a 1.2-1.4x GPU memory reduction, boosting decode throughput by 6.6-8.2x and 1.7-1.9x over FlashAttention2 and vLLM, with minimal performance impact. Our code is available at: https://github.com/thu-nics/MoA

Feng-Qi Cui, Zhen Lin, Xinlong Rao et al.

Dynamic facial expression recognition (DFER) faces significant challenges due to long-tailed category distributions and complexity of spatio-temporal feature modeling. While existing deep learning-based methods have improved DFER performance, they often fail to address these issues, resulting in severe model induction bias. To overcome these limitations, we propose a novel multi-instance learning framework called MICACL, which integrates spatio-temporal dependency modeling and long-tailed contrastive learning optimization. Specifically, we design the Graph-Enhanced Instance Interaction Module (GEIIM) to capture intricate spatio-temporal between adjacent instances relationships through adaptive adjacency matrices and multiscale convolutions. To enhance instance-level feature aggregation, we develop the Weighted Instance Aggregation Network (WIAN), which dynamically assigns weights based on instance importance. Furthermore, we introduce a Multiscale Category-aware Contrastive Learning (MCCL) strategy to balance training between major and minor categories. Extensive experiments on in-the-wild datasets (i.e., DFEW and FERV39k) demonstrate that MICACL achieves state-of-the-art performance with superior robustness and generalization.