#machine learning
Feb 2026
ActTail: Global Activation Sparsity in Large Language Models
ActTail is proposed, a TopK magnitude-based activation sparsity method with global activation sparsity allocation grounded in Heavy-Tailed Self-Regularization (HT-SR) theory that improves both perplexity and downstream task performance at high sparsity compared to uniform allocation.
Wen-Wen Hou, Xin-Yuan Song, Wen-Qian Weng et al.
· arXiv.org · 2 citations