Long-Context Generalization with Sparse Attention
This addresses the challenge of maintaining precise attention in long-context tasks for NLP researchers and practitioners, though it is incremental as it builds on existing sparse attention methods.
The paper tackled the problem of attention dispersion in transformers for long sequences by introducing Adaptive-Scalable Entmax (ASEntmax), which dynamically sparsifies attention to focus on relevant tokens, achieving up to 1000x length extrapolation on synthetic tasks and better long-context generalization in language modeling.
Transformer-based architectures traditionally employ softmax to compute attention weights, which produces dense distributions over all tokens in a sequence. While effective in many settings, this density has been shown to be detrimental for tasks that demand precise focus on fixed-size patterns: as sequence length increases, non-informative tokens accumulate attention probability mass, leading to dispersion and representational collapse. We show in this paper that dynamically sparse attention mechanisms using $α$-entmax can avoid these issues, due to their ability to assign exact zeros to irrelevant tokens. Furthermore, we introduce Adaptive-Scalable Entmax (ASEntmax), which endows $α$-entmax with a learnable temperature parameter, allowing the attention distribution to interpolate between sparse (pattern-focused) and dense (softmax-like) regimes. Our empirical evaluation on synthetic tasks and language modeling demonstrates that ASEntmax substantially outperforms softmax, scalable softmax, and fixed-temperature $α$-entmax baselines, achieving up to 1000$\times$ length extrapolation on synthetic benchmarks and superior long-context generalization on language modeling while preserving short-context performance, including better perplexity trends and higher retrieval accuracies at 8$\times$ training length.