The emergence of sparse attention: impact of data distribution and benefits of repetition

2025-05-23

AI Summary

This work investigates the emergence mechanism of sparse attention—a recently observed capability—in Transformer models. Addressing the lack of understanding regarding *when* and *why* sparse attention emerges during training, we propose a unified theoretical-empirical framework integrating simplified dynamical modeling, controlled experiments on small-scale Transformers, linear regression variants as probing tasks, and context-dependent recall benchmarks. We systematically characterize the dynamic evolution of sparse attention throughout training. Our key finding is that its emergence follows a precise, analytically tractable power-law temporal pattern, jointly governed by task structure, model architecture, and optimizer dynamics. Furthermore, we demonstrate that data repetition serves as a controllable lever to significantly accelerate emergence—reducing required training iterations by up to an order of magnitude. These results establish a quantitative, predictive paradigm for studying capability emergence in large language models, bridging mechanistic analysis with empirical observability.

Technology Category

• Natural Language Processing: (Large) Language Models
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Practical large-scale studies of user experience

Abstract

Emergence is a fascinating property of large language models and neural networks more broadly: as models scale and train for longer, they sometimes develop new abilities in sudden ways. Despite initial studies, we still lack a comprehensive understanding of how and when these abilities emerge. To address this gap, we study the emergence over training of sparse attention, a critical and frequently observed attention pattern in Transformers. By combining theoretical analysis of a toy model with empirical observations on small Transformers trained on a linear regression variant, we uncover the mechanics driving sparse attention emergence and reveal that emergence timing follows power laws based on task structure, architecture, and optimizer choice. We additionally find that repetition can greatly speed up emergence. Finally, we confirm these results on a well-studied in-context associative recall task. Our findings provide a simple, theoretically grounded framework for understanding how data distributions and model design influence the learning dynamics behind one form of emergence.

Problem

• Understanding emergence of sparse attention in Transformers
• Analyzing impact of task structure on emergence timing
• Exploring benefits of repetition in speeding up emergence

Innovation

• Combining theoretical analysis with empirical observations
• Studying sparse attention emergence in Transformers
• Repetition accelerates emergence timing