Demystifying the Slash Pattern in Attention: The Role of RoPE

TL;DR

This paper investigates the emergence of slash attention patterns in large language models, revealing their intrinsic nature and explaining their formation through empirical analysis and theoretical modeling of queries, keys, and Rotary Position Embedding.

Contribution

It provides a comprehensive analysis of slash-dominant heads in LLMs, identifying key conditions and offering a theoretical framework that explains their emergence and generalization.

Findings

SDHs are intrinsic to models and generalize out-of-distribution

Queries and keys are nearly rank-one in SDHs

RoPE's medium- and high-frequency components drive SDHs

Abstract

Large Language Models (LLMs) often exhibit slash attention patterns, where attention scores concentrate along the $\Delta$-th sub-diagonal for some offset $\Delta$. These patterns play a key role in passing information across tokens. But why do they emerge? In this paper, we demystify the emergence of these Slash-Dominant Heads (SDHs) from both empirical and theoretical perspectives. First, by analyzing open-source LLMs, we find that SDHs are intrinsic to models and generalize to out-of-distribution prompts. To explain the intrinsic emergence, we analyze the queries, keys, and Rotary Position Embedding (RoPE), which jointly determine attention scores. Our empirical analysis reveals two characteristic conditions of SDHs: (1) Queries and keys are almost rank-one, and (2) RoPE is dominated by medium- and high-frequency components. Under these conditions, queries and keys are nearly identical across tokens, and interactions between medium- and high-frequency components of RoPE give rise to SDHs. Beyond empirical evidence, we theoretically show that these conditions are sufficient to ensure the emergence of SDHs by formalizing them as our modeling assumptions. Particularly, we analyze the training dynamics of a shallow Transformer equipped with RoPE under these conditions, and prove that models trained via gradient descent exhibit SDHs. The SDHs generalize to out-of-distribution prompts.