Cognitive Mirrors: Exploring the Diverse Functional Roles of Attention Heads in LLM Reasoning

TL;DR

This paper introduces a novel interpretability framework and dataset to analyze attention heads in large language models, revealing their specialized roles in reasoning and their impact on model performance.

Contribution

It proposes CogQA and a multi-class probing method to identify and analyze cognitive attention heads, uncovering their functional specialization and importance in reasoning tasks.

Findings

Attention heads exhibit functional specialization as cognitive heads.

Removing cognitive heads degrades reasoning performance.

Augmenting cognitive heads improves reasoning accuracy.

Abstract

Large language models (LLMs) have achieved state-of-the-art performance in a variety of tasks, but remain largely opaque in terms of their internal mechanisms. Understanding these mechanisms is crucial to improve their reasoning abilities. Drawing inspiration from the interplay between neural processes and human cognition, we propose a novel interpretability framework to systematically analyze the roles and behaviors of attention heads, which are key components of LLMs. We introduce CogQA, a dataset that decomposes complex questions into step-by-step subquestions with a chain-of-thought design, each associated with specific cognitive functions such as retrieval or logical reasoning. By applying a multi-class probing method, we identify the attention heads responsible for these functions. Our analysis across multiple LLM families reveals that attention heads exhibit functional specialization, characterized as cognitive heads. These cognitive heads exhibit several key properties: they are universally sparse, vary in number and distribution across different cognitive functions, and display interactive and hierarchical structures. We further show that cognitive heads play a vital role in reasoning tasks - removing them leads to performance degradation, while augmenting them enhances reasoning accuracy. These insights offer a deeper understanding of LLM reasoning and suggest important implications for model design, training, and fine-tuning strategies.