Probing Large Language Models from A Human Behavioral Perspective

TL;DR

This paper investigates how large language models predict text by comparing their internal mechanisms and prediction patterns to human reading behaviors, revealing similarities and differences in their processing.

Contribution

It introduces a novel approach of probing LLMs from a human behavioral perspective using eye-tracking data, highlighting how LLMs' internal layers relate to human reading patterns.

Findings

LLMs show prediction patterns similar to humans but differ from shallow models.

Layer depth in LLMs correlates with increased semantic encoding in FFN and MHSA.

Logits in FFN increasingly encode word semantics with more layers.

Abstract

Large Language Models (LLMs) have emerged as dominant foundational models in modern NLP. However, the understanding of their prediction processes and internal mechanisms, such as feed-forward networks (FFN) and multi-head self-attention (MHSA), remains largely unexplored. In this work, we probe LLMs from a human behavioral perspective, correlating values from LLMs with eye-tracking measures, which are widely recognized as meaningful indicators of human reading patterns. Our findings reveal that LLMs exhibit a similar prediction pattern with humans but distinct from that of Shallow Language Models (SLMs). Moreover, with the escalation of LLM layers from the middle layers, the correlation coefficients also increase in FFN and MHSA, indicating that the logits within FFN increasingly encapsulate word semantics suitable for predicting tokens from the vocabulary.