Humans and language models diverge when predicting repeating text

TL;DR

This paper investigates the divergence between human and language model predictions in repeated text scenarios, identifying specific attention mechanisms responsible and proposing a modification to align their behaviors more closely.

Contribution

The study reveals how and why language models diverge from human predictions in repeated text tasks and introduces a targeted modification to improve their alignment.

Findings

Humans and GPT-2 predictions align initially but diverge with memory effects.

Specific attention heads cause divergence in language models.

Adding a recency bias improves model-human prediction similarity.

Abstract

Language models that are trained on the next-word prediction task have been shown to accurately model human behavior in word prediction and reading speed. In contrast with these findings, we present a scenario in which the performance of humans and LMs diverges. We collected a dataset of human next-word predictions for five stimuli that are formed by repeating spans of text. Human and GPT-2 LM predictions are strongly aligned in the first presentation of a text span, but their performance quickly diverges when memory (or in-context learning) begins to play a role. We traced the cause of this divergence to specific attention heads in a middle layer. Adding a power-law recency bias to these attention heads yielded a model that performs much more similarly to humans. We hope that this scenario will spur future work in bringing LMs closer to human behavior.