Humans and language models diverge when predicting repeating text
This work addresses a specific gap in modeling human memory effects in language processing, which is incremental but highlights an important limitation in current LMs.
The study identified a scenario where human and language model predictions diverge when text repeats, tracing the cause to specific attention heads and showing that adding a recency bias improves alignment.
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.