Vaidya et al. (2023)

Humans and language models diverge when predicting repeating text

arXiv

Find It

DOI

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.

Reference

Vaidya, A. R., Turek, J., & Huth, A. G. (2023). Humans and language models diverge when predicting repeating text. arXiv preprint arXiv:2310.06408. https://doi.org/10.48550/arxiv.2310.06408

What Rests on This

2 mechanisms cite this source. Each finding below is quoted as it appears on that mechanism's own page.

Non-Associative Learning Memory Systems

Measured in AI

The closest same-task comparison had 100 people and GPT-2 predict the next word of passages repeated up to four times: agreement was high on first reading (r = .87) but fell to r = .05 by the third, because GPT-2 became near-perfect after one repetition while people improved gradually; only one AI model was tested, so no spread across systems can be given
Measured in AI

That comparison cannot be used for a rating: it measures memory for repeated text (closer to repetition priming) rather than a waning or growing response, and GPT-2 could still see the earlier passage while people saw words one at a time and had to remember them, so the AI's better score only favours the side the test was easier for

Forgetting Memory Processes

Measured in AI

The one same-test comparison of retention in the conversation gave 100 Prolific participants and GPT-2 Small (2019) text spans repeated up to three times; people saw only the previous 10 words when predicting, while the model saw everything, and it reached near-perfect prediction after one presentation while people improved modestly (human-model correlation fell from r = .87 to .05); the gap favours the side the test was easier for, so it is not counted
Measured in AI

Comparison Not Possible: every human-AI test of this ability either let the AI keep looking at the material and found it did as well or better than people, or had no documented human sample, and no study places a model's training-based retention (where it cannot look back) against human retention curves or the environmental need statistics people track