This is a repo for a paper "Language Model Grow Less Humanlike beyond Phase Transition" (ACL, 2025). I'm still cleaning up the repo structure, dependencies, etc., but it should be ready before ACL 2025!
$ conda create -n pt-ppp python=3.8
$ conda activate pt-ppp
$ pip install -r requirements.txt
pt-and-ppp/
│ README.md
│ .gitignore
│
├─ src/
│ ├─ data/ # lightweight data-loading helpers
│ │ ├─ configs/ # .json configs for training
│ │ └─ .../ # all other result files
│ │
│ ├─ models/ # model classes
│ ├─ training/ # training loops, trainer classes, etc.
│ ├─ evaluation/ # metrics, analysis utilities
│ └─ utils/
│
├─ notebooks/
│ └─ figures.ipynb # all figures in the paper can be generated here
│
├─ scripts/ # .sh scripts
│ ├─ train.sh
│ └─ evaluate.sh
│
├─ models/ # all pytorch models
│ └─ model-name
│ └─ checkpoint-xxx/
│
└─ data/ # data for pretraining, evaluation, human eye-tracking/self-paced reading data, etc.
├─ README.md # where to download, expected hashes, etc.
├─ configs/ # .json configs for training
└─ .../ # all other result files
See data/README.md for downloading, generating, and preprocessing data.
See scripts/training_sample.sh for how to train a model with and without syntax/copying regularization.
See scripts/evaluation_sample.sh for how to get LM surprisal, SAS and induction head detection, etc.
For
python -m src.evaluation.get_surprisal -m [model_dir] # get surprisal
python -m src.data.hatch -m [model_dir] # put surprisal values back into the reading time data
Rscript src/evaluation/by_head_analysis.R "$model" NULL NULL # run regression analyses and obtain delta log-likelihood
If you use the code in this repo, please cite our paper:
@inproceedings{aoyama-wilcox-2025-language,
title = "Language Models Grow Less Humanlike beyond Phase Transition",
author = "Aoyama, Tatsuya and
Wilcox, Ethan",
editor = "Che, Wanxiang and
Nabende, Joyce and
Shutova, Ekaterina and
Pilehvar, Mohammad Taher",
booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = jul,
year = "2025",
address = "Vienna, Austria",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.acl-long.1214/",
pages = "24938--24958",
ISBN = "979-8-89176-251-0",
abstract = "LMs' alignment with human reading behavior (i.e. psychometric predictive power; PPP) is known to improve during pretraining up to a tipping point, beyond which it either plateaus or degrades. Various factors, such as word frequency, recency bias in attention, and context size, have been theorized to affect PPP, yet there is no current account that explains why such a tipping point exists, and how it interacts with LMs' pretraining dynamics more generally. We hypothesize that the underlying factor is a pretraining phase transition, characterized by the rapid emergence of specialized attention heads. We conduct a series of correlational and causal experiments to show that such a phase transition is responsible for the tipping point in PPP. We then show that, rather than producing attention patterns that contribute to the degradation in PPP, phase transitions alter the subsequent learning dynamics of the model, such that further training keeps damaging PPP."
}