@inproceedings{duan-etal-2026-large,
    title = "Do Large Language Models Acquire Phrase-Based Processing? Evidence from Eye Movements and Model-Brain Alignment After Fine-Tuning",
    author = "Duan, Xufeng  and
      Ma, Zhengwu  and
      Yao, Zhaoqian  and
      Li, Jixing  and
      Cai, Zhenguang",
    editor = "Voigt, Rob  and
      Warstadt, Alex  and
      Feldman, Naomi  and
      Linzen, Tal",
    booktitle = "Proceedings of the Society for Computation in Linguistics 2026",
    month = jul,
    year = "2026",
    address = "San Diego, CA",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.scil-main.44/",
    pages = "464--476",
    ISBN = "979-8-89176-412-5",
    abstract = "Autoregressive large language models (LLMs) process text token-by-token, yet the human language system operates over multi-word units. We ask whether aggregating LLM representations at the phrase level yields a closer correspondence to human reading behavior and language cortex than the default word-level representations, and whether phrase-segmentation fine-tuning amplifies this correspondence. Using Meta-Llama-3.1-8B (base and fine-tuned), we provide three converging lines of evidence. First, phrase-level attention features predict regressive eye-saccade patterns more closely than word-level features; a partial correlation analysis with a shuffled-boundary control indicates that this is not solely an aggregation artifact and that linguistic chunk boundaries explain unique variance beyond word-level attention. Second, fMRI encoding analyses show that fine-tuning selectively improves phrase encoding in left superior temporal gyrus and inferior frontal gyrus, with no improvement for word representations. Third, representational similarity analysis confirms a phrase-specific gain in model-brain geometric alignment. These results identify phrase-level representation as a critical granularity for LLM{--}human correspondence and suggest that targeted training can model human-like compositional processing, linking computational representations to hierarchical theories of language."
}