Do Large Language Models Acquire Phrase-Based Processing? Evidence from Eye Movements and Model-Brain Alignment After Fine-Tuning

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

A study using Meta-Llama-3.1-8B (base and fine-tuned) investigates whether large language models acquire phrase-based processing, aligning with human language systems that operate over multi-word units. Researchers found three converging lines of evidence. First, phrase-level attention features more accurately predict human regressive eye-saccade patterns than word-level features, with linguistic chunk boundaries explaining unique variance beyond simple aggregation. Second, fMRI encoding analyses demonstrated that fine-tuning specifically enhances phrase encoding in the left superior temporal gyrus and inferior frontal gyrus, without improving word representations. Third, representational similarity analysis confirmed a phrase-specific improvement in model-brain geometric alignment. These findings suggest that phrase-level representation is a critical granularity for LLM-human correspondence and that targeted training can foster human-like compositional processing in LLMs.

Key takeaway

For NLP engineers developing advanced language models, this research suggests focusing on phrase-level processing can significantly enhance human-like compositional understanding. If your goal is to improve LLM alignment with human cognitive processes, consider fine-tuning models like Meta-Llama-3.1-8B with phrase-segmentation data. This approach could yield models that better predict human reading behavior and exhibit stronger neural correspondence, moving beyond token-by-token limitations.

Key insights

Targeted fine-tuning enables LLMs to acquire human-like phrase-based processing, improving model-brain alignment.

Principles

• Phrase-level processing improves LLM-human correspondence.
• Linguistic chunk boundaries explain unique variance in reading.
• Targeted training can induce compositional processing.

Method

The study used eye-tracking, fMRI encoding, and representational similarity analysis to compare word-level and phrase-level LLM representations against human brain activity and reading behavior, both pre- and post-fine-tuning.

In practice

• Fine-tune LLMs for phrase-level understanding.
• Aggregate LLM representations at phrase level.
• Consider phrase boundaries for human-like text processing.

Topics

• Large Language Models
• Phrase-Based Processing
• Model-Brain Alignment
• Fine-Tuning
• Eye-Tracking
• fMRI Encoding
• Meta-Llama-3.1-8B

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• Beyond representational alignment with brain-guided language models for robust reasoning — Large language models can be directly enhanced by task-evoked brain signals to improve reasoning performance.
• Probing Semantic Alignment, Lexical Invariance, and Syntactic Influence in LLM Metaphor Processing — LLMs struggle with deep metaphor comprehension, often misinterpreting concepts and relying on lexical associations over context or full syntactic understanding.
• Large Language Models in NLP — LLMs integrate NLP advancements with massive scale to create general-purpose, prompt-driven language systems.
• Disentangling Linguistic Relatedness from Task Alignment in Cross-Lingual Transfer — Cross-lingual fine-tuning primarily improves task-format alignment, not language-specific knowledge transfer, especially for weaker models.
• Few Shades of Supervision for Discourse Segmentation — Classical fine-tuning of LLMs is most effective for EDU segmentation in spontaneous speech.
• Linguistic Symptoms: Augmented Generation by Retrieval and Reasoning in LLMs under Portuguese-English Variation in Medical Contexts — Linguistic variation significantly impacts smaller LLMs in medical contexts, with larger models showing greater cross-language robustness.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.