Language models and brains align due to more than next-word prediction and word-level information

2022/12/01 by Gabriele Merlin, Mariya Toneva, Merlin, Gabriele +1 · 6 citations
Computer Science · #Computation and Language (cs.CL) #FOS: Biological sciences #FOS: Computer and information sciences #Multimodal Machine Learning Applications #Natural Language Processing Techniques #Neurons and Cognition (q-bio.NC) #Topic Modeling

paper · pdf · doi:10.48550/arxiv.2212.00596

openalex publication_date 2022/12/01 · openalex created_date 2022/12/13 · openalex updated_date 2026/07/28

Abstract

Pretrained language models have been shown to significantly predict brain recordings of people comprehending language. Recent work suggests that the prediction of the next word is a key mechanism that contributes to this alignment. What is not yet understood is whether prediction of the next word is necessary for this observed alignment or simply sufficient, and whether there are other shared mechanisms or information that are similarly important. In this work, we take a step towards understanding the reasons for brain alignment via two simple perturbations in popular pretrained language models. These perturbations help us design contrasts that can control for different types of information. By contrasting the brain alignment of these differently perturbed models, we show that improvements in alignment with brain recordings are due to more than improvements in next-word prediction and word-level information.

Cited by

• Brain-tuning Improves Generalizability and Efficiency of Brain Alignment in Speech Models
• Fine-grained Analysis of Brain-LLM Alignment through Input Attribution
• Stacked Regression using Off-the-shelf, Stimulus-tuned and Fine-tuned Neural Networks for Predicting fMRI Brain Responses to Movies (Algonauts 2025 Report)
• Brain-Language Model Alignment: Insights into the Platonic Hypothesis and Intermediate-Layer Advantage
• Language models align with brain regions that represent concepts across modalities
• Do VLMs Align Better with Humans than LLMs during Natural Reading?
• Interpreting Brain Responses to Language with Sparse Features from Language Models

Related

• Fine-grained Analysis of Brain-LLM Alignment through Input Attribution
• Joint processing of linguistic properties in brains and language models
• From Language to Cognition: How LLMs Outgrow the Human Language Network
• Brain-Language Model Alignment: Insights into the Platonic Hypothesis and Intermediate-Layer Advantage
• Brain-Like Language Processing via a Shallow Untrained Multihead Attention Network