# Do Large Language Models Mirror Cognitive Language Processing?

**Authors:** Yuqi Ren, Renren Jin, Tongxuan Zhang, Deyi Xiong  
**arXiv:** [2402.18023](https://arxiv.org/abs/2402.18023)  
**Date:** 2024-02-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.18023v1/

## Key points

- Authors explore capabilities of large language models (LLMs) in simulating cognitive language processing
- Propose novel method using Representational Similarity Analysis (RSA) to evaluate effectiveness of LLMs in simulating cognitive language processing
- Positive correlation between model scaling and LLM-brain similarity, larger models align better with cognitive processes
- Alignment training improves similarity between LLM representations and brain signals
- Performance of different LLM evaluations highly correlated with alignment between LLMs and brain activity

## Summary

In their paper titled "Do Large Language Models Mirror Cognitive Language Processing? ", authors Yuqi Ren, Renren Jin, Tongxuan Zhang, and Deyi Xiong explore the capabilities of large language models (LLMs) in simulating cognitive language processing. LLMs have shown impressive performance in tasks such as text comprehension and logical reasoning, often surpassing human-level abilities. The authors raise the question of whether LLMs truly reflect cognitive language processing and to what extent they resemble it. To address this question, the authors propose a novel method that bridges LLM representations with human cognition signals to evaluate the effectiveness of LLMs in simulating cognitive language processing. They utilize Representational Similarity Analysis (RSA) to measure the alignment between 16 mainstream LLMs and functional magnetic resonance imaging (fMRI) signals of the brain. Through empirical investigation, they analyze various factors such as model scaling, alignment training, and instruction appending on the alignment between LLMs and brain activity. The experimental results reveal interesting insights. Firstly, there is a positive correlation between model scaling and LLM-brain similarity, indicating that larger models tend to better align with cognitive processes. Additionally, alignment training proves to be effective in significantly improving the similarity between LLM representations and brain signals. Furthermore, the authors find that the performance of different LLM evaluations like MMLU and Chatbot Arena is highly correlated with the degree of alignment between LLMs and brain activity. Overall,this study sheds light on how well LLMs simulate cognitive language processing by examining their alignment with fMRI signals from the brain. The findings suggest that factors like model size and specialized training can enhance the ability of LLMs to mirror cognitive processes involved in language comprehension and reasoning tasks.