# Instruction-tuning Aligns LLMs to the Human Brain

**Authors:** Khai Loong Aw, Syrielle Montariol, Badr AlKhamissi, Martin Schrimpf, Antoine Bosselut  
**arXiv:** [2312.00575](https://arxiv.org/abs/2312.00575)  
**Date:** 2023-12-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.00575v1/

## Key points

- Instruction-tuning is a method used to fine-tune large language models (LLMs) for generating human-like responses.
- The study explores the impact of instruction-tuning on LLM-human similarity using brain alignment and behavioral alignment approaches.
- Brain alignment measures the similarity between LLM internal representations and neural activity in the human language system.
- Behavioral alignment examines the similarity between LLM and human behavior during a reading task.
- Instruction-tuning generally enhances brain alignment by an average of 6%, but does not have a similar effect on behavioral alignment.
- Correlations show that brain alignment is strongly positively correlated with model size (r = 0.95) and performance on tasks involving world knowledge (r = 0.81).
- Instruction-tuning improves world knowledge representations in LLMs and their alignment with human brain activity.
- The study provides insights into how instruction-tuning affects LLM-human similarity from neurobiological and behavioral perspectives.
- These findings highlight the importance of incorporating world knowledge to enhance the alignment of language models with human language processing.

## Summary

In the study titled "Instruction-tuning Aligns LLMs to the Human Brain," authors Khai Loong Aw, Syrielle Montariol, Badr AlKhamissi, Martin Schrimpf, and Antoine Bosselut explore the impact of instruction-tuning on the similarity between large language models (LLMs) and human language processing. Instruction-tuning is a popular method used to fine-tune LLMs, allowing them to generate output that closely resembles human responses to natural language queries. To assess the effect of instruction-tuning on LLM-human similarity, the researchers employ two approaches: brain alignment and behavioral alignment. Brain alignment measures the similarity between internal representations of LLMs and neural activity in the human language system. Behavioral alignment examines the similarity between LLM and human behavior during a reading task. The study involves evaluating 25 vanilla (non-instruction-tuned) and instruction-tuned LLMs using three datasets that involve humans reading naturalistic stories and sentences. The results indicate that instruction-tuning generally enhances brain alignment by an average of 6%. However, it does not have a similar effect on behavioral alignment. To further understand the factors influencing LLM-brain alignment, correlations are computed between brain alignment and various model properties such as model size, problem-solving abilities, and performance on tasks requiring world knowledge across different domains. The analysis reveals a strong positive correlation between brain alignment and model size (r = 0.95), as well as performance on tasks involving world knowledge (r = 0.81). Overall, the findings demonstrate that instruction-tuning improves both world knowledge representations in LLMs and their alignment with human brain activity. This suggests that mechanisms responsible for encoding world knowledge in LLMs also contribute to improving their representational similarity to the human brain. The study provides valuable insights into how instruction-tuning affects LLM-human similarity from both neurobiological and behavioral perspectives. These findings have implications for the development and refinement of language models, highlighting the importance of incorporating world knowledge to enhance their alignment with human language processing.