Instruction-tuned large language models misalign with natural language comprehension in humans

Abstract

Transformer-based language models have significantly advanced our understanding of meaning representation in the human brain. Prior research utilizing smaller models like BERT and GPT-2 suggests that "next-word prediction" is a computational principle shared between machines and humans. However, recent advancements in large language models (LLMs) have highlighted the effectiveness of instruction tuning beyond next-word prediction. It remains to be tested whether instruction tuning can further align the model with language processing in the human brain. In this study, we evaluated the self-attention of base and finetuned LLMs of different sizes against human eye movement and functional magnetic resonance imaging (fMRI) activity patterns during naturalistic reading. Our results reveal that increases in model size significantly enhance the alignment between LLMs and brain activity, whereas instruction tuning does not. These findings confirm a scaling law in LLMs' brain-encoding performance and suggest that "instruction-following" may not mimic natural language comprehension in humans.