Large Language Models Align with the Human Brain during Creative Thinking

A team of researchers from Meta and academic institutions published groundbreaking research showing that large language models (LLMs) exhibit neural alignment with human brains during creative thinking tasks. Using functional MRI data from 170 participants performing the Alternate Uses Task (AUT), they measured brain-LLM alignment via Representational Similarity Analysis (RSA), focusing on creativity-related networks like the default mode and frontoparietal networks. The study found that alignment scales with both model size (270M to 72B parameters) and idea originality, with effects strongest early in the creative process.

Crucially, the research demonstrates that post-training objectives selectively reshape LLM representations relative to human creative thought. A creativity-optimized Llama-3.1-8B-Instruct preserved alignment with high-creativity neural responses while reducing alignment with low-creativity ones. Meanwhile, a human behavior fine-tuned model elevated alignment across both high and low creativity responses, and a reasoning-trained variant showed the opposite pattern—suggesting chain-of-thought training steers representations away from creative neural geometry toward analytical processing. These findings provide the first evidence that LLMs can mirror human creative cognition at the neural level.