Large Language Models Align with the Human Brain during Creative Thinking
| Authors | Mete Ismayilzada et al. |
| Year | 2026 |
| HF Upvotes | 1 |
| arXiv | 2604.03480 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
Creative thinking is a fundamental aspect of human cognition, and divergent thinking-the capacity to generate novel and varied ideas-is widely regarded as its core generative engine. Large language models (LLMs) have recently demonstrated impressive performance on divergent thinking tests and prior work has shown that models with higher task performance tend to be more aligned to human brain activity. However, existing brain-LLM alignment studies have focused on passive, non-creative tasks. Here, we explore brain alignment during creative thinking using fMRI data from 170 participants performing the Alternate Uses Task (AUT). We extract representations from LLMs varying in size (270M-72B) and measure alignment to brain responses via Representational Similarity Analysis (RSA), targeting the creativity-related default mode and frontoparietal networks. We find that brain-LLM alignment scales with model size (default mode network only) and idea originality (both networks), with effects strongest early in the creative process. We further show that post-training objectives shape alignment in functionally selective ways: a creativity-optimized Llama-3.1-8B-Instruct preserves alignment with high-creativity neural responses while reducing alignment with low-creativity ones; a human behavior fine-tuned model elevates alignment with both; and a reasoning-trained variant shows the opposite pattern, suggesting chain-of-thought training steers representations away from creative neural geometry toward analytical processing. These results demonstrate that post-training objectives selectively reshape LLM representations relative to the neural geometry of human creative thought.
Engineering Breakdown
Plain English
This paper investigates whether large language models align with human brain activity during creative thinking tasks, specifically divergent thinking measured via the Alternate Uses Task (AUT). The researchers collected fMRI data from 170 participants and compared brain activation patterns against representations from LLMs ranging from 270M to 72B parameters using Representational Similarity Analysis (RSA). Prior work showed that better-performing models align more with passive task brain activity, but this is the first study to examine alignment during active creative cognition. The key finding is that LLM representations correlate with human brain responses during divergent thinking, suggesting that scaling up models may improve their alignment with how humans generate creative ideas.
Core Technical Contribution
The core novelty is applying brain-model alignment methodology to creative tasks rather than passive perception or language understanding. Previous RSA studies focused on non-creative benchmarks where the mapping between model outputs and brain activity was more straightforward; creative divergent thinking introduces open-ended generation where there's no single correct answer. The authors discover that LLMs capture some aspects of the neural representations underlying creative cognition, and they systematically test how this alignment scales with model size (270M to 72B parameters). This is the first quantitative bridge between LLM internal representations and the neural basis of divergent thinking, opening a new evaluation paradigm for assessing model creativity beyond task accuracy.
How It Works
The pipeline starts with 170 human subjects performing the Alternate Uses Task (generating novel uses for everyday objects) while their brain activity is recorded via fMRI, producing voxel-wise activation patterns across brain regions involved in creative cognition. In parallel, the researchers extract intermediate representations (likely from multiple transformer layers) from LLMs of varying sizes when prompted with the same task stimuli. They apply Representational Similarity Analysis (RSA), a neuroscience technique that computes pairwise similarity matrices between brain voxel responses and model layer activations, then correlates these matrices to measure alignment. The key metric is the correlation coefficient between brain RSA matrices and model RSA matrices, aggregated across subjects and targeted to regions hypothesized to support creative thinking. By varying model size and analyzing which layers align best, they build a landscape of how creative cognition maps onto model depth and width.
Production Impact
For engineers building creative AI systems (content generation, ideation tools, design assistance), this provides a novel evaluation metric beyond standard benchmarks: brain alignment as a proxy for human-like creative reasoning. Rather than relying only on task accuracy or human rater preferences, you could use RSA-based brain alignment to select or fine-tune models specifically for creative applications, potentially discovering that certain architectures or training procedures produce more brain-aligned representations without necessarily improving test accuracy. The approach requires access to fMRI data (expensive and limited-scale), so in production you'd likely use this for offline model selection during development, not as a real-time evaluation. This could inform decisions about model scaling for creative tasks — if alignment saturates at 7B parameters but improves at 72B, it justifies the inference cost increase for creative applications where alignment matters for user satisfaction.
Limitations and When Not to Use This
The study's generalizability is constrained by its reliance on fMRI, which has low temporal resolution (2-3 second sampling), potentially missing fast creative processes in the brain; the paper doesn't address whether alignment at the fMRI timescale predicts actual behavioral creativity or human preference. The 170-subject sample, while substantial for neuroscience, doesn't capture the full diversity of human creativity (age, culture, expertise), and it's unclear whether alignment in one population (likely university students in a lab) transfers to other demographics or unconstrained creative settings. The Alternate Uses Task, though a standard creativity measure, is somewhat artificial and may not reflect how creative thinking works in open-ended generative tasks like writing novels or designing products. The paper assumes that brain alignment is a valid goal for LLM training, but doesn't prove that brain-aligned models produce more creative outputs in practical applications or that forcing alignment doesn't harm other capabilities.
Research Context
This work extends the representational alignment literature (pioneered by studies of vision models matching visual cortex) into the cognitive neuroscience domain of creative thinking. It builds on recent findings that larger LLMs show stronger alignment with brain activity on language tasks, but crucially tests whether this relationship holds for generative, non-deterministic creative cognition where there's no ground truth. The paper advances the broader research program of using neuroscience as an evaluation framework for AI systems — complementing benchmarks and human evals with neural data. It opens pathways for using fMRI-based alignment as a fundamental measure of model cognition alignment, potentially influencing how future models are selected or trained if evidence accumulates that brain-aligned models are preferable for creative applications.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::