Large Language Models Align with the Human Brain during Creative Thinking

This study demonstrates that large language models exhibit increasing alignment with human brain activity during creative thinking as model size grows, and that specific post-training objectives (such as creativity optimization, human behavior fine-tuning, or reasoning training) selectively reshape these neural representations to either preserve, enhance, or diverge from the brain's geometry of creative thought.

The Gist

Imagine your brain is a bustling city, and different neighborhoods handle different jobs. Some neighborhoods are great at following strict maps (like solving math problems), while others are like chaotic art studios where wild, new ideas are born.

This paper is like a scientific tour comparing Large Language Models (LLMs)—the super-smart AI chatbots you might know—to the human brain while it's doing something very specific: Creative Thinking.

Here is the story of what they found, broken down simply:

1. The Test: "The Brick Game"

To test creativity, the researchers used a classic game called the Alternate Uses Task.

• The Human Side: They put 170 people in an MRI machine (a giant brain scanner) and showed them a common object, like a brick. They asked the people: "What are some creative, weird ways to use this brick?" (Maybe as a doorstop, a paperweight, or a weapon in a play).
• The AI Side: They fed the exact same "brick" prompts to many different AI models, ranging from tiny ones to massive ones.

2. The Big Discovery: Bigger Isn't Always Better (At First)

The researchers measured how much the AI's "thought process" looked like the human brain's "thought process." They found two main things:

• The "Size" Effect: Generally, bigger AI models (with more "brain power") had thoughts that looked more like human creative thoughts when they were just reading the question. It's like a bigger library having more books on creativity.
• The "Performance" Effect: The AI models that were actually better at coming up with creative answers also had brain patterns that matched humans more closely.

The Catch: This match was strongest when the AI was just thinking about the prompt. Once the AI started writing its answer, the connection to the human brain got weaker.

• Analogy: Imagine a jazz musician. When they are listening to the band and getting ready to play, their mind is in perfect sync with the group. But once they start playing a solo, they might get so carried away with their own technical skills that they drift away from the original group vibe. The AI seems to do the same thing; it gets "too technical" when generating text.

3. The "Training Diet" Matters Most

The most fascinating part of the paper is how training changes the AI's brain. The researchers took a standard AI model and gave it three different "special diets" (post-training objectives) to see how it changed its creative alignment:

• The "Creativity Chef" (Creativity-Optimized): They trained an AI specifically to be more creative.
  • Result: This AI's brain started looking more like the human brain when humans were having "Aha!" moments (high creativity) and less like the human brain when they were having boring ideas. It learned to tune into the "art studio" neighborhood of the brain.
• The "Human Mimic" (Behavior Fine-Tuned): They trained an AI to copy how humans act and talk.
  • Result: This AI matched the human brain for both creative and non-creative ideas. It became a good chameleon.
• The "Logic Puzzle Solver" (Reasoning/Chain-of-Thought): They trained an AI to solve hard logic puzzles and math problems step-by-step.
  • Result: This was the surprise! This AI's brain actually stopped matching human creative thoughts. In fact, it matched the boring human thoughts better than the creative ones.
  • Analogy: It's like training a jazz musician to only play strict classical sheet music. Eventually, they forget how to improvise. The "Logic Solver" AI got so good at following strict rules that it lost the ability to "think like a human artist."

4. The Deep Layers

The researchers also looked at which "layers" of the AI were doing the thinking. They found that the deeper, more complex layers of the AI (the ones that handle abstract concepts) were the ones that matched the human brain best. The early, simple layers (which just understand basic words) didn't match as well. This makes sense because creativity is a high-level, complex activity, not just a simple word game.

The Bottom Line

This paper tells us that AI is getting better at mimicking human creativity, but it's a delicate balance.

• If you just make the AI bigger, it gets closer to human creativity.
• But if you train the AI too hard on logic and strict rules (like math or coding), it actually loses its ability to think like a creative human.
• To build AI that can truly help us with art, science, and new ideas, we need to train it to value diversity and weirdness, not just correct answers.

In short: To make an AI that thinks like a human artist, you have to teach it to be a little messy, not just a perfect logic machine.

Technical Summary

1. Problem Statement

While Large Language Models (LLMs) have demonstrated impressive performance on divergent thinking tasks (e.g., the Alternate Uses Task, or AUT), it remains unclear whether their internal representations align with the neural mechanisms underlying human creativity.

• Gap in Literature: Previous brain-LLM alignment studies have focused almost exclusively on passive language processing (reading/listening) or simple word association. They have not addressed active creative thinking, which involves higher-order cognitive functions and distributed brain networks beyond the classical language system.
• Core Question: Do LLMs with higher creative performance exhibit greater alignment with human brain activity during creative ideation? Furthermore, how do different post-training objectives (e.g., reasoning vs. creativity optimization) affect this neural alignment?

2. Methodology

Data Sources

• Neural Data: fMRI data from 170 healthy participants (from Beaty et al., 2018) performing two tasks:
  1. Alternate Uses Task (AUT): A creative divergent thinking task where participants generate novel uses for objects.
  2. Object Characteristics Task (OCT): A non-creative control task where participants generate typical physical properties of objects.
• Model Data: A diverse set of open-source LLMs (ranging from 270M to 72B parameters) including Gemma-3, Llama-3.1/3.2, Qwen2.5, DeepSeek-R1, and Falcon.
  • Specific variants of Llama-3.1-8B were used to isolate the effects of post-training:
    • Base: Pretrained only.
    • Instruct: Standard instruction tuning.
    • CrPO: Optimized for human creative preferences (novelty, diversity, quality).
    • Minitaur: Fine-tuned to simulate human behavior.
    • DeepSeek-R1 Distill: Fine-tuned with Chain-of-Thought (CoT) reasoning.

Experimental Design & Processing

• Stimuli: 46 object stimuli were presented to both humans and models.
• Representation Extraction:
  • Prompt Stage: Activations extracted after the input prompt (before generation).
  • Generation Stage: Activations extracted after the model generates a response.
  • Token Strategy: Both "last token" and "mean token" representations were evaluated.
• Brain Regions of Interest (ROIs):
  • Default Mode Network (DMN): Core network for creative cognition.
  • Frontoparietal Network (FPN): Involved in cognitive control and creative flexibility.
  • Somatomotor Network: Control network (minimal involvement in high-order cognition).
• Alignment Metric: Representational Similarity Analysis (RSA).
  • Constructed Representational Dissimilarity Matrices (RDMs) for both fMRI data and model activations.
  • Calculated Pearson correlation between RDMs.
  • Scores were normalized by a noise ceiling to account for fMRI reliability.
  • The maximum alignment score across all model layers was selected for each model.

Creativity Partitioning

Human responses were rated by independent raters (1–5 scale). Data was split into High-Creativity (rating ≥ 2.0) and Low-Creativity (rating < 2.0) populations to test if models align differently with varying levels of human creativity.

3. Key Contributions

1. First Active Creative Alignment Study: The first systematic investigation of brain-LLM alignment during an active divergent thinking task (AUT), moving beyond passive language processing.
2. Stage-Dependent Alignment: Discovery that alignment is strong during the prompt processing stage but weakens significantly once models begin generating responses.
3. Post-Training Selectivity: Demonstration that post-training objectives selectively reshape the "neural geometry" of models:
   • Creativity-optimized models align better with high-creativity neural patterns.
   • Reasoning-chain (CoT) trained models align better with low-creativity (analytical) patterns, effectively diverging from creative neural geometry.

4. Key Results

A. Model Size and Performance Correlation

• Prompt Stage: There is a significant positive correlation between model size (parameters) and brain alignment in the DMN ($r=0.58, p<0.05$).
• Performance: Alignment also correlates with AUT task performance ($r=0.51, p<0.05$). Larger, more creative models mirror human neural representations more closely before generation.
• Generation Stage: These correlations weaken or disappear once the model generates a response, suggesting that the generative process introduces representational dynamics that diverge from human creative cognition.

B. Layer Depth Analysis

• Alignment is driven by higher layers of the network.
• There is a significant positive correlation ($r=0.54, p<0.05$) between the relative depth of the best-aligning layer and the alignment score. This confirms that creative thinking is mirrored by abstract, task-sensitive representations rather than early syntactic/lexical processing.

C. Impact of Post-Training Objectives (Llama-3.1-8B Variants)

• Creativity-Optimized (CrPO): Preserves alignment with high-creativity neural responses while reducing alignment with low-creativity responses. This indicates a targeted shift toward the neural geometry of original thought.
• Human Behavior Simulation (Minitaur): Increases alignment with both high and low creativity responses, reflecting a general mimicry of human response patterns.
• Reasoning-Chain (DeepSeek-R1 Distill): Shows a striking reversal. It exhibits negative alignment with high-creativity responses and positive alignment with low-creativity responses. This suggests that Chain-of-Thought training steers representations away from creative ideation toward systematic, analytical processing.
• Control Tasks: No significant correlations were found for the non-creative OCT task or the Somatomotor network, confirming the specificity of the findings to creative cognition.

5. Significance and Implications

• Neural Plausibility of Creativity: The study provides evidence that LLMs can be neurally plausible models of human creative thought, but this alignment is fragile and highly dependent on the training objective.
• The "Convergent Thinking" Trap: Current post-training trends heavily favor convergent thinking (e.g., math, coding, CoT reasoning). The results suggest that these objectives may actively degrade a model's capacity for divergent thinking by shifting its internal representations away from the neural geometry of human creativity.
• Evaluation Paradigm: Brain alignment offers a new, mechanistic lens for evaluating LLMs that goes beyond behavioral benchmarks (which can be gamed). It captures the computational principles underlying human thought.
• Future Directions: To build truly creative AI, future training strategies must explicitly optimize for divergent thinking and diversity, rather than solely focusing on convergent, single-answer reasoning tasks.