Large Language Models Align with the Human Brain during Creative Thinking
Summary: arXiv:2604.03480v1 Announce Type: cross
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. Recent advancements in artificial intelligence have prompted researchers to explore the intersection of human creativity and large language models (LLMs). This article discusses a new study that investigates the alignment of LLMs with human brain activity during creative thinking tasks.
Background
Large language models have shown remarkable performance in various tasks, including those requiring divergent thinking. Previous studies have indicated that models that perform well on these tasks tend to align closely with human brain activity patterns. However, most research has focused on passive, non-creative tasks, leaving a gap in understanding how these models relate to creative cognition.
Research Objectives
The primary goal of this study was to explore brain alignment during creative thinking by utilizing fMRI data from 170 participants engaging in the Alternate Uses Task (AUT). This task is designed to measure participants’ ability to think divergently by generating multiple uses for a common object.
Methodology
- Participants: 170 individuals were recruited for the study.
- Task: Participants performed the Alternate Uses Task (AUT) to generate creative ideas.
- Data Collection: Functional Magnetic Resonance Imaging (fMRI) was used to capture brain responses during the task.
- Model Analysis: Representations from various LLMs, ranging in size from 270 million to 72 billion parameters, were extracted for comparison.
- Alignment Measurement: Representational Similarity Analysis (RSA) was employed to assess the alignment between LLM outputs and brain responses, focusing on creativity-related brain networks.
Key Findings
The study revealed several significant findings regarding the alignment between LLMs and human brain activity during creative thinking:
- Model Size and Brain Alignment: The alignment between LLMs and the brain’s default mode network scaled with the model size, indicating that larger models may better capture the nuances of creative thought.
- Idea Originality: Alignment with brain responses was strongest for original ideas, highlighting the importance of creativity in the processing of information.
- Early Creative Process: The effects of alignment were most pronounced early in the creative process, suggesting that initial idea generation is critical for measuring creativity.
- Post-training Objectives: Different training objectives influenced the alignment. A creativity-optimized model, \texttt{Llama-3.1-8B-Instruct}, maintained high alignment with creative neural responses while reducing alignment with less creative ones.
- Behavior Fine-tuning: A model fine-tuned on human behavior improved alignment across both high and low creativity responses.
- Reasoning-trained Variant: A model trained for reasoning exhibited a contrasting pattern, indicating that such training may divert representations from creative neural processing towards analytical reasoning.
Conclusion
The findings of this study provide critical insights into how large language models can align with the neural mechanisms underlying human creativity. By understanding these alignments, researchers can better design AI systems that foster creative thought, potentially leading to advancements in fields that require innovative problem-solving. The selective reshaping of LLM representations through targeted training objectives highlights the nuanced relationship between artificial intelligence and human cognition.