Large Language Models Align with the Human Brain during Creative Thinking

Mete Ismayilzada; Simone A. Luchini; Abdulkadir Gokce; Badr AlKhamissi; Antoine Bosselut; Antonio Laverghetta; Lonneke van der Plas; Roger E. Beaty

q-bio.NCApr 3, 2026

Large Language Models Align with the Human Brain during Creative Thinking

Authors: Mete Ismayilzada, Simone A. Luchini, Abdulkadir Gokce, Badr AlKhamissi, Antoine Bosselut, Antonio Laverghetta, Lonneke van der Plas, Roger E. Beaty

Organizations: EPFL · Universit`a della Svizzera italiana (USI) · Paris Brain Institute (ICM) · Pennsylvania State University ·

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 is positively associated with model size (default mode network only) and with idea originality (both networks), with these relationships clearest early in the creative process. We further find that post-training objectives are associated with functionally selective differences in alignment: a creativity-optimized Llama-3.1-8B-Instruct retains alignment with high-creativity neural responses while lacking the positive low-creativity alignment present in other variants, but a reasoning-trained variant shows the opposite asymmetry, with negative alignment to high-creativity responses. Together, these results suggest that brain alignment offers an informative lens on how post-training relates to the neural geometry of human creative thought.

Figures & tables

Figure 1: Our high-level brain alignment methodology. Participants perform a creative task (AUT) and a matched non-creative control (OCT) during fMRI, while LLMs receive the same instructions and stimuli. We extract layer-wise activations at the prompt stage and after response generation, and compute alignment as the rank correlation between per-subject fMRI and LLM representational dissimilarity matrices (RDMs). Alignment is evaluated in the default mode (DMN) and frontoparietal (FPN) networks, plus the somatomotor network (SOM) as a control, and then compared between high- and low-creativity human responses.

Figure 2: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using prompt-only model activations.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 3: The distribution of best model layer (measured by relative depth) for DMN AUT brain alignment on prompt-only model activations.

Figure 4: Default Mode Network (DMN) AUT brain alignment results for high and low creativity response populations (rating cutoff = 2.0). The left red and right gray bars indicate positive alignment with high and low creativity responses, respectively. Flipped bars indicate negative alignment. Results for rating cutoffs

1.75

and

2.25

can be found in Appendix Figures 10 and 11 respectively.

Appendix figures & tables45 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Default Mode Network (DMN) AUT Gemma family brain alignment results by model size and task performance using prompt-only model activations.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 6: Default Mode Network (DMN) AUT Gemma family brain alignment results by model size and task performance using prompt+response model activations.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 7: Default Mode Network (DMN) AUT Qwen family brain alignment results by model size and task performance using prompt-only model activations.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 8: Number of low- and high-creativity responses per AUT stimulus at the

2.0

cutoff, sorted by the proportion of high-creativity responses. All 46 stimuli contribute to both buckets.

Figure 9: Per-subject sample sizes in the low- and high-creativity buckets at the

2.0

cutoff. Each point is one subject; the dashed line marks equal counts. The anti-diagonal band reflects the fixed number of AUT trials per subject.

Figure 10: Default Mode Network (DMN) AUT brain alignment results for high and low creativity response populations (rating cutoff =

1.75

). The left red and right gray bars indicate positive alignment with high and low creativity responses, respectively. Flipped bars indicate negative alignment.

Figure 11: Default Mode Network (DMN) AUT brain alignment results for high and low creativity response populations (rating cutoff =

2.25

). The left red and right gray bars indicate positive alignment with high and low creativity responses, respectively. Flipped bars indicate negative alignment.

		last-token			mean-token
model	pop	median	95% CI	alignment	median	95% CI	alignment
crpo-dpo-llama-3.1-8b-instruct	high	+0.320	[-0.119,+0.507]	$\pm$	+0.104	[-0.208,+0.384]	$\pm$
crpo-dpo-llama-3.1-8b-instruct	low	+0.281	[+0.040,+0.530]	$+$	+0.159	[-0.127,+0.453]	$\pm$
crpo-llama-3.1-8b-instruct-cre	high	+0.158	[-0.200,+0.460]	$\pm$	+0.132	[-0.189,+0.415]	$\pm$
crpo-llama-3.1-8b-instruct-cre	low	+0.104	[-0.162,+0.336]	$\pm$	+0.007	[-0.206,+0.097]	$-$
crpo-sft-llama-3.1-8b-instruct	high	+0.270	[+0.010,+0.542]	$+$	+0.018	[-0.273,+0.200]	$\pm$

Appendix

Table 1: Default Mode Network (DMN) AUT alignment results (prompt+response activations) per model and creativity population with rating cutoff=

1.75

		last-token			mean-token
model	pop	median	95% CI	alignment	median	95% CI	alignment
crpo-dpo-llama-3.1-8b-instruct	high	+0.307	[+0.023,+0.673]	$+$	+0.312	[+0.114,+0.620]	$+$
crpo-dpo-llama-3.1-8b-instruct	low	+0.348	[+0.040,+0.726]	$+$	+0.159	[-0.167,+0.453]	$\pm$
crpo-llama-3.1-8b-instruct-cre	high	+0.284	[-0.050,+0.495]	$+$	+0.080	[-0.125,+0.244]	$\pm$
crpo-llama-3.1-8b-instruct-cre	low	+0.077	[-0.256,+0.298]	$\pm$	+0.007	[-0.271,+0.096]	$-$
crpo-sft-llama-3.1-8b-instruct	high	+0.581	[+0.303,+0.789]	$+$	+0.375	[-0.009,+0.770]	$+$

Appendix

Table 2: Default Mode Network (DMN) AUT alignment results (prompt+response activations) per model and creativity population with rating cutoff=

2.0

		last-token			mean-token
model	pop	median	95% CI	alignment	median	95% CI	alignment
crpo-dpo-llama-3.1-8b-instruct	high	+0.467	[-0.049,+1.000]	$+$	+0.091	[-0.282,+0.479]	$\pm$
crpo-dpo-llama-3.1-8b-instruct	low	+0.454	[+0.115,+0.862]	$+$	+0.253	[-0.184,+0.670]	$\pm$
crpo-llama-3.1-8b-instruct-cre	high	+0.366	[-0.228,+0.666]	$\pm$	+0.088	[-0.135,+0.629]	$\pm$
crpo-llama-3.1-8b-instruct-cre	low	+0.112	[-0.183,+0.360]	$\pm$	-0.040	[-0.252,+0.077]	$-$
crpo-sft-llama-3.1-8b-instruct	high	+0.633	[+0.324,+0.848]	$+$	+0.454	[-0.173,+0.769]	$\pm$

Appendix

Table 3: Default Mode Network (DMN) AUT alignment results (prompt+response activations) per model and creativity population with rating cutoff=

2.25

Figure 12: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using prompt-only model activations with last-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 13: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using prompt-only model activations with mean-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 14: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using prompt+response model activations with last-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 15: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using prompt+response model activations with mean-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 16: Frontoparietal Network (FPN) AUT brain alignment results by model size and task performance using prompt-only model activations with last-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 17: Frontoparietal Network (FPN) AUT brain alignment results by model size and task performance using prompt-only model activations with mean-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 18: Frontoparietal Network (FPN) AUT brain alignment results by model size and task performance using prompt+response model activations with last-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 19: Frontoparietal Network (FPN) AUT brain alignment results by model size and task performance using prompt+response model activations with mean-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 20: Somatomotor Network (SOM) AUT brain alignment results by model size and task performance using prompt-only model activations with last-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 21: Somatomotor Network (SOM) AUT brain alignment results by model size and task performance using prompt-only model activations with mean-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 22: Somatomotor Network (SOM) AUT brain alignment results by model size and task performance using prompt+response model activations with last-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 23: Somatomotor Network (SOM) AUT brain alignment results by model size and task performance using prompt+response model activations with mean-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 24: Default Mode Network (DMN) OCT brain alignment results by model size with last-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 25: Default Mode Network (DMN) OCT brain alignment results by model size with mean-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 26: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using empty prompt-only model activations with last-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value. Note that Falcon-40B-Instruct result is missing due to the failure to extract representations from an empty prompt.

Figure 27: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using empty prompt-only model activations with mean-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value. Note that Falcon-40B-Instruct result is missing due to the failure to extract representations from an empty prompt.

Figure 28: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using empty prompt+response model activations with last-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value. Note that Falcon-40B-Instruct result is missing due to the failure to extract representations from an empty prompt.

Figure 29: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using empty prompt+response model activations with mean-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value. Note that Falcon-40B-Instruct result is missing due to the failure to extract representations from an empty prompt.

Figure 30: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using placeholder prompt-only model activations with last-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 31: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using placeholder prompt-only model activations with mean-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 32: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using placeholder prompt+response model activations with last-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 33: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using placeholder prompt+response model activations with mean-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 34: The distribution of best model layer (measured by relative depth) for DMN AUT brain alignment on prompt+response model activations.

Figure 35: The distribution of best model layer (measured by relative depth) and its correlation with DMN AUT brain alignment on prompt-only model activations with last-token pooling.

Figure 36: The distribution of best model layer (measured by relative depth) and its correlation with DMN AUT brain alignment on prompt-only model activations with mean-token pooling.

Figure 37: The distribution of best model layer (measured by relative depth) and its correlation with DMN AUT brain alignment on prompt+response model activations with last-token pooling.

Figure 38: The distribution of best model layer (measured by relative depth) and its correlation with DMN AUT brain alignment on prompt+response model activations with mean-token pooling.

Figure 39: Default Mode Network (DMN) AUT brain alignment results by mean response generation length using prompt+response model activations with last-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 40: Default Mode Network (DMN) AUT brain alignment results by mean response generation length using prompt+response model activations with mean-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 41: Default Mode Network (DMN) AUT brain alignment results by median response generation length using prompt+response model activations with last-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 42: Default Mode Network (DMN) AUT brain alignment results by median response generation length using prompt+response model activations with mean-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 43: Frontoparietal network (FPN) OCT brain alignment results by model size with last-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 44: Frontparietal network (FPN) OCT brain alignment results by model size with mean-token pooling.

r

and

p

correspond to the Pearson correlation coefficient and p-value.

Figure 45: Distribution of mean AUT ratings per human response.

Model	Mean	Median	Std
llama-3.2-1b-instruct	325.27	309.50	87.26
llama-3.2-3b-instruct	328.85	314.00	91.82
llama-3.1-8b	148.34	8.00	292.45
llama-3.1-8b-instruct	264.61	260.50	61.98
crpo-llama-3.1-8b-instruct-cre	129.39	28.00	273.30
crpo-sft-llama-3.1-8b-instruct	15.71	9.00	65.67

Appendix

Table 4: Generation length (in words) statistics per model over the AUT responses.

LLM Evaluation

Large language model evaluation measures what models can do, how reliably they do it, and whether benchmark scores predict real-world performance.

See all 2,408 papers

CreativityNeuro: Steering Language Model Weights to Improve Divergent Thinking and Reduce Mode Collapse
Jul 1, 2026Samuel Schapiro, Core Francisco Park, Felix Sosa +1Language Model Steering Computational Creativity
Univeristy of Illinois, Urbana-Champaign · Center for Brain Science, Harvard University · CBS-NTT Program in Physics of Intelligence, Harvard University +2
Assessing the Creativity of Large Language Models: Testing, Limits, and New Frontiers
May 13, 2026Samuel Schapiro, Alexi Gladstone, Jonah Black +1Creativity Assessment LLM Evaluation
University of Illinois Urbana-Champaign
Playing with Words, Improving with Rewards: Training Language Models for Creative Association
May 27, 2026Vijeta Deshpande, Namrata Shivagunde, Sherin Muckatira +5Computational Creativity
University of Massachusetts Lowell · Dartmouth College · University of Amsterdam +2

Jul 1, 2026J/K move · Enter open · S save

CreativityNeuro: Steering Language Model Weights to Improve Divergent Thinking and Reduce Mode Collapse

Samuel Schapiro, Core Francisco Park, Felix Sosa, Lav R. Varshney

Univeristy of Illinois, Urbana-Champaign · Center for Brain Science, Harvard University · CBS-NTT Program in Physics of Intelligence, Harvard University+2

Language Model Steering Computational Creativity Mode Collapse

Open paper AskPDF arXiv

Divergent thinking is a crucial aspect of creativity, yet large language models (LLMs) tend to consistently generate similar responses to open-ended questions, in what has been termed the artificial hivemind effect. Here, we introduce CreativityNeuro, a data-free method for enhancing divergent thinking in LLMs via contrastive weight steering. We evaluate our method across multiple creativity assessments and report several main findings. On the Divergent Association Task (DAT), a vocabulary-space creativity test, CreativityNeuro improves performance by up to 14 human percentile points. Next, in a large-scale human evaluation (N=720) on the Alternative Uses Test (AUT) and the Task Task, CreativityNeuro achieves significant improvements in originality, surprise, and creativity, transferring to longer-form and more open-ended tasks. Importantly, we find that across all three tasks, CreativityNeuro demonstrably reduces measures of mode collapse. Moreover, activation steering achieves comparable performance to CreativityNeuro on the DAT, but it does not transfer to the AUT and Task Task, demonstrating the effectiveness of weight-space steering in generalizing to unseen tasks. In conclusion, CreativityNeuro improves divergent thinking and reduces mode collapse without requiring behavioral data, re-training, or gradient-based fine-tuning, providing a straightforward way to enhance LLM performance in creative domains.

Abstract

LLM Evaluation

Explore similar work

CreativityNeuro: Steering Language Model Weights to Improve Divergent Thinking and Reduce Mode Collapse

Assessing the Creativity of Large Language Models: Testing, Limits, and New Frontiers

Playing with Words, Improving with Rewards: Training Language Models for Creative Association