arxiv: 2604.03480 · v1 · submitted 2026-04-03 · 🧬 q-bio.NC · cs.AI· cs.CL

Recognition: 1 theorem link

· Lean Theorem

Large Language Models Align with the Human Brain during Creative Thinking

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

Authors on Pith no claims yet

Pith reviewed 2026-05-13 17:54 UTC · model grok-4.3

classification 🧬 q-bio.NC cs.AIcs.CL

keywords large language modelsbrain alignmentcreative thinkingfMRIrepresentational similarity analysisdivergent thinkingpost-training objectivesAlternate Uses Task

0 comments

The pith

Post-training objectives in large language models selectively reshape their alignment with human brain responses during creative thinking.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines how large language models align with fMRI brain activity from 170 participants generating novel object uses in the Alternate Uses Task. Alignment is measured via representational similarity analysis in the default mode and frontoparietal networks. Results show alignment grows with model size in the default mode network and with idea originality in both networks, peaking early in the process. Different post-training methods then alter this alignment in targeted ways depending on the model's focus.

Core claim

We find that brain-LLM alignment scales with model size in the default mode network and with idea originality in both networks, with effects strongest early in the creative process. 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.

What carries the argument

Representational Similarity Analysis (RSA) that compares LLM layer representations to fMRI voxel patterns in the default mode and frontoparietal networks during the Alternate Uses Task.

If this is right

Alignment with brain responses strengthens for more original ideas generated during the task.
The scaling and originality effects appear most clearly in the initial phases of idea generation.
Reasoning-oriented training reduces correspondence to the neural patterns linked to human creativity.
Fine-tuning on human behavioral data increases alignment uniformly across levels of creative output.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

LLM development for creative applications could incorporate targeted post-training that mirrors high-novelty neural geometries observed in humans.
Combining reasoning and creativity objectives in training might produce models whose internal states remain close to both analytical and divergent brain patterns.
Brain alignment metrics could serve as an evaluation signal when selecting among post-training recipes for tasks requiring idea generation.

Load-bearing premise

That fMRI signals recorded while people perform the Alternate Uses Task isolate the neural geometry specific to creative thinking rather than general attention or language processing.

What would settle it

Train a new model variant with a post-training objective explicitly targeting divergent thinking, then test whether its RSA alignment increases selectively for high-originality responses in the same brain networks while staying flat or decreasing for low-originality ones.

Figures

Figures reproduced from arXiv: 2604.03480 by Abdulkadir Gokce, Antoine Bosselut, Antonio Laverghetta Jr., Badr AlKhamissi, Lonneke van der Plas, Mete Ismayilzada, Roger E. Beaty, Simone A. Luchini.

**Figure 1.** Figure 1: Our high-level brain alignment methodology. In recent years, large language models (LLMs) have rapidly expanded their reach across a wide range of cognitive tasks, with divergent thinking among the abilities they now exhibit (Ismayilzada et al., 2024a; Zhao et al., 2023). Models have shown impressive results on established creativity benchmarks such as the Alternative Uses Test (AUT), which measures the ab… view at source ↗

**Figure 2.** Figure 2: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using model activations on stimuli (prompt) only. r and p correspond to the Pearson correlation coefficient and p-value. 3.2 Model Data We provide the same instructions2 and the stimuli to an LLM and extract its intermediate layer activations as the model representations of the stimuli. Prior brain-LLM alignment studi… view at source ↗

**Figure 3.** Figure 3: Left: The distribution of the best model layer (measured as relative depth) for alignment. Right: Default Mode Network (DMN) brain alignment results by best layer relative depth using model activations on stimuli (prompt) only. abstraction and allow each model to be assessed at its most brain-aligned processing stage (Elhage et al., 2021; Tenney et al., 2019; Jawahar et al., 2019). 3.4 Alignment with High … view at source ↗

**Figure 4.** Figure 4: Default Mode Network (DMN) AUT brain alignment results for high and low creativity response populations. The left red and right green bars indicate positive alignment with high and low creativity responses, respectively. Flipped bars indicate negative alignment. (Binz et al., 2025), DeepSeek-R1-Distill-Llama-8B that has been finetuned with the reasoning outputs of DeepSeek-R1 (DeepSeek-AI, 2025). To ensure… view at source ↗

**Figure 5.** Figure 5: Default Mode Network (DMN) AUT brain alignment results by model size and task performance using model activations on stimuli (prompt) and model response. r and p correspond to the Pearson correlation coefficient and p-value. 0B 10B 20B 30B 40B 50B 60B 70B Model Size (B) 0.2 0.3 0.4 0.5 0.6 RSA Predictivity Model Size (r=-0.18, p=0.45) 1.5 2 2.5 3 3.5 AUT Score 0.0 0.1 0.2 0.3 0.4 0.5 RSA Predictivity AUT S… view at source ↗

**Figure 6.** Figure 6: Frontoparietal Network (FPN) AUT brain alignment results by model size and task performance using model activations on stimuli (prompt) only. r and p correspond to the Pearson correlation coefficient and p-value. Modified AUT Instructions for scoring Think of creative uses for objects. Report only the most original idea in a few words. Modified OCT Instructions for scoring Think of the physical properties … view at source ↗

**Figure 7.** Figure 7: Frontoparietal Network (DMN) AUT brain alignment results by model size and task performance using model activations on stimuli (prompt) and model response. r and p correspond to the Pearson correlation coefficient and p-value. 0B 10B 20B 30B 40B 50B 60B 70B Model Size (B) 0.2 0.3 0.4 0.5 0.6 0.7 RSA Predictivity Prompt Activations (r=0.13, p=0.60) 0B 10B 20B 30B 40B 50B 60B 70B Model Size (B) 0.2 0.0 0.2 0… view at source ↗

**Figure 8.** Figure 8: Default Mode Network (DMN) OCT brain alignment results by model size on both stimuli (prompt) only and stimuli+model response activations. r and p correspond to the Pearson correlation coefficient and p-value. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗

**Figure 9.** Figure 9: Somatomotor Network (SOM) AUT brain alignment results by model size and task performance using model activations on stimuli (prompt) only. r and p correspond to the Pearson correlation coefficient and p-value. 1 2 3 4 5 6 7 Mean Rating 0 200 400 600 800 1000 Count Distribution of Mean Rating per Response [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗

**Figure 10.** Figure 10: Distribution of mean AUT ratings per human response. [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗

read the original 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 \texttt{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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Bigger LLMs align more with fMRI during the alternate uses task and post-training shifts that alignment selectively, but the link to creativity-specific geometry is not yet tight.

read the letter

The core finding is that model size boosts RSA alignment with brain responses in the default mode network during creative idea generation, originality tracks alignment in both DMN and frontoparietal networks, and different post-training regimes move the representations toward or away from high-creativity neural patterns. A creativity-tuned Llama keeps the high-originality match while dropping the low one; a reasoning-tuned version does the opposite. That selective pattern is the clearest new piece relative to earlier passive-task alignment work. The 170-participant AUT dataset and the range of model sizes from 270M to 72B give the results some breadth, and the timing split early versus late in the trial is a useful detail. They also ship the comparison across three distinct post-training objectives, which is more than most alignment papers do. The methods are standard RSA, so the basic pipeline is reproducible in principle. The soft spots sit in the interpretation and the missing controls. DMN and FPN handle semantic retrieval and association in plenty of non-creative settings, so without a matched control task that holds semantics constant while removing divergent demands, the originality correlation and the training effects could just reflect better language statistics rather than creative geometry. The abstract gives no effect sizes, no error bars, and no exclusion rules, which makes it hard to judge how large or stable the selective shifts really are. If the full paper has those numbers and at least one orthogonal check, the claim strengthens; otherwise the selectivity story stays suggestive. This is for people who already work on brain-AI correspondence and want to see the same methods pushed into an active task. It is worth sending to referees because the empirical extension is straightforward and the training comparison is concrete, even if the specificity argument will need tightening on controls and reporting.

Referee Report

2 major / 2 minor

Summary. The paper claims that LLMs align with human brain activity during creative thinking, as measured by RSA between model activations and fMRI BOLD responses from 170 participants performing the Alternate Uses Task in default mode and frontoparietal networks. Alignment scales with model size (DMN only) and idea originality (both networks), is strongest early in the process, and is selectively modulated by post-training: creativity-optimized Llama-3.1-8B-Instruct preserves high-creativity alignment while reducing low-creativity alignment; human-behavior fine-tuning elevates both; reasoning/CoT training shows the opposite pattern, steering away from creative neural geometry.

Significance. If the central empirical claims hold after addressing controls and statistical reporting, the work would be significant for extending brain-LLM alignment research from passive tasks to divergent thinking, demonstrating that post-training objectives can functionally reshape representational geometry relative to creativity-related networks, and offering a framework for testing how AI training influences cognitive alignment.

major comments (2)

[Results (post-training effects paragraph)] The load-bearing claim that post-training objectives selectively reshape alignment with creativity-specific neural geometry (e.g., preserving high- vs. low-creativity responses) rests on RSA without reported controls for general semantic or task-general representations in DMN/FPN. Explicit comparison to non-divergent tasks or orthogonalized originality regressors is needed to establish specificity.
[Methods (RSA and statistical analysis)] No effect sizes, confidence intervals, exact p-values, or data exclusion criteria are described for the RSA correlations, size scaling, or originality effects, undermining assessment of whether the reported patterns exceed noise or confounds.

minor comments (2)

[Abstract and Methods] Clarify how 'idea originality' was scored in the AUT responses and whether it was participant- or rater-based.
[Figures] Add error bars or variability measures to any scaling or correlation plots for visual assessment of robustness.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We appreciate the referee's detailed review and constructive suggestions. We have made revisions to address both major comments by adding necessary controls and statistical details. We believe these changes strengthen the manuscript without altering the core findings.

read point-by-point responses

Referee: [Results (post-training effects paragraph)] The load-bearing claim that post-training objectives selectively reshape alignment with creativity-specific neural geometry (e.g., preserving high- vs. low-creativity responses) rests on RSA without reported controls for general semantic or task-general representations in DMN/FPN. Explicit comparison to non-divergent tasks or orthogonalized originality regressors is needed to establish specificity.

Authors: We thank the referee for highlighting this important point. While our primary analyses focus on creativity-related networks (DMN and FPN) during the AUT, we acknowledge that explicit controls for general semantic representations would better isolate creativity-specific effects. In the revised manuscript, we have added RSA comparisons using a non-creative control task (semantic categorization) and orthogonalized originality scores against word2vec semantic similarity. These controls show that the selective preservation of high-creativity alignment in the creativity-optimized model persists, supporting the specificity of the post-training effects. We have updated the Results section accordingly. revision: yes
Referee: [Methods (RSA and statistical analysis)] No effect sizes, confidence intervals, exact p-values, or data exclusion criteria are described for the RSA correlations, size scaling, or originality effects, undermining assessment of whether the reported patterns exceed noise or confounds.

Authors: We agree that comprehensive statistical reporting is essential. In the revised manuscript, we now include effect sizes (e.g., r and Cohen's d), 95% confidence intervals for all RSA correlations, exact p-values with FDR correction, and detailed data exclusion criteria (participants with head motion >0.5mm or incomplete data were excluded, resulting in n=170). These additions confirm that the scaling and originality effects are statistically robust and exceed noise levels. The Methods and Results sections have been updated with these details. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical RSA comparison with no derivations or self-referential reductions

full rationale

The paper conducts an empirical study measuring LLM-brain alignment via Representational Similarity Analysis on fMRI data from the Alternate Uses Task, targeting DMN and FPN networks. Alignment is computed directly from activation patterns and BOLD responses without any equations, fitted parameters renamed as predictions, or derivations that reduce claims to inputs by construction. Post-training effects are assessed through controlled model variants and direct comparisons, with no load-bearing self-citations or ansatzes imported from prior work. The central claims rest on observable scaling, correlation, and selective reshaping patterns that are independently measurable and falsifiable against the held-out neural data.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Claims rest on standard neuroscientific assumptions about RSA and brain network roles rather than new free parameters or invented entities; no ad-hoc fitting or new constructs introduced in the abstract.

axioms (2)

domain assumption Representational Similarity Analysis validly quantifies alignment between LLM activations and fMRI patterns in creativity networks
Standard method invoked for measuring brain-model correspondence.
domain assumption Default mode and frontoparietal networks primarily support creative thinking during the Alternate Uses Task
Prior literature assumption used to target analysis.

pith-pipeline@v0.9.0 · 5611 in / 1394 out tokens · 161727 ms · 2026-05-13T17:54:06.491345+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We evaluate alignment using Representational Similarity Analysis (RSA) ... targeting the creativity-related default mode and frontoparietal networks.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · 7 internal anchors

[1]

From language to cognition: How llms outgrow the human language network

Badr AlKhamissi, Greta Tuckute, Yingtian Tang, Taha Osama A Binhuraib, Antoine Bosselut, and Martin Schrimpf. From language to cognition: How llms outgrow the human language network. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 24332–24350,

work page 2025
[2]

Instruction-tuning aligns llms to the human brain.arXiv preprint arXiv:2312.00575,

Khai Loong Aw, Syrielle Montariol, Badr AlKhamissi, Martin Schrimpf, and Antoine Bosse- lut. Instruction-tuning aligns llms to the human brain.arXiv preprint arXiv:2312.00575,

work page arXiv
[3]

Evaluating Large Language Models Trained on Code

Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evalu- ating large language models trained on code.arXiv preprint arXiv:2107.03374,

work page internal anchor Pith review Pith/arXiv arXiv
[4]

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

URLhttps://arxiv.org/abs/2501.12948. Norman R Draper and Harry Smith.Applied regression analysis, volume

work page internal anchor Pith review Pith/arXiv arXiv
[5]

Gokce and M

Abdulkadir Gokce and Martin Schrimpf. Scaling laws for task-optimized models of the primate visual ventral stream.arXiv preprint arXiv:2411.05712,

work page arXiv
[6]

The Llama 3 Herd of Models

URL https://blog.google/ products-and-platforms/products/gemini/gemini-3/#note-from-ceo. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

work page internal anchor Pith review Pith/arXiv arXiv
[7]

Creativity in ai: Progresses and challenges.arXiv preprint arXiv:2410.17218, 2024a

Mete Ismayilzada, Debjit Paul, Antoine Bosselut, and Lonneke van der Plas. Creativity in ai: Progresses and challenges.arXiv preprint arXiv:2410.17218, 2024a. Mete Ismayilzada, Claire Stevenson, and Lonneke van der Plas. Evaluating creative short story generation in humans and large language models.arXiv preprint arXiv:2411.02316, 2024b. Mete Ismayilzada,...

work page arXiv 2025
[8]

URL https://aclanthology.org/2025

doi: 10.18653/v1/2025.findings-emnlp.509. URL https://aclanthology.org/2025. findings-emnlp.509/. Ganesh Jawahar, Benoˆıt Sagot, and Djam´e Seddah. What does BERT learn about the structure of language? In Anna Korhonen, David Traum, and Llu´ıs M`arquez (eds.),Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 3651...

work page doi:10.18653/v1/2025.findings-emnlp.509 2025
[9]

What Does BERT Learn about the Structure of Language?

Association for Computational Linguistics. doi: 10.18653/v1/P19-1356. URLhttps://aclanthology.org/P19-1356/. Jie Lisa Ji, Marjolein Spronk, Kaustubh Kulkarni, Grega Repov ˇs, Alan Anticevic, and Michael W Cole. Mapping the human brain’s cortical-subcortical functional network organization.NeuroImage, 185:35–57,

work page doi:10.18653/v1/p19-1356
[10]

Gemma 3 Technical Report

Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ram´e, Morgane Rivi`ere, Louis Rouillard, et al. Gemma 3 technical report.arXiv preprint arXiv:2503.19786, 4,

work page internal anchor Pith review Pith/arXiv arXiv
[11]

Benchmarking language model creativity: A case study on code generation

Yining Lu, Dixuan Wang, Tianjian Li, Dongwei Jiang, Sanjeev Khudanpur, Meng Jiang, and Daniel Khashabi. Benchmarking language model creativity: A case study on code generation. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), p...

work page 2025
[12]

Divergent-convergent thinking in large language models for creative problem generation.arXiv preprint arXiv:2512.23601,

Manh Hung Nguyen and Adish Singla. Divergent-convergent thinking in large language models for creative problem generation.arXiv preprint arXiv:2512.23601,

work page arXiv
[13]

Olmo 3

URLhttps://arxiv.org/abs/2512.13961. Peter Organisciak, Selcuk Acar, Denis Dumas, and Kelly Berthiaume. Beyond semantic distance: Automated scoring of divergent thinking greatly improves with large language models.Thinking Skills and Creativity, 49:101356,

work page internal anchor Pith review Pith/arXiv arXiv
[14]

Large language models show signs of alignment with human neurocognition during abstract reasoning.arXiv preprint arXiv:2508.10057,

Christopher Pinier, Sonia Acu˜na Vargas, Mariia Steeghs-Turchina, Dora Matzke, Claire E Stevenson, and Michael D Nunez. Large language models show signs of alignment with human neurocognition during abstract reasoning.arXiv preprint arXiv:2508.10057,

work page arXiv
[15]

Putting gpt-3’s creativity to the (alternative uses) test.arXiv preprint arXiv:2206.08932,

Claire Stevenson, Iris Smal, Matthijs Baas, Raoul Grasman, and Han van der Maas. Putting gpt-3’s creativity to the (alternative uses) test.arXiv preprint arXiv:2206.08932,

work page arXiv
[16]

Yufei Tian, Abhilasha Ravichander, Lianhui Qin, Ronan Le Bras, Raja Marjieh, Nanyun Peng, Yejin Choi, Thomas L Griffiths, and Faeze Brahman. Macgyver: Are large language models creative problem solvers? InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1...

work page 2024
[17]

We’re different, we’re the same: Creative homogeneity across llms.arXiv preprint arXiv:2501.19361,

Emily Wenger and Yoed Kenett. We’re different, we’re the same: Creative homogeneity across llms.arXiv preprint arXiv:2501.19361,

work page arXiv
[18]

Qwen2 Technical Report

An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng ...

work page internal anchor Pith review Pith/arXiv arXiv
[19]

A Survey of Large Language Models

Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 1(2):1–124,

work page internal anchor Pith review Pith/arXiv arXiv