Pith. sign in

REVIEW 5 major objections 6 minor 47 references

A Unified Geometric Space for Topological Alignment Between Transformer-Based Models and Human Brain Networks

T0 review · 5 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A seven-dimensional Brain-like Space maps 151 Transformer models by attention-head topology to seven resting-state brain networks.

desk verdict Large, well-intentioned mapping effort, but the RoPE model results are built on borrowed positional embeddings and the content-free assumption is untested, so the headline arc geometry is provisional. read the letter →

arxiv 2510.24342 v3 pith:L2EG3NGT submitted 2025-10-28 cs.AI

Silin Chen , Yuzhong Chen , Caiwei Wang , Zifan Wang , Junhao Wang , Zifeng Jia , Keith M Kendrick , Tuo Zhang
show 4 more authors
Lin Zhao Dezhong Yao Tianming Liu Xi Jiang
This is my paper · ORCID
classification cs.AI
keywords Brain-likeSpaceTransformerattentiontopologyfunctionalbrainnetworksbrain-AIalignmentgraph-theoreticsimilaritypositionalencodingself-supervisedvisionmodelsresting-statefMRI
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that the internal organization of any Transformer-based model can be placed on a single geometric axis shared with the human brain, without running any input through the model. It does this by comparing the graph topology of each attention head with seven canonical resting-state brain networks and arranging the 151 models in a seven-dimensional "Brain-like Space." The reported result is an arc-shaped distribution: models trained for global semantic abstraction, and language-dominant models, sit near higher-order cognitive networks, while models trained to reconstruct local detail sit near sensory networks. If the claim holds, it would give researchers a task-free, modality-agnostic way to ask whether an artificial system organizes information the way a brain does, and to compare vision, language, and multimodal models on the same terms.

What carries the argument

The load-bearing construction is the spatial attention graph, defined without any input content as the softmax of position-based query-key products in Eq. (2), with the absolute positional embedding matrix and learned query and key matrices as the only inputs; for rotary-encoding models the paper borrows absolute embeddings and rescales them to match weight statistics. Each graph is described by five graph-theoretic metrics, and the cosine similarity of that five-dimensional signature to the signatures of the seven canonical functional brain networks becomes one coordinate. The seven coordinates place every attention head as a point in the Brain-like Space, on which PCA and k-means reveal the arc and four clusters.

What would settle it

Rebuild the space using content-dependent attention matrices obtained by feeding real tokens through the same 151 models, keeping the same five graph metrics and seven network targets; if the reported arc and four clusters do not reproduce, the positional-only graph is not capturing the organization the paper claims to measure.

Watch

Extended reading notes

Core claim

On the paper's own terms, every attention head of every Transformer model is embedded as a point in a seven-dimensional space whose coordinates are cosine similarities between a five-metric graph signature (clustering, modularity, degree spread, shortest paths, global efficiency) and the corresponding signatures of seven functional brain networks extracted from resting-state fMRI of 1,042 participants. Across 62,480 attention heads from 151 models, the points form a continuous arc that the authors divide into four clusters of increasing brain-likeness. Language-dominant models concentrate in the most brain-like cluster; vision models split by pretraining, with global-semantic variants (DINO, DINOv3, BEiT, AugReg-trained ViTs) near the cognitive end and local-reconstructive variants (MAE, DeiT3, DINOv2) near the sensory end. Multimodal models with rotary positional encoding fuse their vision and language components toward the cognitive end, while learnable-position models separate them. Fine-tuning and instruction tuning change the positions only slightly, and brain-likeness scores are not significantly correlated with ImageNet Top-1 accuracy (r = 0.266, p = 0.156).

Load-bearing premise

An attention graph is drawn only from position information and learned query and key weights, with no actual input content, based on the assumption that a model's spatial interaction structure is set by its positional encoding; if content-dependent attention dominates real Transformers, the space measures positional schemes instead of model organization.

Editorial extensions

If this is right

  • Any Transformer-based model can be scored for brain-likeness without processing any input, so vision, language, and multimodal models become directly comparable on a single organizational axis.
  • Training objectives that push global semantic abstraction—self-distillation with semantic consistency, masked token prediction, Mixup-based augmentation—move vision models toward higher-order cognitive networks, while pixel-reconstruction and CNN-distillation objectives push them toward sensory networks.
  • RoPE-based multimodal models show deep fusion of vision and language components in the space, whereas models with learnable positional encodings show a division of labor, with vision components shifting to less brain-like clusters.
  • Brain-likeness is not a proxy for ImageNet performance: the reported correlation is weak and non-significant (r = 0.266, p = 0.156), so organizational similarity and downstream task accuracy can part ways.
  • Fine-tuning and instruction tuning leave a model's brain-like organization largely unchanged, suggesting that the organization is set during pretraining.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension, not stated by the paper, is to use the space as an architecture-level diagnostic: if randomly initialized Transformers already occupy meaningful positions, the arc is set by architecture and positional encoding rather than learning.
  • The same construction could be applied to non-attention architectures by substituting their spatial interaction graphs; if CNNs' fixed receptive fields land near the sensory end, the global-sensory dichotomy would look like a property of architecture rather than training alone.
  • Because the alignment score tracks training objectives but not ImageNet accuracy, it could serve as a cheap, task-free monitor for drift toward local-detail processing during fine-tuning or distillation.
  • The non-significant accuracy correlation does not establish a causal trade-off between brain-likeness and performance; it only shows the two measures can vary independently in the sampled vision models.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper introduces 'Brain-like Space', a seven-dimensional space in which each attention head of a Transformer model is embedded by the cosine similarities between five graph-theoretic metrics of its spatial attention graph and those of seven canonical functional brain networks derived from resting-state fMRI. Analyzing 151 Transformer-based models with 62,480 attention head graphs, the authors report a continuous arc-shaped geometry with four clusters (C1–C4) reflecting increasing brain-likeness. They attribute the distribution to data augmentation, training objectives, distillation, and positional encoding schemes, and report that the brain-likeness score does not correlate significantly with ImageNet accuracy (r=0.266, p=0.156). The framework is proposed as a task-free, modality-agnostic measure of organizational similarity between artificial and biological systems.

Significance. If the framework withstood scrutiny, it would provide a novel, input-free and modality-agnostic tool for comparing Transformer architectures with human brain networks, and the scale of the empirical survey (151 models, 62,480 head graphs, 1042 subjects) is a clear strength. The paper makes code and data availability explicit, and its descriptive observations (e.g., the global-semantic vs. local-reconstruction dichotomy) are internally consistent. However, the central claims currently rest on an unvalidated content-free assumption about positional encoding, a borrowed-positional-embedding substitution for RoPE models, and circular analysis choices (PC1, thresholds, k); these issues must be addressed before the results can be interpreted as evidence about intrinsic model organization.

major comments (5)
  1. [Methods, Eq. (2)] The spatial attention graph is constructed from positional embeddings and query/key matrices only, with no input content. The paper explicitly adopts the hypothesis from ref [47] that a model's intrinsic spatial interaction structure is primarily determined by its positional encoding scheme. This assumption is load-bearing: if content-dependent attention contributes substantially to the topology of real Transformer representations, the Brain-like Space measures a property of positional embeddings rather than of model organization. The manuscript does not validate this assumption (e.g., by comparing the content-free graphs with content-dependent attention maps on a subset of models), and Sec. 3.6 does not list it as a limitation. Please provide such a validation or reframe the claims accordingly.
  2. [Methods, 'Construction of Spatial Attention Graphs' and Eqs. (3)-(4)] For RoPE-based models, the paper substitutes a borrowed absolute positional embedding Pbase from GPT-2 or ViT-base, linearly interpolated and rescaled, instead of using the model's own RoPE. Because Pbase is identical for all RoPE language models (and separately for all RoPE vision encoders), the reported convergence of RoPE LLMs near C3/C4 in Sec. 2.5 and the 'deep fusion' effect of RoPE may be artifacts of the shared borrowed Pbase rather than evidence about the models' own organization. Please either use the actual RoPE (e.g., with a fixed probe input) or report a sensitivity analysis over different Pbase choices; without this, the positional-encoding conclusions are not supported.
  3. [Sec. 2.6, Eq. (14)] The brain-likeness score is defined as the sum of projections onto PC1 of the Brain-like Space, and PC1 is fit to all 151 models' attention heads. Likewise, the matched-head thresholds (0.8, lowered to 0.6) and the k=4 clustering are selected using the same dataset. This creates a circularity: the 'gradual increase' from C1 to C4 and the model rankings are partly determined by the very data being scored. Please use leave-one-model-out or nested construction of PC1 and thresholds, or otherwise demonstrate that the ranking is stable when the axis is defined independently.
  4. [Sec. 2.3 and Fig. 2e] The matched-head threshold is lowered from 0.8 to 0.6 specifically for DINOv2 and MAE, and the resulting matches are 'almost exclusively' VIS. A lowered post-hoc threshold does not provide evidence for a specific brain-network alignment; at the standard threshold these models have no matches. The claim of 'reduced alignment in DINOv2 compared to its predecessors' is therefore weaker than stated. Please report matches at a fixed threshold or justify the threshold change a priori.
  5. [Methods, Eq. (12)] The 'standardized vector' used in the cosine similarity is not defined. It is unclear whether standardization is z-scoring across the five graph metrics, across heads, or across brain networks, and this choice affects every similarity value and hence the entire Brain-like Space. Please specify the standardization procedure and justify its invariance properties.
minor comments (6)
  1. [Title] The arXiv title 'A Unified Geometric Space for Topological Alignment Between Transformer-Based Models and Human Brain Networks' differs from the first-page title 'A Unified Geometric Space Bridging AI Models and the Human Brain'; please align them.
  2. [Eq. (1)] The same symbol W is used on both sides of the definition; please use a different symbol for the normalized edge weights to avoid confusion.
  3. [Eq. (5)] The notation '~A(l) ij' is not defined, and '10 −5' for epsilon and delta should be written as '10^{-5}'.
  4. [Sec. 2.1] The k-means clustering input is not specified: please state whether clustering is performed on the seven-dimensional vectors or on the PC scores, and describe the feature space used for the silhouette analysis.
  5. [Sec. 2.6 and Abstract] The p-value for the ImageNet correlation is reported as 0.156 in the abstract and 0.1555 in Sec. 2.6; please standardize the rounding.
  6. [Sec. 2.6] The criteria for selecting the 30 representative vision models are not given; please describe the selection procedure or state that it is a convenience sample.

Circularity Check

3 steps flagged · score 5.0 of 10

RoPE-model convergence is partly manufactured by a shared borrowed GPT-2/ViT-base Pbase; the graph construction rests on a self-cited ansatz, and the brain-likeness score is a PC1-fitted quantity, so the headline geometry is only partially self-contained.

  1. fitted input called prediction [Methods, 'Construction of Spatial Attention Graphs in Artificial Neural Network Models' (Eqs 2-4); Results 2.5, Fig. 3a]
    "For models using the RoPE positional encoding scheme that is inherently coupled with input content, we adopt a pragmatic approach by using absolute positional embeddings from the GPT-2 or ViT-base as the base embedding matrix Pbase ... As shown in Fig. 3a, eight representative RoPE-based LLMs ... exhibit a similar distribution in the Brain-like Space with their centroids located near the boundary between C3 and C4."

    In Eq (2), the spatial attention graph is built from P, W_Q, and W_K. For all RoPE LLMs the paper inserts the same GPT-2 absolute positional embedding as Pbase, and for all RoPE vision encoders the same ViT-base Pbase (interpolated and rescaled). Consequently, the reported convergence of RoPE LLMs near C3/C4 and the high C4 assignment of LMM-vision-RoPE are partly produced by this shared borrowed input rather than by each model's own organization. The later explanation that RoPE 'providing a continuous, shared geometric prior for all tokens ... enabling deep fusion' invokes a mechanism that was never entered into Eq (2) for these models, so the RoPE effect is, at least in part, a renaming of the common Pbase rather than an empirical effect of the models' own RoPE.

  2. fitted input called prediction [Methods, 'Brain-like Space Construction and Brain-likeness Score Calculation' (Eq 14); Results 2.6, Fig. 4a]
    "the brain-likeness score is defined and calculated by summing the projection values of all attention heads in a model onto the first principal component axis of the Brain-like Space: brain-likeness score = Σ_l Σ_h (s^(l,h) · u_PC1) ... Correlation between the brain-likeness score and ImageNet-1k Top-1 accuracy across 30 representative vision models."

    The brain-likeness score is defined as the sum of projections onto u_PC1, the first principal component of the Brain-like Space. That space is constructed from the very 151 models analyzed, including the 30 vision models later correlated with ImageNet-1k accuracy. Thus the score is a data-fitted summary, not an independent measure, and the reported correlation (r=0.266, p=0.156) is an in-sample correlation between a PC1-derived score and accuracy on the same models used to fit PC1. The paper presents this as evidence about the relationship between brain-likeness and task performance, but by construction the score cannot serve as an independent predictor.

1 more flagged steps
  1. ansatz smuggled in via citation [Methods, 'Construction of Spatial Attention Graphs in Artificial Neural Network Models' (Eq 2); Discussion 3.6]
    "Following the hypothesis in the previous study [47], we assume that the intrinsic spatial interaction structure of an artificial neural network model is primarily determined by its positional encoding scheme."

    This is the sole justification for dropping input content from Eq (2), and it is attributed to ref [47], a prior work whose authors overlap with the present paper. The paper provides no independent validation or sensitivity analysis for this hypothesis, and Sec 3.6 lists limitations without flagging the content-free assumption or the foreign-Pbase substitution for RoPE models. The Sec 2.5 conclusion that positional encoding schemes shape brain-like organization is therefore substantially inherited from a self-cited ansatz rather than demonstrated from the present data.

full rationale

The brain-network data (rs-fMRI, Yeo-7 networks), the model weights from timm/Transformers, and the ImageNet accuracies are external, so the arc-shaped geometry is not wholly circular and retains independent descriptive content. However, two load-bearing quantities are defined using author-inserted shared inputs or in-sample fits: all RoPE models are assigned a common borrowed GPT-2/ViT-base Pbase in Eq (2), making their apparent convergence partly an artifact of that shared input; and the brain-likeness score of Eq (14) projects onto a PC1 fitted on the same 151 models, so the correlation with ImageNet accuracy is not an out-of-sample test. The graph construction also rests on the same-group ansatz from ref [47] that spatial interaction structure is primarily determined by positional encoding, which is asserted rather than validated. These issues are partial rather than total circularity: the global-vs-local pretraining-paradigm results and the external brain-network similarities are not reduced to their inputs by construction. Score 5 reflects one central result (RoPE convergence/deep fusion) that is substantially manufactured by a shared input, plus a data-fitted score and a load-bearing self-citation.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The central claim depends on several domain assumptions and data-fitted choices, most importantly the content-free attention graph construction (Eq 2, based on self-cited ref [47]) and the PCA and threshold definitions of brain-likeness. The external brain data provide some grounding, but the distinctive findings are obtained from internal, data-fitted components.

free parameters (5)
  • Matched-head similarity threshold = 0.8 (default), 0.6 (DINOv2 and MAE)
    Used to decide when an attention head is matched to a brain network; the lower threshold is applied post hoc for DINOv2 and MAE in Sec 2.3, Fig 2e.
  • Number of clusters k = 4
    Chosen by Silhouette analysis on the same 62,480 heads (Sec 2.1, Fig 1b); cluster boundaries are data-dependent.
  • PC1 axis of Brain-like Space = First principal component of all attention-head vectors
    The brain-likeness score is the sum of projections onto this fitted axis (Eq 14); the axis is not an independent reference and would change with the model cohort.
  • Positional embedding scaling factor k = (sigma_Q + sigma_K) / (2 sigma_P)
    Ad hoc rescaling in Eqs (3) and (4) to attach a borrowed absolute positional embedding to RoPE models; no theoretical justification is given.
  • Borrowed positional embedding P_base = GPT-2 (length 50) for language, ViT-base (length 197) for vision, linearly interpolated
    RoPE is content-coupled, so the paper substitutes an absolute embedding from another model (Methods, Construction of Spatial Attention Graphs); the choice is pragmatic rather than derived.
assumptions (5)
  • ad hoc to paper The intrinsic spatial interaction structure of a Transformer model is primarily determined by its positional encoding scheme.
    Invoked before Eq (2) in Methods; the supporting citation is ref [47], whose authors overlap with the present paper, and no content-dependent validation is provided.
  • domain assumption The seven Yeo-7 networks, mapped from the DK atlas via Dice coefficients, faithfully represent the canonical intrinsic connectivity networks of the human brain.
    Used to define the seven dimensions of the Brain-like Space in Sec 2.1 and Methods 'Functional Brain Network Mapping'; the atlas choice is standard but still a modeling decision.
  • domain assumption Five graph-theoretic metrics (clustering coefficient, modularity, degree standard deviation, path length, global efficiency) capture the organizational topology of both brain and model graphs.
    Methods 'Graph-Theoretic Metric Definition'; these metrics are standard, but their sufficiency for brain-AI similarity is assumed rather than established.
  • domain assumption Cosine similarity of the five-metric feature vectors is a valid topological similarity measure across graphs of different sizes.
    Methods Eq (12); graphs range from 50 to 197 nodes for models and 136 nodes for brains, and no size correction or standardization is specified.
  • domain assumption Softmax normalization of brain connectivity (Eq 1) makes brain graphs formally comparable to attention maps.
    Methods 'Functional Brain Network Mapping'; this design choice affects all downstream cosine similarities and is not independently validated.
invented entities (2)
  • Brain-like Space
    purpose: A seven-dimensional embedding in which each dimension is the cosine similarity between an attention-head graph and one of the seven functional brain networks (Eq 13); it situates models for task-free comparison.
    It is a proposed metric construction introduced here; no external benchmark or falsifiable prediction independently validates it beyond the paper's own analyses.
  • Brain-likeness score
    purpose: Sum over all attention heads of the projection of each head's seven-dimensional vector onto PC1 of the Brain-like Space (Eq 14); used to rank models and correlate with ImageNet accuracy.
    The score depends on PCA axes fit to the same 62,480 heads, making it an internal, data-fitted measure with no independent calibration.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Unified Geometric Space for Topological Alignment Between Transformer-Based Models and Human Brain Networks." pith.science (2026). https://pith.science/paper/L2EG3NGT

@misc{pith2026251024342,
  author       = {Pith},
  title        = {Pith review of: A Unified Geometric Space for Topological Alignment Between Transformer-Based Models and Human Brain Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L2EG3NGT}},
  note         = {Machine review of arXiv:2510.24342}
}
read the original abstract

Whether artificial neural networks organize information comparably to the human brain remains unclear. Prior brain--AI alignment studies are constrained by specific inputs and tasks, limiting cross-modal comparison. Here we introduce a brain--model topological alignment space, mapping Transformer attention topology onto human intrinsic connectivity networks (ICNs) to enable task-free, modality-agnostic comparison. Analyzing 151 Transformer-based models with 62,480 attention head graphs, we observe a continuous arc-shaped distribution reflecting varying alignment. Models optimized for global semantics aligned with higher-order ICNs, while local-detail models aligned with sensory ICNs. Non-intuitive findings include reduced alignment in DINOv2 compared to its predecessors and a counterintuitive scaling inversion in distilled DeiT models, while fine-tuning and instruction tuning had limited effect. Alignment scores showed no significant correlation with ImageNet accuracy (r = 0.266, p = 0.156). This work offers a quantitative framework for comparing the organizational principles of artificial and biological systems.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 16 canonical work pages

  1. [47]

    3-Augment

    Chen, Y., Xiao, Z., Du, Y., Zhao, L., Zhang, L., Wu, Z., Zhu, D., Zhang, T., Y ao, D., Hu, X., et al. : A unified and biologically plausible relational graph representa- tion of vision transformers. IEEE Transactions on Neural Networks and Learnin g Systems (2024) https://doi.org/10.1109/tnnls.2023.3342810 30 A Unified Geometric Space Bridging AI Models and...

    arXiv 2024

  2. [1]

    Elsevier, New York and London (1973)

    Jerison, H.: Evolution of the Brain and Intelligence. Elsevier, New York and London (1973)

    work page 1973

  3. [2]

    The bulletin of mathematical biophysics 5(4), 115–133 (1943) https: //doi.org/10.1007/bf02478259

    McCulloch, W.S., Pitts, W.: A logical calculus of the ideas immane nt in nervous activity. The bulletin of mathematical biophysics 5(4), 115–133 (1943) https: //doi.org/10.1007/bf02478259

    doi:10.1007/bf02478259 1943

  4. [3]

    The Journal of physiology 148(3), 574 (1959) https://doi.org/10.1113/ jphysiol.1959.sp006308

    Hubel, D.H., Wiesel, T.N.: Receptive fields of single neurones in the cat’s striate cortex. The Journal of physiology 148(3), 574 (1959) https://doi.org/10.1113/ jphysiol.1959.sp006308

    work page 1959

  5. [4]

    Proceedings of the IEEE 86(11), 2278–2324 (1998) https: //doi.org/10.1109/5.726791

    LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning ap plied to document recognition. Proceedings of the IEEE 86(11), 2278–2324 (1998) https: //doi.org/10.1109/5.726791

    doi:10.1109/5.726791 1998

  6. [5]

    nature 521(7553), 436–444 (2015) https://doi.org/10.1145/3355047.3359415

    LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436–444 (2015) https://doi.org/10.1145/3355047.3359415

    work page arXiv 2015

  7. [6]

    : Coupling visual semantics of artificial neural networks and human brain function via synchronized activations

    Zhao, L., Dai, H., Wu, Z., Xiao, Z., Zhang, L., Liu, D.W., Hu, X., Jiang, X., Li, S., Zhu, D., et al. : Coupling visual semantics of artificial neural networks and human brain function via synchronized activations. IEEE Transacti ons on Cognitive and Developmental Systems 16(2), 584–594 (2023) https://doi.org/10. 1109/tcds.2023.3287184

    arXiv 2023

  8. [7]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Liu, X., Zhou, M., Shi, G., Du, Y., Zhao, L., Wu, Z., Liu, D., Liu, T., Hu, X.: Coupling artificial neurons in bert and biological neurons in the human br ain. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, pp. 8888–8896 (2023). https://doi.org/10.1609/aaai.v37i7.26068

    work page doi:10.1609/aaai.v37i7.26068 2023

Show all 47 references
  1. [8]

    Nature Machine Intelligence 6(12), 1467–1477 (2024) https://doi.org/10.1038/ s42256-024-00925-4

    Mischler, G., Li, Y.A., Bickel, S., Mehta, A.D., Mesgarani, N.: Conte xtual fea- ture extraction hierarchies converge in large language models and the brai n. Nature Machine Intelligence 6(12), 1467–1477 (2024) https://doi.org/10.1038/ s42256-024-00925-4

    2024

  2. [9]

    In: Advances in Neural Information Processing S ystems (2025)

    An, J., Song, Y., Yang, R., Ding, N., Lu, L., Wang, Y., Wang, W., Zhuang, C., Wang, Q., Fang, F.: Hierarchical frequency tagging probe (hftp): A un ified approach to investigate syntactic structure representations in large language mod- els and the human brain. In: Advances in ...

    2025

  3. [10]

    Nature Machine Intelligence, 1–15 (2025) https://doi.org/10

    Doerig, A., Kietzmann, T.C., Allen, E., Wu, Y., Naselaris, T., Kay, K., Charest, I.: High-level visual representations in the human brain are aligned wi th large language models. Nature Machine Intelligence, 1–15 (2025) https://doi.org/10. 1038/s41467-024-53147-y

    2025

  4. [11]

    Nature computational science, 1–11 (2025) https://doi.org/10.1038/s43588-025-00863-0

    Gao, C., Ma, Z., Chen, J., Li, P., Huang, S., Li, J.: Increasing alignmen t of large language models with language processing in the human brain. Nature computational science, 1–11 (2025) https://doi.org/10.1038/s43588-025-00863-0

    2025 doi

  5. [12]

    Nature human behaviour, 1–15 (2025) https://doi.org/10.1038/s41562-025-02105-9

    Goldstein, A., Wang, H., Niekerken, L., Schain, M., Zada, Z., Aubrey, B ., Sheffer, T., Nastase, S.A., Gazula, H., Singh, A., et al.: A unified acoustic-to-spee ch- to-language embedding space captures the neural basis of natural language processing in everyday conversations. Na...

    2025 doi

  6. [13]

    Nature Machine Intelligence, 1–16 (2025) https://doi.org/10.1038/s42256-025-01049-z

    Du, C., Fu, K., Wen, B., Sun, Y., Peng, J., Wei, W., Gao, Y., Wang, S., Zhang, C., Li, J., et al.: Human-like object concept representations emerge n aturally in multimodal large language models. Nature Machine Intelligence, 1–16 (2025) https://doi.org/10.1038/s42256-025-01049-z

    2025 doi

  7. [14]

    arXiv preprint arXiv:2106.10270 (2021) arXiv:2106.10270 [cs.CV]

    Steiner, A., Kolesnikov, A., Zhai, X., Wightman, R., Uszkoreit, J., Beyer, L.: How to train your vit? data, augmentation, and regularization in vision transfor mers. arXiv preprint arXiv:2106.10270 (2021) arXiv:2106.10270 [cs.CV]

    2021 arXiv

  8. [15]

    In: Eu ropean Conference on Computer Vision, pp

    Touvron, H., Cord, M., J´ egou, H.: Deit iii: Revenge of the vit. In: Eu ropean Conference on Computer Vision, pp. 516–533 (2022). https://doi.org/10.1007/ 978-3-031-20053-3 30 . Springer

    2022

  9. [16]

    Science Advanc es 10(39) (2024) https://doi.org/10.1126/sciadv.adl1776

    Prince, J.S., Alvarez, G.A., Konkle, T.: Contrastive learning ex plains the emer- gence and function of visual category-selective regions. Science Advanc es 10(39) (2024) https://doi.org/10.1126/sciadv.adl1776

    2024 doi

  10. [17]

    Proceedings of the National Academy of Sciences 118(3) (2021) https://doi.org/ 10.1073/pnas.2014196118

    Zhuang, C., Yan, S., Nayebi, A., Schrimpf, M., Frank, M.C., DiCarl o, J.J., Yamins, D.L.K.: Unsupervised neural network models of the ventral v isual stream. Proceedings of the National Academy of Sciences 118(3) (2021) https://doi.org/ 10.1073/pnas.2014196118

    2021 doi

  11. [18]

    Nature Communications 13(1) (2022) https://doi.org/10.1038/s41467-022-28091-4

    Konkle, T., Alvarez, G.A.: A self-supervised domain-general learn ing framework for human ventral stream representation. Nature Communications 13(1) (2022) https://doi.org/10.1038/s41467-022-28091-4

    2022 doi

  12. [19]

    arXiv preprint (2025) arXiv:2508.18226 [cs.AI]

    Raugel, J., Szafraniec, M., Vo, H.V., Couprie, C., Labatut, P., Bojanow ski, P., Wyart, V., King, J.-R.: Disentangling the factors of convergence betwe en brains and computer vision models. arXiv preprint (2025) arXiv:2508.18226 [cs.AI]

    2025 arXiv

  13. [20]

    In: Proceedings 27 of the IEEE/CVF International Conference on Computer Vision, pp

    Caron, M., Touvron, H., Misra, I., J´ egou, H., Mairal, J., Bojanowski, P ., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings 27 of the IEEE/CVF International Conference on Computer Vision, pp. 9650–9660 (2021). https://doi.org/10.1109/ic...

    2021

  14. [21]

    arXiv preprint (2021) arXiv:2106.08254 [cs.CV]

    Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-training of image tran sformers. arXiv preprint (2021) arXiv:2106.08254 [cs.CV]

    2021 arXiv

  15. [22]

    arXiv preprint (2022) arXiv:2208.06366 [cs.CV]

    Peng, Z., Dong, L., Bao, H., Ye, Q., Wei, F.: Beit v2: Masked image modeli ng with vector-quantized visual tokenizers. arXiv preprint (2022) arXiv:2208.06366 [cs.CV]

    2022 arXiv

  16. [23]

    arXiv preprint (2023) arXiv:2304.07193 [cs.CV]

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalid ov, V., Fer- nandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning rob ust visual features without supervision. arXiv preprint (2023) arXiv:2304.07193 [cs.CV]

    2023 arXiv

  17. [24]

    In: Proceedings of the IEEE/CVF Conferen ce on Computer Vision and Pattern Recognition, pp

    He, K., Chen, X., Xie, S., Li, Y., Doll´ ar, P., Girshick, R.: Masked au toencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conferen ce on Computer Vision and Pattern Recognition, pp. 16000–16009 (2022). https://doi. org/10.1109/cvpr52688.2022.01553

    2022

  18. [25]

    OpenAI Blog

    Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I ., et al.: Language models are unsupervised multitask learners. OpenAI Blog. Acces sed: 2023-01-01 (2019). https://openai.com/blog/better-language-models/

    2019

  19. [26]

    Devlin, J., Chang, M.-W., Lee, K., Toutanova, K.: Bert: Pre-train ing of deep bidirectional transformers for language understanding. In: Proceedin gs of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolog...

    2019 doi

  20. [27]

    : Learning transferable visual models from natural language supervision

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sast ry, G., Askell, A., Mishkin, P., Clark, J., et al. : Learning transferable visual models from natural language supervision. In: International Conference on Machine Lear ning, pp. 8748–8763 (2021). PMLR...

    2021 arXiv

  21. [28]

    In: Internati onal Con- ference on Machine Learning, pp

    Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre- training for unified vision-language understanding and generation. In: Internati onal Con- ference on Machine Learning, pp. 12888–12900 (2022). PMLR. https://arxiv.org/ abs/2201.12086

    2022 arXiv

  22. [29]

    Beyer, L., H´ enaff, O.J., Kolesnikov, A., Zhai, X., Oord, A.v.d.: Are we done with imagenet? arXiv preprint (2020) arXiv:2006.07159 [cs.CV]

    2020 arXiv

  23. [30]

    Margulies, D.S., Ghosh, S.S., Goulas, A., Falkiewicz, M., Huntenb urg, J.M., Langs, G., Bezgin, G., Eickhoff, S.B., Castellanos, F.X., Petrides, M. , et al. : Situating the default-mode network along a principal gradient of macros cale cor- tical organization. Proceedings of th...

    2016 doi

  24. [31]

    In: W orld Con- ference on Explainable Artificial Intelligence, pp

    Dorszewski, T., Tˇ etkov´ a, L., Jenssen, R., Hansen, L.K., Wickstrøm, K.K.: From colors to classes: Emergence of concepts in vision transformers. In: W orld Con- ference on Explainable Artificial Intelligence, pp. 28–47 (2025). Springer. https: //arxiv.org/abs/2503.24071

    2025 arXiv

  25. [32]

    arXiv preprint (2024) arXiv:2405.14880 [cs.CV]

    Pan, X., Philip, A., Xie, Z., Schwartz, O.: Dissecting query-ke y interaction in vision transformers. arXiv preprint (2024) arXiv:2405.14880 [cs.CV]

    2024 arXiv

  26. [33]

    arXiv preprint (2022) arXiv:2206.07682 [cs.CL]

    Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogat ama, D., Bosma, M., Zhou, D., Metzler, D., et al.: Emergent abilities of lar ge language models. arXiv preprint (2022) arXiv:2206.07682 [cs.CL]

    2022 arXiv

  27. [34]

    : Instruction- tuning aligns llms to the human brain

    Aw, K.L., Montariol, S., AlKhamissi, B., Schrimpf, M., Bosselut, A. : Instruction- tuning aligns llms to the human brain. CoRR (2023) arXiv:2312.00575 [cs.CL]

    2023 arXiv

  28. [35]

    Schrimpf, M., Kubilius, J., Hong, H., Majaj, N.J., Rajalingham, R. , Issa, E.B., Kar, K., Bashivan, P., Prescott-Roy, J., Geiger, F., et al.: Brain-s core: Which artificial neural network for object recognition is most brain-like? Bi oRxiv, 407007 (2018) https://doi.org/10.1101/407007

    2018 doi

  29. [36]

    arXiv preprint (2025) arXiv:2504.16940 [q-bio.NC]

    Linsley, D., Feng, P., Serre, T.: Better artificial intelligenc e does not mean better models of biology. arXiv preprint (2025) arXiv:2504.16940 [q-bio.NC]

    2025 arXiv

  30. [37]

    CoRR (2024) arXiv:2406.01538 [cs.CL]

    Feghhi, E., Hadidi, N., Song, B., Blank, I., Kao, J.C.: What are large language models mapping to in the brain? a case against over-reliance on brain sc ores. CoRR (2024) arXiv:2406.01538 [cs.CL]

    2024 arXiv

  31. [38]

    Neuron 95(2), 245–258 (2017) https://doi.org/10

    Hassabis, D., Kumaran, D., Summerfield, C., Botvinick, M.: Neurosc ience- inspired artificial intelligence. Neuron 95(2), 245–258 (2017) https://doi.org/10. 1016/j.neuron.2017.06.011

    2017

  32. [39]

    arXiv preprint (2025) arXiv:2507.10722 [q-bio.NC]

    Shankar, S., Pan, Y., Jiang, H., Liu, Z., Darbandi, M.R., Lorenzo, A., Chen , J., Hasan, M.M., Zidan, A.H., Gelman, E., et al.: Bridging brains and machines : A unified frontier in neuroscience, artificial intelligence, and neu romorphic systems. arXiv preprint (2025) arXiv:2507...

    2025 arXiv

  33. [40]

    : The wu-minn human connectome project: an overview

    Van Essen, D.C., Smith, S.M., Barch, D.M., Behrens, T.E., Yac oub, E., Ugurbil, K., Consortium, W.-M.H., et al. : The wu-minn human connectome project: an overview. Neuroimage 80, 62–79 (2013) https://doi.org/10.1016/j.neuroimage. 2013.05.041

    2013 doi

  34. [41]

    : Increasing diversity in connectomics with the chinese human connectome project

    Ge, J., Yang, G., Han, M., Zhou, S., Men, W., Qin, L., Lyu, B., Li, H., Wan g, H., Rao, H., et al. : Increasing diversity in connectomics with the chinese human connectome project. Nature Neuroscience 26(1), 163–172 (2023) https://doi.org/ 10.1038/s41593-022-01215-1 29

    2023 doi

  35. [42]

    : The minimal preprocessing pipelines for the human connectome project

    Glasser, M.F., Sotiropoulos, S.N., Wilson, J.A., Coalson, T.S., Fisc hl, B., Ander- sson, J.L., Xu, J., Jbabdi, S., Webster, M., Polimeni, J.R., et al. : The minimal preprocessing pipelines for the human connectome project. Neuroi mage 80, 105–124 (2013) https://doi.org/10.101...

    2013 doi

  36. [43]

    Psychoradiology 1(1), 23–41 (2021) https://doi.org/10.1093/psyrad/ kkab002

    Jiang, X., Zhang, T., Zhang, S., Kendrick, K.M., Liu, T.: Fundamental functional differences between gyri and sulci: implications for brain function , cognition, and behavior. Psychoradiology 1(1), 23–41 (2021) https://doi.org/10.1093/psyrad/ kkab002

    2021 doi

  37. [44]

    : The organiza- tion of the human cerebral cortex estimated by intrinsic functional c onnectivity

    Yeo, B.T., Krienen, F.M., Sepulcre, J., Sabuncu, M.R., Lashkar i, D., Hollinshead, M., Roffman, J.L., Smoller, J.W., Z¨ ollei, L., Polimeni, J.R., et al. : The organiza- tion of the human cerebral cortex estimated by intrinsic functional c onnectivity. Journal of neurophysiolog...

    2011

  38. [45]

    , Ramachandran, S.C., Pisner, D.A., Frank, P.F., Lemmer, A.D., Nikolaidis, A., Vogelst ein, J.T.: Standardizing human brain parcellations

    Lawrence, R.M., Bridgeford, E.W., Myers, P.E., Arvapalli, G.C. , Ramachandran, S.C., Pisner, D.A., Frank, P.F., Lemmer, A.D., Nikolaidis, A., Vogelst ein, J.T.: Standardizing human brain parcellations. Scientific Data 8(1) (2021) https://doi. org/10.1038/s41597-021-00849-3

    2021 doi

  39. [46]

    Neuroimage 237, 118164 (2021) https: //doi.org/10.1016/j.neuroimage.2021.118164

    Dworetsky, A., Seitzman, B.A., Adeyemo, B., Neta, M., Coalson, R.S., P etersen, S.E., Gratton, C.: Probabilistic mapping of human functional brain net works identifies regions of high group consensus. Neuroimage 237, 118164 (2021) https: //doi.org/10.1016/j.neuroimage.2021.118164

    2021

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.