Pith. sign in

REVIEW 4 major objections 5 minor 61 references

A Pre-trained Framework for Multilingual Brain Decoding Using Non-invasive Recordings

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A single pre-trained multilingual semantic space enables brain decoding across languages, subjects, and imaging modalities.

desk verdict A credible integration of known components with a headline fairness claim that needs a matched-data control before it can be trusted. read the letter →

arxiv 2506.03214 v1 pith:NHV7FXZE submitted 2025-06-03 q-bio.NC cs.AIcs.CL

Yi Guo , Yihang Dong , Michael Kwok-Po Ng , Shuqiang Wang This is my paper
classification q-bio.NCcs.AIcs.CL
keywords braindecodingmultilingualnon-invasiveneuroimagingunifiedsemanticspacepre-trainedmodellinguisticfairnessmultimodalfusionbrain-computerinterface
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

This paper proposes UniDecoder, a brain-decoding framework that claims to be the first to reconstruct heard speech from non-invasive brain recordings across multiple languages, multiple subjects, and multiple imaging modalities with a single shared model. The central move is to project fMRI, MEG, and EEG signals into a unified semantic space defined by a pre-trained multilingual model, so that brain activity recorded during English, Chinese, French, and Dutch listening becomes comparable in one representation. If the claim holds, brain-computer interfaces need not be rebuilt per language or per person, and data collected from well-resourced languages could be used to improve decoding for underrepresented ones. The paper validates the framework on 159 participants and reports improved fairness scores when other-language data is added under data-limited conditions.

What carries the argument

The argument is carried by the unified semantic space defined by a pre-trained multilingual model (PMM): Bloom-1.1B's token embeddings serve as the common regression and contrastive target for every brain recording. A self-adaptive semantic mapper (SSM) converts each modality's input into this space, using modality-specific pre-trained encoders, an inter-subject harmonization (ISH) module that applies a per-subject linear transform, and weighted feature-level fusion that combines fMRI and MEG. Decoding proceeds by semantic-consistency beam search, which scores each candidate token sequence by summing the PMM's token probability, a penalized mean-squared error against the brain-predicted embedding, and a cosine-similarity reward, with the final text detokenized through the PMM's tokenizer. The composite training loss (directional cosine, token MSE, and CLIP contrastive terms) is what forces recordings from different languages, subjects, and modalities into the same representational geometry, and the new BGEScore metric grounds evaluation in multilingual embeddings.

What would settle it

Fix a data-limited language, add a matched amount of training data from native speakers of that same language, holding the number of subjects, modality, and sample count equal to the cross-lingual addition. If same-language data yields the same or a larger fairness-score gain, the cross-lingual enhancement attributed to the unified semantic space collapses into a plain data-volume effect that any multi-subject framework would produce.

Watch

Extended reading notes

Core claim

The paper's claim is that diverse non-invasive brain recordings can be mapped into a single semantic space shared across languages, subjects, and neuroimaging modalities, and that natural-language text can be generated from that space in any supported language. UniDecoder aligns word-level brain-derived embeddings with the embeddings produced by a pre-trained multilingual model (Bloom-1.1B) using a composite loss of directional cosine similarity, token-level mean squared error, and a CLIP-style contrastive term, then generates text with a semantic-consistency beam search that balances the model's language probability against agreement with the predicted semantic embeddings. The same representation space lets a language-recognition module identify the stimulus language from brain activity alone, lets a per-subject harmonization transformation narrow the gap between shared and subject-specific decoding, and lets fMRI and MEG features be fused, with fusion outperforming either modality alone. Under simulated data scarcity, the paper reports that incorporating data from speakers of other languages raises decoding performance for the target language, and it interprets this cross-lingual enhancement as evidence of linguistic fairness.

Load-bearing premise

The linguistic-fairness claim rests on the assumption that adding brain data from other languages helps a target language through the shared semantic space rather than simply by giving the model more training examples; the cross-lingual experiments are not compared against an equally sized addition of same-language data.

Editorial extensions

If this is right

  • One shared model could replace per-language and per-subject decoders, lowering the cost of moving a brain-computer interface to a new language.
  • Brain data collected from high-resource languages could be reused to support decoding for low-resource languages, reducing the data barrier for underrepresented user groups.
  • Sites could contribute fMRI, MEG, or EEG recordings to a common decoding pipeline regardless of which device they own.
  • The stimulus language can be read off the brain-derived semantic representation itself, so the system can select the output language without user-supplied metadata.
  • Combining fMRI with MEG in the shared space improved word-level accuracy over either modality alone, suggesting that multi-device recording is a practical route to stronger decoding.

Reading between the lines

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

  • The fairness result would be decisive only after a same-language control of equal size; until then, the improvement from cross-lingual enhancement could plausibly come from the extra training volume rather than from the shared semantic space.
  • If the shared PMM space is the real transfer mechanism, the size of the cross-lingual gain should track the PMM's tokenizer and vocabulary coverage per language, so languages with high subword fragmentation, like the Dutch case documented here, should systematically benefit less; that ordering is testable across many languages.
  • Because the outputs stay at the semantic level and the authors explicitly note the lack of exact word-level reconstruction, a natural next step is to fuse the PMM space with phonological or acoustic embeddings to add decoding precision without renouncing multilingual generality.
  • The framework's architecture also suggests a distributed route the paper only sketches: the shared semantic space could serve as the substrate for federated training across sites, so that a rare-language corpus never needs to be centralized.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces UniDecoder, a unified framework that maps non-invasive brain recordings (fMRI, MEG, EEG) into a semantic space defined by the multilingual model Bloom-1.1B, then decodes natural-language text via semantic-consistency beam search. The framework is evaluated on four public datasets covering 159 participants and four languages (English, Chinese, French, Dutch), with additional analyses of multi-subject harmonization, multimodal fusion, language identification, SHAP-based cortical contributions, and cross-lingual enhancement under data-limited conditions. The central claims are that UniDecoder generalizes across languages, subjects, and modalities, and that training on data from multiple languages improves decoding for low-resource languages, thereby promoting linguistic fairness.

Significance. If the central claims were fully supported, the work would be a substantial advance: it is, to my knowledge, the first demonstration of open-vocabulary multilingual decoding from non-invasive recordings, and the use of a pre-trained multilingual model as a shared semantic target is a natural and promising design that avoids the circularity of evaluating on the same embedding model used for training (training uses Bloom-1.1B; evaluation uses BGE-M3). The paper also benefits from a large participant pool, multiple modalities, several linguistic similarity metrics, and an explicit limitation section. However, the most novel claim—cross-lingual enhancement and linguistic fairness—rests on a single experiment (Fig. 5b) that lacks an equivalent-volume same-language control and does not report variance, so the enhancement may be a data-quantity artifact rather than evidence for the shared semantic space.

major comments (4)
  1. [Results, 'UniDecoder enables multi-subject brain decoding and promotes linguistic fairness' (Fig. 5b)] The cross-linguistic enhancement comparison adds other-language data to the 40% target-language training set and compares against a 40% baseline with no added data. Because the added data increases total training set size, the observed fairness-score gains (0.89 to 0.93, 0.90 to 0.92, 0.90 to 0.94) can be explained entirely by data volume. A matched control that adds an equivalent number of same-language samples (e.g., from held-out subjects of the target language) is required to attribute the improvement to cross-lingual alignment in the PMM semantic space; without it, the abstract's 'cross-lingual mapping enhancement' and 'linguistic fairness' conclusions are not established.
  2. [Results, 'UniDecoder enables multi-subject brain decoding and promotes linguistic fairness' (Fig. 5)] Fig. 5 reports single-point fairness scores with no error bars, confidence intervals, or significance tests, despite the protocol using 'randomly reducing 60%' of training samples. The score differences are small (0.02 to 0.04) and may fall within sampling noise; the authors should repeat the data-limited split with multiple random seeds and report paired statistics (e.g., bootstrap or permutation tests) for both the intra-linguistic and cross-linguistic comparisons.
  3. [Discussion section] The claim that 'the PMM's coverage of over 200 languages [17] suggests that UniDecoder could be readily adapted' is inconsistent with the Methods statement that Bloom-1.1B supports 46 languages. Reference [17] concerns scaling machine translation to 200 languages, not Bloom's tokenizer or vocabulary; this overstates the scalability evidence and should be corrected to describe the actual PMM capability or replaced with a PMM that verifiably covers more languages.
  4. [Results, 'Multimodal fusion improves brain decoding performance' (Fig. 6)] The significant multimodal-fusion benefit is based on repeated-measures ANOVA on SMN4Lang (12 participants, one language), while the aggregated four-dataset comparison in Fig. 6c is descriptive only. Since fusion benefit is central to the multimodal-generalization claim, the authors should either add a cross-dataset statistical analysis (e.g., a mixed-effects model with dataset as a random effect) or restrict the claim of significant fusion benefits to the dataset where it is tested.
minor comments (5)
  1. [Methods, 'Language similarity metrics'] The random baseline used for z-score normalization draws 'randomly sampled vectors' (Eq. 7) but does not specify the distribution or whether these vectors are sampled from the same embedding space; please specify this detail.
  2. [Fig. 3e] The language identification accuracy is shown without details on the classifier, train/test split, or chance level; please provide these details.
  3. [Datasets section] The four languages are not uniformly distributed across datasets: SMN4Lang is Chinese-only, Broderick is English-only, and SparrKULee is Dutch-only; only LPPC is multilingual. The paper should clarify that cross-language generalization within a single dataset is assessed only for LPPC.
  4. [General] No code or data availability statement is provided, which limits reproducibility of the framework and of the evaluation metrics, including the proposed BGEScore implementation.
  5. [Methods, 'UniDecoder'] The hyperparameters beta1, beta2, beta3, lambda1, lambda2, and the fusion weights alpha_m are fixed without a sensitivity analysis; providing a short sensitivity or ablation would strengthen confidence in the reported performance.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular steps identified; the cross-lingual enhancement claim has a missing-control issue that is a confound, not a circular reduction.

full rationale

UniDecoder's training target is the Bloom-1.1B embedding matrix derived from stimulus texts (Methods: 'We then applied PMM to extract semantic features from these texts'), while decoding is evaluated against held-out reference text using BGE-M3-based BGEScore plus WER/BLEU-1/METEOR; the evaluation model is separate from the training target, so there is no definitional circularity. The SSM is trained with a composite loss aligning brain representations to PMM embeddings, but the reported decoding success is measured on unseen recordings against text not used in training. No load-bearing premise rests on a self-citation: the references to the corresponding author's prior patents and brain-network papers are peripheral technical citations, and the PMM/brain-alignment premise is supported by independent external work (references [31-34], [44], etc.). The main methodological concern—the cross-lingual fairness result (Fig. 5b) compares adding other-language data against a no-additional-data baseline without a matched same-language data-volume control—is a missing-baseline confound, not a circular reduction; the improved fairness score is an empirical outcome of training on more data, not an identity with the training objective. Therefore no circular step is identifiable.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

The central framework relies on the assumed validity of the PMM semantic space as a common target for brain recordings from different languages and modalities. The hyperparameters are chosen by hand or empirical validation, and the evaluation adopts a newly introduced BGEScore metric based on a separate embedding model (BGE-M3). No new physical or theoretical entities are introduced; the main conceptual construct is the unified semantic space.

free parameters (9)
  • Loss weight beta1 = 0.4
    Weight for the directional cosine similarity loss in Eq. 2, set based on empirical validation.
  • Loss weight beta2 = 0.3
    Weight for the token-level MSE loss in Eq. 2, set based on empirical validation.
  • Loss weight beta3 = 0.3
    Weight for the CLIP contrastive loss in Eq. 2, set based on empirical validation.
  • Beam search weight lambda1 = 0.3
    Penalty weight for embedding MSE in Eq. 6, a tunable hyperparameter.
  • Beam search weight lambda2 = 0.7
    Weight for cosine similarity in Eq. 6, a tunable hyperparameter.
  • Beam width = 15
    Number of candidate sequences kept during beam search, chosen to increase exploration.
  • Context window = 15 tokens
    Number of tokens per segment for semantic feature extraction from the PMM.
  • fMRI TR lag = 4 TRs
    Applied to compensate for hemodynamic delay in fMRI data.
  • Fusion weights alpha_m = not specified
    Weights for modality fusion in Eq. 1; the paper does not state whether these are learned or hand-set.
assumptions (4)
  • domain assumption PMM embeddings capture semantic content shared across languages that is decodable from non-invasive brain recordings.
    The entire framework aligns brain recordings to Bloom-1.1B embeddings (Methods, UniDecoder). If these embeddings do not align with neural representations, the mapping cannot work.
  • domain assumption The random baseline normalization used for z-scoring metrics provides a valid measure of decoding quality.
    Equation 7 in Methods defines normalized similarity relative to 200 random generations. This assumes random generations form a meaningful null distribution.
  • domain assumption The 26 language-related ROIs from the Destrieux atlas contain sufficient information for fMRI decoding.
    Data preprocessing restricts fMRI to these ROIs; if informative regions are excluded, decoding performance would be underestimated, but claims of generalization depend on this selection.
  • ad hoc to paper Bloom-1.1B supports the four languages tested with usable tokenization.
    Methods state Bloom was chosen for 46-language support; the paper later notes Dutch tokenization is poor (Extended Data Fig. 2), and Discussion claims 200-language coverage citing NLLB, not Bloom.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Pre-trained Framework for Multilingual Brain Decoding Using Non-invasive Recordings." pith.science (2026). https://pith.science/paper/NHV7FXZE

@misc{pith2026250603214,
  author       = {Pith},
  title        = {Pith review of: A Pre-trained Framework for Multilingual Brain Decoding Using Non-invasive Recordings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NHV7FXZE}},
  note         = {Machine review of arXiv:2506.03214}
}
read the original abstract

Brain-computer interfaces (BCIs) with speech decoding from brain recordings have broad application potential in fields such as clinical rehabilitation and cognitive neuroscience. However, current decoding methods remain limited to single-language, single-subject, and single neuroimaging modality settings, restricting their clinical applicability and generalizability. Here we propose a joint multilingual, multi-subject and multimodal decoding framework. It maps diverse brain recordings into a unified semantic space defined by a pre-trained multilingual model (PMM), enabling decoding across multiple languages, multiple subjects and multiple neuroimaging modalities. The proposed framework is validated using non-invasive brain recordings from 159 participants across four languages. Experimental results show that it exhibits strong generalization across multilingual, multi-subject, and multimodal settings. More importantly, the proposed framework can promote linguistic fairness, which is vital for underrepresented languages in BCI applications. The unified semantic space enables cross-lingual mapping enhancement, allowing the framework to boost the decoding performance of underrepresented languages, thereby promoting linguistic fairness. Overall, the proposed framework establishes a new potential paradigm for brain decoding, opening new paths for broader applications of BCI.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 59 canonical work pages

  1. [17]

    R.et al.Scaling neural machine translation to 200 languages.Nature 630, 841–846 (2024)

    Costa-juss` a, M. R.et al.Scaling neural machine translation to 200 languages.Nature 630, 841–846 (2024)

    work page 2024

  2. [1]

    S.et al.An accurate and rapidly calibrating speech neuroprosthesis.N

    Card, N. S.et al.An accurate and rapidly calibrating speech neuroprosthesis.N. Engl. J. Med.391, 609–618 (2024). 13

    work page 2024

  3. [2]

    K., Chartier, J

    Anumanchipalli, G. K., Chartier, J. & Chang, E. F. Speech synthesis from neural decod- ing of spoken sentences.Nature568, 493–498 (2019)

    work page 2019

  4. [3]

    Chen, X.et al.A neural speech decod- ing framework leveraging deep learning and speech synthesis.Nat. Mach. Intell.6, 467– 480 (2024)

    work page 2024

  5. [4]

    Commun.16, 3376 (2025)

    Cai, J.et al.Natural language process- ing models reveal neural dynamics of human conversation.Nat. Commun.16, 3376 (2025)

    work page 2025

  6. [5]

    W., Rotondo, A

    Mathis, M. W., Rotondo, A. P., Chang, E. F., Tolias, A. S. & Mathis, A. Decoding the brain: From neural representations to mechanistic models.Cell187, 5814–5832 (2024)

    work page 2024

  7. [6]

    & Shengye, H

    Wang, S., Yu, W., Chenchen, X. & Shengye, H. Visualization method for evaluating brain addiction traits, apparatus, and medium (2024). US Patent 12,093,833

    work page 2024

  8. [7]

    A.et al.Neuroprosthesis for decoding speech in a paralyzed person with anarthria.N

    Moses, D. A.et al.Neuroprosthesis for decoding speech in a paralyzed person with anarthria.N. Engl. J. Med.385, 217–227 (2021)

    work page 2021

Show all 61 references
  1. [8]

    43, 114924 (2024)

    Zhang, D.et al.A brain-to-text framework for decoding natural tonal sentences.Cell Rep. 43, 114924 (2024)

    2024

  2. [9]

    R., Avansino, D

    Willett, F. R., Avansino, D. T., Hochberg, L. R., Henderson, J. M. & Shenoy, K. V. High- performance brain-to-text communication via handwriting.Nature593, 249–254 (2021)

    2021

  3. [10]

    B., Littlejohn, K

    Silva, A. B., Littlejohn, K. T., Liu, J. R., Moses, D. A. & Chang, E. F. The speech neu- roprosthesis.Nat. Rev. Neurosci.25, 473–492 (2024)

    2024

  4. [11]

    L., Piech, D

    Shen, K., Chen, O., Edmunds, J. L., Piech, D. K. & Maharbiz, M. M. Translational opportunities and challenges of invasive elec- trodes for neural interfaces.Nat. Biomed. Eng 7, 424–442 (2023)

    2023

  5. [12]

    InChinese Conference on Biometric Recognition375–384 (CCBR 2023)

    Li, Y.et al.Generative AI enables the detec- tion of autism using EEG signals. InChinese Conference on Biometric Recognition375–384 (CCBR 2023)

    2023

  6. [13]

    & Huth, A

    Tang, J., LeBel, A., Jain, S. & Huth, A. G. Semantic reconstruction of continuous lan- guage from non-invasive brain recordings. Nat. Neurosci.26, 858–866 (2023)

    2023

  7. [14]

    & King, J.-R

    D´ efossez, A., Caucheteux, C., Rapin, J., Kabeli, O. & King, J.-R. Decoding speech perception from non-invasive brain recordings. Nat. Mach. Intell.5, 1097–1107 (2023)

    2023

  8. [15]

    & Huth, A

    Tang, J. & Huth, A. G. Semantic lan- guage decoding across participants and stim- ulus modalities.Curr. Biol.35, 1023–1032 (2025)

    2025

  9. [16]

    Stavisky, S. D. Restoring speech using brain- computer interfaces.Annu. Rev. Biomed. Eng.27, 29–54 (2025)

    2025

  10. [18]

    B.et al.A bilingual speech neuro- prosthesis driven by cortical articulatory rep- resentations shared between languages.Nat

    Silva, A. B.et al.A bilingual speech neuro- prosthesis driven by cortical articulatory rep- resentations shared between languages.Nat. Biomed. Eng.8, 977–991 (2024)

    2024

  11. [19]

    J.et al.Decoding individual identity from brain activity elicited in imagin- ing common experiences.Nat

    Anderson, A. J.et al.Decoding individual identity from brain activity elicited in imagin- ing common experiences.Nat. Commun.11, 5916 (2020)

    2020

  12. [20]

    & Wang, S

    Kong, H., Pan, J., Shen, Y. & Wang, S. Adversarial learning based structural brain- network generative model for analyzing mild cognitive impairment. InChinese Conference on Pattern Recognition and Computer Vision 361–375 (PRCV 2022)

    2022

  13. [21]

    & Wang, S

    Gong, C., Chen, X., Mughal, B. & Wang, S. Addictive brain-network identification by spa- tial attention recurrent network with feature selection.Brain Informatics10, 2 (2023)

    2023

  14. [22]

    K.et al.Representation of internal speech by single neurons in human supramarginal gyrus.Nat

    Wandelt, S. K.et al.Representation of internal speech by single neurons in human supramarginal gyrus.Nat. Hum. Behav.8, 1136–1149 (2024)

    2024

  15. [23]

    & Wang, S

    You, S., Shen, Y., Wu, G. & Wang, S. Brain MR images super-resolution with the consis- tent features. InInternational Conference on Machine Learning and Computing501–506 (ICMLC ’22)

  16. [24]

    Wang, S.et al.Generative AI enables EEG super-resolution via spatio-temporal adaptive diffusion learning.IEEE Transactions on Consumer ElectronicsDOI–10 (2025)

    2025

  17. [25]

    EEG and MEG: rele- vance to neuroscience.Neuron80, 1112–1128 (2013)

    Lopes da Silva, F. EEG and MEG: rele- vance to neuroscience.Neuron80, 1112–1128 (2013). 14

    2013

  18. [26]

    Neurosci.23, 1484– 1495 (2020)

    Bijsterbosch, J.et al.Challenges and future directions for representations of functional brain organization.Nat. Neurosci.23, 1484– 1495 (2020)

    2020

  19. [27]

    M.et al.Dynamic statistical para- metric mapping: Combining fmri and meg for high-resolution imaging of cortical activity

    Dale, A. M.et al.Dynamic statistical para- metric mapping: Combining fmri and meg for high-resolution imaging of cortical activity. Neuron26, 55–67 (2000)

    2000

  20. [28]

    & Zhang, W

    Wang, S., Yanyan, S. & Zhang, W. Enhanced generative adversarial network and target sample recognition method (2024). US Patent 12,154,036

    2024

  21. [29]

    Jing, C.et al.Estimating addiction-related brain connectivity by prior-embedding graph generative adversarial networks.IEEE Trans- actions on Cybernetics54, 5026–5039 (2024)

    2024

  22. [30]

    A., Fouragnan, E., Retzler, C

    Pisauro, M. A., Fouragnan, E., Retzler, C. & Philiastides, M. G. Neural correlates of evi- dence accumulation during value-based deci- sions revealed via simultaneous eeg-fmri.Nat. Commun.8, 15808 (2017)

    2017

  23. [31]

    A., Bickel, S., Mehta, A

    Mischler, G., Li, Y. A., Bickel, S., Mehta, A. D. & Mesgarani, N. Contextual feature extraction hierarchies converge in large lan- guage models and the brain.Nat. Mach. Intell.6, 1467–1477 (2024)

    2024

  24. [32]

    Y., Kay, K., Naselaris, T., Tarr, M

    Wang, A. Y., Kay, K., Naselaris, T., Tarr, M. J. & Wehbe, L. Better models of human high-level visual cortex emerge from natural language supervision with a large and diverse dataset.Nat. Mach. Intell.5, 1415–1426 (2023)

    2023

  25. [33]

    Neurosci.26, 2213–2225 (2023)

    Li, Y.et al.Dissecting neural computations in the human auditory pathway using deep neu- ral networks for speech.Nat. Neurosci.26, 2213–2225 (2023)

    2023

  26. [34]

    Commun.15, 2768 (2024)

    Goldstein, A.et al.Alignment of brain embed- dings and artificial contextual embeddings in natural language points to common geometric patterns.Nat. Commun.15, 2768 (2024)

    2024

  27. [36]

    https://doi.org/10.1109/TMI.2025

    Yao, W.et al.CATD: Unified representation learning for EEG-to-fMRI cross-modal gener- ation.IEEE Transactions on Medical Imaging (2025). https://doi.org/10.1109/TMI.2025. 3550206

    2025 doi

  28. [37]

    Lundberg, S. M. & Lee, S.-I. A unified approach to interpreting model predictions. InAdvances in Neural Information Processing Systems, 4766–4775 (NeurIPS 2017)

    2017

  29. [38]

    & Zong, C

    Wang, S., Zhang, X., Zhang, J. & Zong, C. A synchronized multimodal neuroimaging dataset for studying brain language process- ing.Sci. Data9, 590 (2022)

    2022

  30. [39]

    Data9, 530 (2022)

    Li, J.et al.Le Petit Prince multilingual naturalistic fMRI corpus.Sci. Data9, 530 (2022)

    2022

  31. [40]

    P., Anderson, A

    Broderick, M. P., Anderson, A. J., Di Liberto, G. M., Crosse, M. J. & Lalor, E. C. Elec- trophysiological correlates of semantic dissim- ilarity reflect the comprehension of natural, narrative speech.Curr. Biol.28, 803–809 (2018)

    2018

  32. [41]

    Accou, B.et al.Sparrkulee: A speech-evoked auditory response repository from KU Leuven, containing the EEG of 85 participants.Data 9, 94 (2024)

    2024

  33. [42]

    & Zhu, W

    Papineni, K., Roukos, S., Ward, T. & Zhu, W. J. BLEU: a method for automatic evalu- ation of machine translation. InProceedings of the 40th annual meeting of the Association for Computational Linguistics311–318 (ACL 2002)

    2002

  34. [43]

    & Lavie, A

    Banerjee, S. & Lavie, A. METEOR: An auto- matic metric for mt evaluation with improved correlation with human judgments. InPro- ceedings of the ACL Workshop on Intrin- sic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization 65–72 (ACL 2005)

    2005

  35. [44]

    InFindings of the Association for Computational Linguistics 2318–2335 (ACL 2024)

    Chen, J.et al.M3-embedding: Multi-linguality, multi-functionality, multi- granularity text embeddings through self-knowledge distillation. InFindings of the Association for Computational Linguistics 2318–2335 (ACL 2024)

    2024

  36. [45]

    Neurosci.25, 1014–1019 (2022)

    Malik-Moraleda, S.et al.An investigation across 45 languages and 12 language families reveals a universal language network.Nat. Neurosci.25, 1014–1019 (2022)

    2022

  37. [46]

    Fedorenko, E., Ivanova, A. A. & Regev, T. I. The language network as a natural kind within the broader landscape of the human brain. 15 Nat. Rev. Neurosci.25, 289–312 (2024)

    2024

  38. [47]

    Jing, C.et al.Addiction-related brain net- works identification via graph diffusion recon- struction network.Brain Informatics11, 1 (2024)

    2024

  39. [48]

    J.et al.Algorithmic fairness in arti- ficial intelligence for medicine and healthcare

    Chen, R. J.et al.Algorithmic fairness in arti- ficial intelligence for medicine and healthcare. Nat. Biomed. Eng.7, 719–742 (2023)

    2023

  40. [49]

    US Patent 12,254,684

    Wang, S., Yu, W., Chen, Z.et al.Smart diagnosis assistance method to solve results of inaccurate classification of image, and ter- minal based on medical images (2025). US Patent 12,254,684

    2025

  41. [50]

    Advocating for neurodata privacy and neurotechnology regulation.Nat

    Yuste, R. Advocating for neurodata privacy and neurotechnology regulation.Nat. Protoc. 18, 2869–2875 (2023)

    2023

  42. [51]

    Goldstein, A.et al.A unified acoustic-to- speech-to-language embedding space captures the neural basis of natural language processing in everyday conversations.Nat. Hum. Behav. 1–15 (2025)

    2025

  43. [52]

    Preprint atbioRxivhttps: //doi.org/10.1101/2021.07.09.451638 (2021)

    Feczko, E.et al.Adolescent Brain Cognitive Development (ABCD) community MRI col- lection and utilities. Preprint atbioRxivhttps: //doi.org/10.1101/2021.07.09.451638 (2021)

    2021 doi

  44. [53]

    F.et al.The minimal prepro- cessing pipelines for the Human Connectome Project.Neuroimage80, 105–124 (2013)

    Glasser, M. F.et al.The minimal prepro- cessing pipelines for the Human Connectome Project.Neuroimage80, 105–124 (2013)

    2013

  45. [54]

    & Halgren, E

    Destrieux, C., Fischl, B., Dale, A. & Halgren, E. Automatic parcellation of human corti- cal gyri and sulci using standard anatomical nomenclature.Neuroimage53, 1–15 (2010)

    2010

  46. [55]

    G., De Heer, W

    Huth, A. G., De Heer, W. A., Griffiths, T. L., Theunissen, F. E. & Gallant, J. L. Natu- ral speech reveals the semantic maps that tile human cerebral cortex.Nature532, 453–458 (2016)

    2016

  47. [56]

    InProceed- ings of the 40th International Conference on Machine Learning, PMLR 202, 28492–28518 (2023)

    Radford, A.et al.Robust speech recognition via large-scale weak supervision. InProceed- ings of the 40th International Conference on Machine Learning, PMLR 202, 28492–28518 (2023)

    2023

  48. [57]

    L.et al.Bloom: a 176b- parameter open-access multilingual language model

    Scao, T. L.et al.Bloom: a 176b- parameter open-access multilingual language model. Preprint at https://arxiv.org/abs/ 2211.05100 (2022)

    2022 arXiv

  49. [58]

    & Luo, Z

    Zheng, Y., Zhang, R., Zhang, J., Ye, Y. & Luo, Z. Llamafactory: Unified efficient fine-tuning of 100+ language models. InProceedings of the Annual Meeting of the Association for Computational Linguistics, 400–410 (2024)

    2024

  50. [59]

    Sudlow, C.et al.UK Biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age.PLoS Med.12, e1001779 (2015)

    2015

  51. [60]

    In37th Conference on Neural Information Processing Systems (NeurIPS 2024)

    Wang, G.et al.EEGPT: Pretrained trans- former for universal and reliable representa- tion of eeg signals (2024). In37th Conference on Neural Information Processing Systems (NeurIPS 2024)

    2024

  52. [61]

    InInternational Conference on Machine Learning8748–8763 (PMLR 2021)

    Radford, A.et al.Learning transferable visual models from natural language supervi- sion. InInternational Conference on Machine Learning8748–8763 (PMLR 2021)

    2021

  53. [62]

    & Ney, H

    Tillmann, C. & Ney, H. Word reordering and a dynamic programming beam search algorithm for statistical machine translation.Comput. Linguist.29, 97–133 (2003). Competing interests The authors declare no competing interests. 16 Extended Data T able 1Datasets Dataset Language Mod...

    2003

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.