Pith. sign in

REVIEW 3 major objections 4 minor 6 references

Brains and language models converge on a shared conceptual space across different languages

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Encoding models for one language predict brain responses to the same story in two other languages, implying a shared conceptual space across English, Chinese, and French.

desk verdict The cross-language uBERT encoding transfer as written is missing the embedding-space alignment it logically requires, which is a bigger problem than the translation-equivalence concern. read the letter →

arxiv 2506.20489 v1 pith:C5K3S4JQ submitted 2025-06-25 q-bio.NC

Zaid Zada , Samuel A Nastase , Jixing Li , Uri Hasson This is my paper · ORCID
classification q-bio.NC
keywords cross-languageneuralrepresentationvoxelwiseencodingmodelsmultilinguallanguagenaturalisticfMRIsharedconceptualspacenarrativecomprehensionintersubjectcorrelationdefault-modenetwork
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 asks whether the neural representation of story meaning is shared across speakers of different languages, rather than merely occupying overlapping brain regions. Using fMRI recordings of native listeners who heard the same audiobook in English, Chinese, or French, the authors trained voxelwise encoding models that map language-model word embeddings onto each listener's brain activity. The central result is that a model trained on one language generalizes to listeners hearing the same story in another language, with significant predictive performance in high-level language and default-mode regions. The paper also shows that separately trained unilingual language models converge on a common embedding geometry, strongest in middle layers, and that languages closer to a listener's native language predict that listener's brain activity better. If correct, these findings indicate that speakers of different languages share a partly language-agnostic conceptual space, and that language models trained on different languages discover the same structure.

What carries the argument

The central machinery is the voxelwise encoding model, a ridge-regression model that predicts each brain voxel's BOLD time series from contextual word embeddings extracted from a language model. To compare across languages, the authors downsample the actual and predicted BOLD responses to the sentence level and regress out low-level confounds including word rate, syllable rate, acoustic RMS energy, onset strength, framewise displacement, and TR count. For unilingual models, embedding spaces are aligned with a Procrustes rotation learned on half the sentences and evaluated on the other half; multilingual models are compared directly because they share a single embedding space. Whisper supplies a second stream of speech embeddings from its encoder and word embeddings from its decoder, allowing the paper to separate shared speech features from shared word-level meaning.

What would settle it

A concrete test would use the same three languages with the story's sentences translated to carry different meanings while matching low-level acoustics; if encoding models trained on one language still predict the other-language brains as well as they do for the original story, cross-language generalization would not be evidence for a shared conceptual space.

Watch

Extended reading notes

Core claim

The paper reports that language models trained on different languages converge onto a similar embedding space, especially in the middle layers, and that this shared geometry can be used to predict neural activity across language groups. In the brain analysis, a voxelwise encoding model trained on English embeddings and English listeners' BOLD responses predicts the BOLD responses of Chinese and French listeners hearing the same story, and the same holds for every language pair. The generalization is strongest in high-level language areas and default-mode regions, and is weaker in early auditory cortex and superior temporal gyrus, where language-specific speech processing likely dominates. The paper concludes that the neural representation of meaning is at least partly shared across speakers of different languages, and that language models trained on separate corpora converge on this shared meaning.

Load-bearing premise

The three professionally translated audiobook versions are meaning-equivalent at the sentence level, and regressing out word rate, syllable rate, acoustic RMS, onset strength, framewise displacement, and TR count removes all language-specific low-level cues that could otherwise explain cross-language prediction.

Editorial extensions

If this is right

  • An encoding model trained on one language can be applied, without retraining, to predict where and when another language's listeners engage with the same narrative.
  • Unilingual language models trained on separate corpora contain a common geometric core, strongest in middle layers, and the degree of this geometric convergence predicts how well embeddings from one language transfer to another language's brain activity.
  • Brain responses to a story index linguistic relatedness: embeddings from languages closer to a listener's native language predict that listener's brain activity better, so neural data can serve as a behavioral measure of perceived language similarity.
  • Shared speech features across languages are present in Whisper's late encoder layers, and phoneme classifiers trained on one language's speech embeddings identify shared phonemes in other languages' speech.
  • The findings imply that high-level language and default-mode regions encode narrative meaning in a form that is partly independent of the particular sounds, scripts, and syntax of a given language.

Reading between the lines

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

  • If the shared space is driven by plot and event structure rather than word-level semantics, cross-language prediction should survive within-sentence word-order scrambling but degrade when sentence order is shuffled; this is testable with the same corpus.
  • The same encoding framework could be turned into a translation-evaluation metric: candidate translations that better predict a target-language listener's brain activity would score higher on neural naturalness.
  • The 58-language analysis is biased toward Indo-European languages, so the family-tree gradient is most reliable for Germanic and Romance languages; extending to Sinitic, Dravidian, and Turkic families with native corpora is the direct next test.
  • Untrained or randomly initialized language models should not reproduce the cross-language generalization, making the trained-versus-untrained contrast a formal benchmark for any future claim of shared conceptual structure.
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

3 major / 4 minor

Summary. The paper uses the Le Petit Prince multilingual fMRI corpus (English, Chinese, and French listeners) together with unilingual BERT, multilingual BERT, and Whisper embeddings to ask whether a conceptual space is shared across languages, both in language models and in the brain. The authors report three main results: (1) sentence embeddings from separately trained unilingual BERT models can be aligned by a learned rotation and show above-baseline cross-language similarity, peaking in middle layers; (2) voxelwise encoding models trained on one language's LM embeddings and brain responses generalize to brain responses of listeners of the other two languages, especially in high-level language and default-mode regions; and (3) in a multilingual model, embeddings from languages more similar to a listener's native language better predict that listener's brain activity, and Whisper speech embeddings contain cross-language phoneme information. The discussion interprets these findings as evidence for a partially shared conceptual representation across languages and across human brains and language models.

Significance. If the central claims hold, the paper would be a valuable contribution to the cross-language neuroscience of language: it uses a naturalistic, three-language fMRI dataset, proposes a concrete encoding-model framework for testing cross-language generalization, and includes a genuine held-out Procrustes alignment for embedding comparison. The phoneme-probing analysis is a true cross-language test, and the mBERT zero-shot generalization across 58 languages is a novel and falsifiable prediction. The paper also ships its code and uses an open dataset, which aids reproducibility. However, the reported unilingual embedding similarity is quite small (average r = 0.115), and the exact procedure for the central uBERT cross-language encoding transfer is ambiguous, which bears directly on the strength of the headline claim. These issues are addressable but currently leave the main conclusion less crisply supported than the abstract suggests.

major comments (3)
  1. [Methods, 'Evaluating encoding models across languages'] The central uBERT cross-language result (Fig. 3C) is ambiguous about what is actually fed to the trained encoding weights at test time. The text says 'we evaluated predictions from the English model' against French and Chinese BOLD, which suggests the predictions are generated from English uBERT embeddings of the held-out English sentences, not from French or Chinese uBERT embeddings. If that is the case, no Procrustes alignment of separately trained uBERT spaces is needed, but then the claim should be stated as 'English LM features predict French/Chinese BOLD,' not as 'encoding models trained on one language generalize to another language.' If instead the weights were applied to French/Chinese uBERT embeddings, then the Methods omit the required rotation alignment for the encoding pipeline, and the train/test separation of that alignment would need to be verified. Because the abstract and Fig. 3C hinge on this point, please state explicitly which embeddings were used to generate the cross-language predictions and, if alignment was used, describe it.
  2. [Results, 'Unilingual models and brains converge'] The evidence for the claim that unilingual LMs 'converge on a similar embedding space' rests on a reported average test correlation of r = 0.115 across layers and language pairs after subtracting an untrained baseline. This is a small effect, and the paper does not report confidence intervals, permutation tests, or the distribution across the 825 held-out sentences and 768 dimensions. Since this is one of the two pillars of the title claim, please provide inferential statistics and effect sizes that establish that the held-out similarity is reliably above baseline, rather than only reporting the mean.
  3. [Results, Fig. 3C and Fig. S2] The difference map between within- and across-language encoding performance (Fig. S2) is used to argue that the across-language effect is only slightly weaker than the within-language effect. However, the whole-brain correlation between within- and across-language unthresholded maps (r = 0.974) is reported without any uncertainty or test of whether the small differences are reliable. More importantly, the evaluation against group-averaged BOLD from a different scanner (French participants) may introduce systematic noise; the authors do not discuss how scanner differences could affect the cross-language generalization magnitudes. Please address this explicitly, for example by reporting the correlation separately for each language pair and scanner.
minor comments (4)
  1. [Fig. 3 caption] The caption states 'pFDR > .05' for thresholded encoding maps; this should presumably be 'pFDR < .05.'
  2. [Discussion, limitations] The paper acknowledges the use of a single children's book and professionally translated audiobooks as limitations, which is appropriate; please also mention that the sentence-level averaging to 1,649 points limits temporal resolution and that the cross-language comparison is therefore at a coarse timescale.
  3. [Methods, 'Translating the story'] The GPT-4o translation into 55 languages is described, but no check of translation quality beyond manual inspection of samples is reported; a sentence-level back-translation consistency measure would strengthen the 58-language analysis in Fig. 4.
  4. [Results, Fig. 4C] The correlations between language-family closeness and encoding performance (r = 0.786, r = 0.869) are reported with p-values but without correction for the multiple language-family comparisons or for the fact that the same subjects and voxel mask are reused across the 58 languages.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central cross-language encoding transfer is a held-out predictive test, and the cited self-citations are not load-bearing.

full rationale

The central claim—that encoding models trained on one language's LM embeddings generalize to held-out brain responses of listeners of another language—is a genuine out-of-sample prediction. Weights are fit on eight runs of one language and evaluated on a held-out ninth run against the other language group's average BOLD; predictions are correlated with actual BOLD after regressing out low-level confounds. The Procrustes rotation for uBERT sentence-embedding similarity is learned on half the sentences and evaluated on the remaining half, so the reported similarity is a held-out quantity. No load-bearing step reduces to its own inputs by construction. The paper's self-citations (e.g., Zada et al. 2024, 2025; Honey et al. 2012) are precedents or methodological background, not the justification for the cross-language result. Two concerns are noted but are not circularity. First, the uBERT cross-language encoding transfer (Fig. 3C) requires a coordinate alignment between separately trained BERT embedding spaces, yet the Methods describe Procrustes alignment only for the similarity analysis; if the alignment was applied for encoding, it is an omitted essential step, and if not, the transfer is unexplained—this is a reproducibility and correctness flaw, not a circular reduction. Second, in the 58-language analysis, both the embedding-similarity axis and the zero-shot transfer performance derive from the same mBERT embedding space, so the correlation between them is partly shaped by the model's geometry; however, the transfer performance still correlates with held-out brain data, and the authors include an untrained-embedding control, so the claim is not equivalent to its input. Overall, the derivation chain is self-contained against external benchmarks and the central prediction is independent of the authors' prior results.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or theoretical entities. It relies on standard domain assumptions about translation equivalence, intersubject averaging, confound regression, and the semantic validity of LM embeddings. The only fitted quantities are the ridge penalty and the Procrustes rotation, both handled with held-out evaluation.

free parameters (2)
  • Ridge regularization strength (per voxel) = Selected by nested cross-validation over 10^1 to 10^19
    Standard regularization for high-dimensional encoding models; tuned on training data and not conceptually load-bearing.
  • Procrustes rotation matrix (per language pair) = Learned on 825 training sentences per pair
    Used only in the embedding-similarity analysis; evaluated on held-out sentences, so it does not inflate the reported similarity.
assumptions (5)
  • domain assumption The English, Chinese, and French audiobook translations preserve the conceptual meaning of the original story.
    Manual sentence alignment explicitly assumes 'each sentence has the same meaning across each language' (Methods, Design and stimuli). This is load-bearing for cross-language generalization.
  • domain assumption Averaging BOLD responses across subjects isolates shared, stimulus-driven neural responses.
    Used to produce group-average targets for encoding evaluation, following Hasson et al. (2004) and Nastase et al. (2019).
  • domain assumption Sentence-level averaging and regressing word rate, syllable rate, acoustic RMS, onset strength, framewise displacement, and TR count removes language-specific low-level confounds.
    This is the main control for non-conceptual cross-language similarity; if insufficient, cross-language generalization could be inflated (Methods, Evaluating encoding models across languages).
  • domain assumption LM contextual embeddings capture conceptual and semantic features relevant to human neural representations.
    Borrowed from prior work (Schrimpf et al., 2021; Goldstein et al., 2022); the paper does not independently derive this.
  • standard math Procrustes alignment on a training half gives an unbiased estimate of embedding similarity on the test half.
    Used to compute unilingual embedding similarities (Methods, Computing similarity between word embeddings).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Brains and language models converge on a shared conceptual space across different languages." pith.science (2026). https://pith.science/paper/C5K3S4JQ

@misc{pith2026250620489,
  author       = {Pith},
  title        = {Pith review of: Brains and language models converge on a shared conceptual space across different languages},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C5K3S4JQ}},
  note         = {Machine review of arXiv:2506.20489}
}
read the original abstract

Human languages differ widely in their forms, each having distinct sounds, scripts, and syntax. Yet, they can all convey similar meaning. Do different languages converge on a shared neural substrate for conceptual meaning? We used language models (LMs) and naturalistic fMRI to identify neural representations of the shared conceptual meaning of the same story as heard by native speakers of three languages: English, Chinese, and French. We found that LMs trained on entirely different languages converge onto a similar embedding space, especially in the middle layers. We then aimed to find if a similar shared space exists in the brains of different native speakers of the three languages. We trained voxelwise encoding models that align the LM embeddings with neural responses from one group of subjects speaking a single language. We then used the encoding models trained on one language to predict the neural activity in listeners of other languages. We found that models trained to predict neural activity for one language generalize to different subjects listening to the same content in a different language, across high-level language and default-mode regions. Our results suggest that the neural representations of meaning underlying different languages are shared across speakers of various languages, and that LMs trained on different languages converge on this shared meaning. These findings suggest that, despite the diversity of languages, shared meaning emerges from our interactions with one another and our shared world.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

6 extracted references · 3 canonical work pages

  1. [13]

    Correia, J., Formisano, E., Valente, G., Hausfeld, L., Jansma, B., & Bonte, M. (2014). Brain-Based Translation: fMRI Decoding of Spoken Words in Bilinguals Reveals Language-Independent Semantic Representations in Anterior Temporal Lobe. The Journal of Neuroscience , 34 (1), 332–338. https://doi.org/10.1523/JNEUROSCI.1302-13.2014 14. De Deyne, S., Navarro,...

    work page arXiv 2014

  2. [24]

    Evans, N., & Levinson, S. C. (2009). The myth of language universals: Language diversity and its importance for cognitive science. Behavioral and Brain Sciences , 32 (5), 429–448. https://doi.org/10.1017/S0140525X0999094X 25. Fedorenko, E., Piantadosi, S. T., & Gibson, E. A. F. (2024). Language is primarily a tool for communication rather than thought. Na...

    arXiv doi:10.48550/arxiv.2405.07987 2009

  3. [36]

    Joshi, P., Santy, S., Budhiraja, A., Bali, K., & Choudhury, M. (2020). The State and Fate of Linguistic Diversity and Inclusion in the NLP World. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , 6282–6293. https://doi.org/10.18653/v1/2020.acl-main.560 37. Kundu, P., Inati, S. J., Evans, J. W., Luh, W.-M., & Bandett...

    arXiv doi:10.48550/arxiv.1309.4168 2020

  4. [55]

    T., Muller, D

    Piantadosi, S. T., Muller, D. C. Y., Rule, J. S., Kaushik, K., Gorenstein, M., Leib, E. R., & Sanford, E. (2024). Why concepts are (probably) vectors. Trends in Cognitive Sciences , S1364661324001712. https://doi.org/10.1016/j.tics.2024.06.011 56. Popham, S. F., Huth, A. G., Bilenko, N. Y., Deniz, F., Gao, J. S., Nunez-Elizalde, A. O., & Gallant, J. L. (2...

    doi:10.1016/j.tics.2024.06.011 2024

  5. [67]

    Same Neurons, Different Languages: Probing Morphosyntax in Multilingual Pre-trained Models

    Stańczak, K., Ponti, E., Hennigen, L. T., Cotterell, R., & Augenstein, I. (2022). Same Neurons, Different Languages: Probing Morphosyntax in Multilingual Pre-trained Models (arXiv:2205.02023). arXiv. https://doi.org/10.48550/arXiv.2205.02023 68. Stephens, G. J., Silbert, L. J., & Hasson, U. (2010). Speaker–listener neural coupling underlies successful comm...

    work page Pith review arXiv doi:10.48550/arxiv.2205.02023 2022

  6. [76]

    C., Jiang, S., & Srinivasan, M

    Xu, Y., Duong, K., Malt, B. C., Jiang, S., & Srinivasan, M. (2020). Conceptual relations predict colexification across languages. Cognition , 201 , 104280. https://doi.org/10.1016/j.cognition.2020.104280 77. Yamashita, M., Kubo, R., & Nishimoto, S. (2025). Conversational content is organized across multiple timescales in the brain. Nature Human Behaviour ....

    work page arXiv 2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.