Brains and algorithms partially converge in natural language processing
Open accessAbstract
Deep learning algorithms trained to predict masked words from large amount of text have recently been shown to generate activations similar to those of the human brain. However, what drives this similarity remains currently unknown. Here, we systematically compare a variety of deep language models to identify the computational principles that lead them to generate brain-like representations of sentences. Specifically, we analyze the brain responses to 400 isolated sentences in a large cohort of 102 subjects, each recorded for two hours with functional magnetic resonance imaging (fMRI) and magnetoencephalography (MEG). We then test where and when each of these algorithms maps onto the brain responses. Finally, we estimate how the architecture, training, and performance of these models independently account for the generation of brain-like representations. Our analyses reveal two main findings. First, the similarity between the algorithms and the brain primarily depends on their ability to predict words from context. Second, this similarity reveals the rise and maintenance of perceptual, lexical, and compositional representations within each cortical region. Overall, this study shows that modern language algorithms partially converge towards brain-like solutions, and thus delineates a promising path to unravel the foundations of natural language processing.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The finding that Item 1 corresponds approximately to the 2/3 layer is consistent with previous studies suggesting that the 2/3 layer is optimal for predicting human brain activity.
“This finding is consistent with previous studies, which suggest that the 2 3 layer is optimal for predicting human brain activity (Schrimpf et al., 2021; Caucheteux and King, 2022; Caucheteux et al., 2023).”
From Can "consciousness" be observed from large language model (LLM) internal states? Dissecting LLM representations obtained from Theory of Mind test with Integrated Information Theory and Span Representation analysis (2025) · cites this paper as Caucheteux and King, 2022
The Results section, “Compositional embeddings best predict brain responses” (page unavailable), reports that middle transformer layers outperform input and output layers in predicting MEG and fMRI responses. It defines the middle layers as spanning one-half to three-quarters of model depth, which includes two-thirds, but does not establish the two-thirds layer specifically as optimal. (LLM, conf=0.86)
Experimental — accuracy not yet measuredThe 2/3 layer has been reported to exhibit strong alignment with human brain activity.
“as this layer has been reported to exhibit strong alignment with human brain activity (Schrimpf et al., 2021; Caucheteux and King, 2022; Caucheteux et al., 2023).”
From Can "consciousness" be observed from large language model (LLM) internal states? Dissecting LLM representations obtained from Theory of Mind test with Integrated Information Theory and Span Representation analysis (2025) · cites this paper as Caucheteux and King, 2022
The Results section (“Compositional embeddings best predict brain responses”; page unavailable) finds that middle transformer layers—defined as halfway to three-quarters through the model, a range that includes the two-thirds layer—map onto brain activity better than input or output layers. It does not single out the two-thirds layer, and the Discussion (page unavailable) says the model-to-brain correlations are very low, so “strong alignment” overstates the measured relationship. (LLM, conf=0.84)
Experimental — accuracy not yet measuredThe intermediate-to-deep layers (l = 2/3 n_layers) of each LLM merit particular attention due to their demonstrated superior predictive capabilities for human brain activity.
“Additionally, we focused on the intermediate-to-deep layers ( 𝑙 = 2 3 𝑛 𝑙𝑎𝑦𝑒𝑟𝑠 ) of each LLM, as suggested by Caucheteux et al. (2023). These layers merit particular attention due to their demonstrated superior predictive capabilities for human brain activity (Schrimpf et al., 2021; Caucheteux and King, 2022; Caucheteux et al., 2023).”
From Can "consciousness" be observed from large language model (LLM) internal states? Dissecting LLM representations obtained from Theory of Mind test with Integrated Information Theory and Span Representation analysis (2025) · cites this paper as Caucheteux and King, 2022
Recent studies have increasingly explored LLM representations as analogs to neuroimaging data, aligning and comparing the human brain with the LLM brain during natural language processing.
“Recent studies have increasingly explored LLM representations as analogs to neuroimaging data, aligning and comparing the human brain with the LLM”
From Can "consciousness" be observed from large language model (LLM) internal states? Dissecting LLM representations obtained from Theory of Mind test with Integrated Information Theory and Span Representation analysis (2025) · cites this paper as Caucheteux and King, 2022
LLM representations are being increasingly studied as analogs of neuroimaging data when comparing and aligning human brain activity with LLM 'brain' activity in natural language processing.
“While LLM representations are being increasingly studied as analogs of neuroimaging data when comparing and aligning human brain activity with LLM”
From Can "consciousness" be observed from large language model (LLM) internal states? Dissecting LLM representations obtained from Theory of Mind test with Integrated Information Theory and Span Representation analysis (2025) · cites this paper as Caucheteux and King, 2022
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
The citation lists haven’t been fetched yet
References 80 · Cited by 318
Research works printed but not linked: 15
- [1]Turing, A. M. Parsing the Turing Test 23–65 (Springer, 2009).
- [4]Vaswani, A. et al. Attention is all you need. In Proceedings on NIPS (Cornell University, 2017).
- [5]Devlin, J., Chang, M., Lee, K. & Toutanova, K. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) (2019).
- [6]Lample, G. & Conneau, A. Cross-lingual language model pretraining. In Adv. Neural Inf. Process. Syst. (2019).
- [7]Brown, T. B. et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems (2020).
- [12]Marcus, G. Deep learning: a critical appraisal. Preprint at https://arXiv.org/1801.00631 (2018).
- [14]Mikolov, T., Chen, K., Corrado, G. & Dean, J. Efficient estimation of word representations in vector space. Preprint at https://arxiv.org/1301.3781 (2013).
- [28]Caucheteux, C., Gramfort, A. & King, J.-R. Disentangling syntax and semantics in the brain with deep networks. ICML 2021-38th International Conference on Machine Learning (2021).
- [31]Jain, S. & Huth, A. in Advances in Neural Information Processing Systems (eds Bengio, S. et al.) vol. 31, 6628–6637 (Curran Associates, Inc., 2018).
- [32]Athanasiou, N., Iosif, E. & Potamianos, A. Neural activation semantic models: computational lexical semantic models of localized neural activations. In Proceedings of the 27th International Conference on Computational Linguistics 2867–2878 (Association for Computational Linguistics, 2018).
- [33]Toneva, M. & Wehbe, L. Interpreting and improving natural-language processing (in machines) with natural language-processing (in the brain). Advances in Neural Information Processing Systems 32 (2019).
- [69]Reddy, A. J. & Wehbe, L. Syntactic representations in the human brain: beyond effort-based metrics. Preprint at bioRXiv (2021).
- [71]Kell AJE Yamins DLK Shook EN Norman-Haignere SV McDermott JH A task-optimized neural network replicates human auditory behavior, predicts brain responses, and reveals a cortical processing hierarchy Neuron 2018 98 630–644 10.1016/j.neuron.2018.03.044 29681533 · DOI · Paper page
- [81]Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT (2019).
- [93]Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S. & Dean, J. Distributed Representations of Words and Phrases and their Compositionality. Advances in Neural Information Processing Systems 3111–3119 (MIT Press, 2013).
- Authors
- Caucheteux, C., & King, J.-R.
- Affiliations
- Commissariat à l'Énergie Atomique et aux Énergies Alternatives (FR)
- Centre National de la Recherche Scientifique (FR)
- Corresponding author
- Caucheteux, Charlotte
- King, Jean-Rémi
- Venue
- Communications Biology
- Published
- 2022-02-16
- Volume
- 5
- Issue
- 1
- Pages
- 134-134
- ISSN
- 2399-3642 (ISSN-L)
- PMC
- PMC8850612
- OpenAlex
- W4212828284
- Cited by
- 318
- References
- 80
- Citations by year
- 2026: 62 · 2025: 82 · 2024: 84 · 2023: 66 · 2022: 20 · 2021: 2 · 2019: 1 · 2013: 1
- FWCI
- 32.48
- Type
- journal article
- Language
- en
- Update notices
- correction (2023-04-11)
- Open access
- gold
- doi.org
- Locations
- Communications Biology (publishedVersion, cc-by)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- HAL (submittedVersion, other-oa)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- Agence Nationale de la Recherche (ANR-17-EURE-, ANR-17-EURE-0017, ANR-17)
- Fondation Fyssen (17-EURE-0017, ANR-17-EURE-0017)
- Topics
- Life Sciences › Neuroscience › Cognitive Neuroscience › Neurobiology of Language and Bilingualism
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Social Sciences › Psychology › Developmental and Educational Psychology › Reading and Literacy Development
- Keywords
- language models, natural language processing, word prediction, functional MRI, magnetoencephalography, sentence processing, perceptual representation, lexical representation, compositional representations, deep learning
- MeSH terms
- Algorithms, Brain, Humans, Language, Magnetic Resonance Imaging, Natural Language Processing
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
