Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 40 of the 40 references it cites were found in a registry or an index; 31 of the 48 sources its claims cite were checked for support.

Incremental accumulation of linguistic context in artificial and biological neural networks

Open access

Published version of a preprint

Abstract

Accumulated evidence suggests that Large Language Models (LLMs) are beneficial in predicting neural signals related to narrative processing. The way LLMs integrate context over large timescales, however, is fundamentally different from the way the brain does it. In this study, we show that unlike LLMs that apply parallel processing of large contextual windows, the incoming context to the brain is limited to short windows of a few tens of words. We hypothesize that whereas lower - level brain areas process short contextual windows, higher-order areas in the default-mode network (DMN) engage in an online incremental mechanism where the incoming short context is summarized and integrated with information accumulated across long timescales. Consequently, we introduce a novel LLM that instead of processing the entire context at once, it incrementally generates a concise summary of previous information. As predicted, we found that neural activities at the DMN were better predicted by the incremental model, and conversely, lower-level areas were better predicted with short-context-window LLM.

Overview

CoChat Summary

TL;DR

The paper argues that the human brain integrates linguistic context incrementally, whereas standard LLMs process large text windows in parallel. Using fMRI from 219 participants listening to narratives, it finds that short LLM context windows (~32 tokens) best predict brain activity, and an incremental-context LLM that combines a short window with a self-generated running summary improves prediction in higher-order default mode network areas.

Key findings

  • Neural encoding performance improved as context window size increased only up to 32 tokens, then decreased or plateaued for larger windows; direct comparison of 32 tokens vs. up to 2048 tokens yielded 2594 significant voxels, all favoring the short window (Max = 0.028, Mean = 0.01, SD = 0.004).
  • The incremental-context model significantly outperformed the MAX-tokens model in many parietal, temporal, and frontal areas (5023 significant voxels; Max = 0.02, Mean = 0.008, SD = 0.001), with no voxels significantly favoring the MAX-tokens model.
  • Compared with the 32-token model, the incremental model was better in DMN areas including precuneus, TPJ, and medial prefrontal cortex (679 significant voxels; Max = 0.02, Mean = 0.006, SD = 0.003), while the 32-token model was better in superior temporal gyrus, primary auditory cortex, and inferior frontal gyrus (435 significant voxels; Min = −0.01, Mean = −0.003, SD = 0.0007).
  • Spectral analysis showed high-frequency power negatively correlated with the incremental-vs-short model difference (r = −0.63, p < 0.0001), and low-frequency power positively correlated with it (r = 0.66, p < 0.0001).

Methods

  • fMRI data came from the Narratives dataset: 297 scans from 219 individuals (ages 18–53, mean 22.1 ± 3.1 years, 171 reported female) listening to 8 long spoken stories totaling 15,978 tokens; analyses used 9258 stimulus-locked voxels after inter-subject correlation filtering.
  • Word embeddings were extracted from GPT-NeoXT (20B parameters) with context windows of 8, 16, 32, 64, 128, 256, 512, and up to 2048 tokens; neural encoding used voxel-wise linear regression with PCA to 32 dimensions and 5-fold cross-validation.
  • The incremental-context model concatenated the last 32 tokens with a self-generated summary updated every 50 tokens (summary max 50 tokens; total input ≤90 tokens); spectral analysis used Welch PSD, with high-frequency power integrated from 0.08–0.33 Hz and low-frequency power from 0–0.02 Hz.

Limitations

  • The paper states it does not guarantee the incremental model is completely cognitively plausible: it is unclear whether the brain aggregates long-term context as discrete words or continuously, and whether summarization itself is cognitively plausible; alternative aggregation methods (key sentences, leading keywords) were substantially inferior in their tests, but future studies may find better alternatives.
  • The paper notes that large-context LLMs tend to lose information in the middle of the window, so the incremental model might perform better because it overcomes this issue rather than because it is more cognitively plausible.
  • Evident from the text: the study uses only 8 narratives and passive listening, and it does not assess the incremental model’s performance on NLP tasks related to long texts.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41467-025-56162-9

Cites

40 of 40 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Contextual processing in the brain vs. large language models. a The neuroanatomical hierarchical organization according to multiple timescales of processing. Partially adapted from ref. 18 with the authors’ permission. b Our proposed neural mechanism of integrating long-term contextual information at the top level of the timescale hierarchy. c The baseline implementation of contextual integration via Large Language Models (LLMs). The model is exposed to the entire incoming context window and processes it in a parallel manner. d Our proposed alternative model of contextual integration via LLM. Instead of processing the entire context window at once, the incremental-context LLM is applied sequentially along the story. The LLM accumulates long-term contextual information by generating a concise summary of the past, and, at each step, integrating this summary with the incoming context window and updating the summary to be used in the next step (see more details in Fig. 3 and S3).
Fig. 1 Contextual processing in the brain vs. large language models. a The neuroanatomical hierarchical organization according to multiple timescales of processing. Partially adapted from ref. 18 with the authors’ permission. b Our proposed neural mechanism of integrating long-term contextual information at the top level of the timescale hierarchy. c The baseline implementation of contextual integration via Large Language Models (LLMs). The model is exposed to the entire incoming context window and processes it in a parallel manner. d Our proposed alternative model of contextual integration via LLM. Instead of processing the entire context window at once, the incremental-context LLM is applied sequentially along the story. The LLM accumulates long-term contextual information by generating a concise summary of the past, and, at each step, integrating this summary with the incoming context window and updating the summary to be used in the next step (see more details in Fig. 3 and S3).CC BY-NC-ND · PubMed Central
Fig. 2 The effect of the size of the context window (in # of tokens) fed into the LLM on its ability to predict the neural signals. a Cortical maps for different window sizes showing voxels where the neural encoder score (Pearson’s r calculated between the predicted and the original signals, averaged across 297 fMRI scans) was statistically significant using a non-parametric Wilcoxon signed-rank test and FDR correction. b Averaged r-scores by window size for five different voxels, each located in a different language-related region of interest (ROI) in the left hemisphere. Error bars represent the 95% confidence interval of the mean as calculated via 10,000-iterations bootstrap analysis. Anatomical locations are provided as MNI coordinates. c A cortical map showing ROIs that were predicted significantly better using a window size of 32 tokens than with a window size of up to 2048 tokens (red areas), or vice versa (green areas). A1 primary auditory cortex, STG superior temporal gyrus, IFG inferior frontal gyrus, TPJ temporoparietal junction. Source data are provided as a Source Data file.
Fig. 2 The effect of the size of the context window (in # of tokens) fed into the LLM on its ability to predict the neural signals. a Cortical maps for different window sizes showing voxels where the neural encoder score (Pearson’s r calculated between the predicted and the original signals, averaged across 297 fMRI scans) was statistically significant using a non-parametric Wilcoxon signed-rank test and FDR correction. b Averaged r-scores by window size for five different voxels, each located in a different language-related region of interest (ROI) in the left hemisphere. Error bars represent the 95% confidence interval of the mean as calculated via 10,000-iterations bootstrap analysis. Anatomical locations are provided as MNI coordinates. c A cortical map showing ROIs that were predicted significantly better using a window size of 32 tokens than with a window size of up to 2048 tokens (red areas), or vice versa (green areas). A1 primary auditory cortex, STG superior temporal gyrus, IFG inferior frontal gyrus, TPJ temporoparietal junction. Source data are provided as a Source Data file.CC BY-NC-ND · PubMed Central
Fig. 3 Schematic illustrations depicting the process of extracting contextual embedding representations for a single token using both the baseline short-term context model and the incremental context model. a The short-term, N-tokens (N = 32) model. The word embedding representation of the token i is extracted by providing the LLM with that token, as well as the preceding N tokens. b The long-term, incremental context model. To extract the word embedding representation for the same token i, the input to the model included both the short N-token window (as in the N-tokens model) and a concise natural language summary generated by the model, which was based on information from the long-term context. This long-term contextual summary is updated as the model progresses through the story. It is generated based on the text that appeared before the short N-token window, as well as on the summary generated at the previous update step by the model. See Fig. S3 for more details.
Fig. 3 Schematic illustrations depicting the process of extracting contextual embedding representations for a single token using both the baseline short-term context model and the incremental context model. a The short-term, N-tokens (N = 32) model. The word embedding representation of the token i is extracted by providing the LLM with that token, as well as the preceding N tokens. b The long-term, incremental context model. To extract the word embedding representation for the same token i, the input to the model included both the short N-token window (as in the N-tokens model) and a concise natural language summary generated by the model, which was based on information from the long-term context. This long-term contextual summary is updated as the model progresses through the story. It is generated based on the text that appeared before the short N-token window, as well as on the summary generated at the previous update step by the model. See Fig. S3 for more details.CC BY-NC-ND · PubMed Central
Fig. 4 Comparisons between our incremental context model, the baseline long-term full-transformer (up to 2048 tokens) model, and the baseline short-term context model (32 tokens). a, b Cortical maps show the differences between the models in the neural encoder scores. The following comparisons were made: Incremental Context vs. MAX tokens (a), and Incremental Context vs. 32 tokens (b). The maps display only the voxels that demonstrate a significant difference using a non-parametric Wilcoxon signed-rank test and FDR correction. c Bar plots display the averaged neural encoder results (y axis) for selected voxels located at different language-related ROIs, depending on the model used for embedding representations (32 tokens/MAX-tokens/Incremental Context). Error bars represent the 95% confidence interval of the mean as calculated via 10,000-iterations bootstrap analysis. A1 primary auditory cortex, STG superior temporal gyrus, IFG inferior frontal gyrus, TPJ temporoparietal junction. Source data are provided as a Source Data file.
Fig. 4 Comparisons between our incremental context model, the baseline long-term full-transformer (up to 2048 tokens) model, and the baseline short-term context model (32 tokens). a, b Cortical maps show the differences between the models in the neural encoder scores. The following comparisons were made: Incremental Context vs. MAX tokens (a), and Incremental Context vs. 32 tokens (b). The maps display only the voxels that demonstrate a significant difference using a non-parametric Wilcoxon signed-rank test and FDR correction. c Bar plots display the averaged neural encoder results (y axis) for selected voxels located at different language-related ROIs, depending on the model used for embedding representations (32 tokens/MAX-tokens/Incremental Context). Error bars represent the 95% confidence interval of the mean as calculated via 10,000-iterations bootstrap analysis. A1 primary auditory cortex, STG superior temporal gyrus, IFG inferior frontal gyrus, TPJ temporoparietal junction. Source data are provided as a Source Data file.CC BY-NC-ND · PubMed Central
Fig. 5 The complementary spectral analysis. a The cortical map of the ΔrIncrementalContext−32tokens (abbreviated as ΔrIncCon−32T) scores replicated from Fig. 3b but with no statistical thresholding. This map shows the overall pattern of the cortical hierarchy between voxels that are better predicted by the long-term Incremental context model (more blue areas) and voxels that are better predicted by the short-term (32 tokens) model (more red areas). b The power spectral density (PSD) calculated for the BOLD signal within four important ROIs: inferior parietal lobule (IPL), dorsolateral prefrontal cortex (DLPFC), the Precuneus, and the primary auditory area. Source data are provided as a Source Data file. c, d Cortical maps showing the power of high (c) and low (d) frequencies across the cortex. These measurements were extracted from the PSD curve as described in the Method. e, f Scatter plots representing the correlation between the ΔrIncrementalContext−32tokens scores (Fig. 4a) and the high (e) or low (f) frequency power of the signal. The statistical significance of the correlations was tested using Fisher’s Z test for correlation coefficients.
Fig. 5 The complementary spectral analysis. a The cortical map of the ΔrIncrementalContext−32tokens (abbreviated as ΔrIncCon−32T) scores replicated from Fig. 3b but with no statistical thresholding. This map shows the overall pattern of the cortical hierarchy between voxels that are better predicted by the long-term Incremental context model (more blue areas) and voxels that are better predicted by the short-term (32 tokens) model (more red areas). b The power spectral density (PSD) calculated for the BOLD signal within four important ROIs: inferior parietal lobule (IPL), dorsolateral prefrontal cortex (DLPFC), the Precuneus, and the primary auditory area. Source data are provided as a Source Data file. c, d Cortical maps showing the power of high (c) and low (d) frequencies across the cortex. These measurements were extracted from the PSD curve as described in the Method. e, f Scatter plots representing the correlation between the ΔrIncrementalContext−32tokens scores (Fig. 4a) and the high (e) or low (f) frequency power of the signal. The statistical significance of the correlations was tested using Fisher’s Z test for correlation coefficients.CC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

7 Valid · 4 Not valid · 8 Not checked · 53 Own results and interpretations

45 more claims · 3 with their cited sources checked

Create a free account to see all 72 claims and to verify the 17 references still open.

See all claims — free account
  • It may state two separate claims as one. 12
  • Probably not a substantial claim. 13
  • The quote only partly states this claim. 1

Create a free account to see these claims.

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?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 35 references

Research works printed but not linked: 3

  1. Schrimpf, M. et al. Artificial neural networks accurately predict language processing in the brain. BioRxiv https://www.biorxiv.org/content/10.1101/2020.06.26.174482v1 (2020). · doi:10.1101/2020.06.26.174482v1
  2. Radford, A. et al. Language models are unsupervised multitask learners. OpenAI Blog 1, 9 (2019).
  3. Jain, S. & Huth, A. Incorporating context into language encoding models for fMRI. In: Advances in Neural Information Processing Systems (eds. Bengio, S. et al.) vol. 31 (Curran Associates, Inc., 2018). · DOI · Paper page

Cited by

See all 26

Metadata

Authors
Tikochinski, R., Goldstein, A., Meiri, Y., Hasson, U., & Reichart, R.
Affiliations
  • Technion – Israel Institute of Technology (IL)
  • Hebrew University of Jerusalem (IL)
  • Princeton University (US)
Corresponding author
Tikochinski, Refael
Venue
Nature Communications
Published
2025-01-18
Volume
16
Issue
1
Pages
803-803
ISSN
2041-1723 (ISSN-L)
OpenAlex
W4406513948
Cited by
24
References
35
Citations by year
2026: 10 · 2025: 14
FWCI
13.68
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by-nc-nd)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Council for Higher Education
  • National Institutes of Health
  • National Institute on Deafness and Other Communication Disorders (R01DC022534)
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Neurobiology of Language and Bilingualism
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Neural dynamics and brain function
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Neuroscience and Music Perception
Keywords
linguistic context, incremental processing, narrative processing, oral narrative, large language models, neural activity prediction, functional MRI, artificial neural networks
MeSH terms
Adult, Brain, Brain Mapping, Female, Humans, Language, Linguistics, Magnetic Resonance Imaging, Male, Models, Neurological, Speech Perception, Neural Networks, Computer, Young Adult

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.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-05 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

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

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Incremental accumulation of linguistic context in artificial and biological neural networks. https://app.cochat.ai/papers/10.1038/s41467-025-56162-9
The paper
Tikochinski, R., Goldstein, A., Meiri, Y., Hasson, U., & Reichart, R. (2025). Incremental accumulation of linguistic context in artificial and biological neural networks. Nature Communications. https://doi.org/10.1038/s41467-025-56162-9

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.