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 accessPublished 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
Cites
Statistics
Figures





Verification
Claims the paper makes
7 Valid · 4 Not valid · 8 Not checked · 53 Own results and interpretations
Some cited sources support this claim and some don’t — read each source’s verdict below.
In the paper
“Previous studies in the neural encoding literature all used LLMs that were only aware of the short-term context, i.e., where the context window size was as large as several tens of words1–10.”
§ Incremental accumulation of linguistic context in artificial and biological neural networks > Discussion
What the sources say
- Partly supportedcites [1] Pereira, F. et al. Toward a universal decoder of linguistic meaning from brain activation. Nat. Commun. 9, 963 (2018). sourceExperimental — accuracy not yet measured
The Methods section (“Semantic vectors”; page unavailable) says the paper used GloVe vectors based on word co-occurrences within a sentence or a 5–10-word window, supporting the short-context aspect. It also reports testing skip-thought vectors, which represent a sentence using its preceding and succeeding sentences. The paper does not establish that all previous neural-encoding studies used LLMs or that their context windows reached several tens of words. (LLM, conf=0.85)
- Not checked yetcites [2] Schwartz, D., Toneva, M. & Wehbe, L. Inducing brain-relevant bias in natural language processing models. Adv. Neural Inf. Process. Syst. 32 (2019).
- Not checked yetcites [3] 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).
- Disputedcites [4] Caucheteux, C., Gramfort, A. & King, J.-R. Model-based analysis of brain activity reveals the hierarchy of language in 305 subjects. In: EMNLP 2021-Conference on Empirical Methods in Natural Language Processing (2021). sourceExperimental — accuracy not yet measured
The “Extracting deep models’ activations” section (page unavailable) says this neural-encoding study fed GPT-2 sequences of up to 256 BPE tokens in memory. That is contrary to the claim that all previous studies limited their models’ context windows to at most several tens of words, although the paper reports tokens rather than a word count. (LLM, conf=0.78)
Checked by a reasoning model — may be wrong.
- Disputedcites [5] Caucheteux, C., Gramfort, A. & King, J.-R. Evidence of a predictive coding hierarchy in the human brain listening to speech. Nat. Hum. Behav. 7, 430–441 (2023). sourceExperimental — accuracy not yet measured
The Methods section, “Activations of deep language models” (page unavailable), says this study supplied GPT-2 with up to 1,024 previous tokens per word—not a context window limited to several tens of words. The separate seven-word forecast window described in Methods is an addition to the encoding model, not the LLM’s context window. (LLM, conf=0.90)
Checked by a reasoning model — may be wrong.
- Disputedcites [6] Caucheteux, C., Gramfort, A. & King, J.-R. Deep language algorithms predict semantic comprehension from brain activity. Sci. Rep. 12, 16327 (2022). sourceExperimental — accuracy not yet measured
The Materials and methods section (GPT-2 activations; page unavailable) says this study used a sliding context window of 1,024 tokens, not merely several tens of words. The Discussion (page unavailable) also reports effects of attention to context more than 1,000 words away. This cited study therefore contradicts the claim that all previous neural-encoding studies used only short-term context. (LLM, conf=0.90)
Checked by a reasoning model — may be wrong.
- Disputedcites [7] Huth, A. G., de Heer, W. A., Griffiths, T. L., Theunissen, F. E. & Gallant, J. L. Natural speech reveals the semantic maps that tile human cerebral cortex. Nature 532, 453–458 (2016). sourceExperimental — accuracy not yet measured
The Methods (‘Semantic model construction’; page not supplied) describe a 985-dimensional word-co-occurrence feature space, with a 15-word window used to build its statistics. ‘Estimation and validation of semantic voxel-wise models’ says those word features were used in regularized linear regression, not an LLM with a short-term context window. The 15-word co-occurrence window does not establish the claimed LLM context window. (LLM, conf=0.87)
Checked by a reasoning model — may be wrong.
- Disputedcites [8] Jain, S. & Huth, A. Incorporating context into language encoding models for fMRI. Adv. Neural Inf. Process. Syst. 31, (2018). sourceExperimental — accuracy not yet measured
The Introduction (p. 1) says previous language-encoding studies used word-level embedding vectors that treated each word independently of context—not LLMs with short context windows. Section 4.1 (p. 4) describes the cited paper’s own LSTM model, which used up to 20 previous words, but that does not establish the claim about all previous studies. (LLM, conf=0.88)
Checked by a reasoning model — may be wrong.
- Not checked yetcites [9] Tikochinski, R., Goldstein, A., Yeshurun, Y., Hasson, U. & Reichart, R. Perspective changes in human listeners are aligned with the contextual transformation of the word embedding space. Cereb. Cortex 33, 7830–7842 (2023). source
- Disputedcites [10] Goldstein, A. et al. Shared computational principles for language processing in humans and deep language models. Nat. Neurosci. 25, 369–380 (2022). sourceExperimental — accuracy not yet measured
The Results section (“Contextual representation”; page not provided) says this study supplied GPT-2 with all preceding words up to 1,024 tokens. That contradicts the claim that all cited neural-encoding studies limited model context to several tens of words. (LLM, conf=0.90)
Checked by a reasoning model — may be wrong.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Large Language Models (LLMs) have shown success in predicting neural signals associated with narrative processing”
- Also stated in: § Introduction · boosted
“Recent works in cognitive neuroscience have demonstrated the power of computational large language models (LLMs) in predicting language-evoked neural signals among humans1–10.”
§ Incremental accumulation of linguistic context in artificial and biological neural networks > Abstract
What the sources say
- Partly supportedcites [1] Pereira, F. et al. Toward a universal decoder of linguistic meaning from brain activation. Nat. Commun. 9, 963 (2018). sourceExperimental — accuracy not yet measured
The Results section on sentence decoding describes successful decoding of brain activation elicited by written passages, including 24 first- or third-person narratives in experiment 3. But the Discussion says this paper’s model predicts a stimulus’s semantic representation from imaging data, not neural signals from a large language model. The claim therefore overstates both the type of model and the direction of prediction. (LLM, conf=0.87)
- Not checked yetcites [2] Schwartz, D., Toneva, M. & Wehbe, L. Inducing brain-relevant bias in natural language processing models. Adv. Neural Inf. Process. Syst. 32 (2019).
- Not checked yetcites [3] 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).
- Supportedcites [4] Caucheteux, C., Gramfort, A. & King, J.-R. Model-based analysis of brain activity reveals the hierarchy of language in 305 subjects. In: EMNLP 2021-Conference on Empirical Methods in Natural Language Processing (2021). source
The Model-based analysis section says deep language models tend to linearly predict brain responses to language and describes fitting a mapping from model activations to brain recordings. The Additional information section identifies GPT-2 as the model used and says the recordings came from subjects listening to narratives. (LLM, conf=0.91)
- Supportedcites [5] Caucheteux, C., Gramfort, A. & King, J.-R. Evidence of a predictive coding hierarchy in the human brain listening to speech. Nat. Hum. Behav. 7, 430–441 (2023). source
The Results section (“Deep language models map onto brain activity”) reports using GPT-2 activations to predict fMRI signals from people listening to short stories and evaluating those predictions on held-out data. This supports the claim about language models predicting neural signals during narrative processing. (LLM, conf=0.93)
- Supportedcites [6] Caucheteux, C., Gramfort, A. & King, J.-R. Deep language algorithms predict semantic comprehension from brain activity. Sci. Rep. 12, 16327 (2022). source
The Materials and methods section ('Linear mapping between GPT-2 and the brain') says the authors used GPT-2 activations to predict held-out fMRI scans recorded while subjects listened to narratives. The Discussion reports a reliable relationship between the model’s mapping onto brain responses and story comprehension. (LLM, conf=0.93)
- Partly supportedcites [7] Huth, A. G., de Heer, W. A., Griffiths, T. L., Theunissen, F. E. & Gallant, J. L. Natural speech reveals the semantic maps that tile human cerebral cortex. Nature 532, 453–458 (2016). sourceExperimental — accuracy not yet measured
The ‘Estimation and validation of semantic voxel-wise models’ section reports good predictions of BOLD responses to a new narrative story. But it describes word co-occurrence features and linear regression, not a large language model. The prediction claim is supported; attributing it to LLMs is not. (LLM, conf=0.90)
- Partly supportedcites [8] Jain, S. & Huth, A. Incorporating context into language encoding models for fMRI. Adv. Neural Inf. Process. Syst. 31, (2018). sourceExperimental — accuracy not yet measured
The Introduction (p. 2) reports that an LSTM language model predicted fMRI responses to natural narrative speech better than word-embedding models. The paper identifies its model as an LSTM language model, not a Large Language Model; the “LLMs” characterization is therefore not established by this source. (LLM, conf=0.86)
- Not checked yetcites [9] Tikochinski, R., Goldstein, A., Yeshurun, Y., Hasson, U. & Reichart, R. Perspective changes in human listeners are aligned with the contextual transformation of the word embedding space. Cereb. Cortex 33, 7830–7842 (2023). source
- Supportedcites [10] Goldstein, A. et al. Shared computational principles for language processing in humans and deep language models. Nat. Neurosci. 25, 369–380 (2022). source
In Results > Contextual representation (page unavailable), the paper reports that GPT-2 contextual embeddings improved an encoding model’s ability to predict neural responses to words in the story. This supports the claim that a language model successfully predicted neural signals associated with narrative processing. (LLM, conf=0.94)
boostedstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § Introduction · boosted
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Finally, areas along the default mode network (DMN), located at the top of the temporal integration hierarchy, can integrate the paragraphs into a coherent narrative by integrating information accumulated over hundreds of seconds as the story unfolds, with relevant past information stored in long-term memories (Fig. 1a)18–20.”
§ Incremental accumulation of linguistic context in artificial and biological neural networks > Introduction
What the sources say
- Supportedcites [18] Hasson, U., Chen, J. & Honey, C. J. Hierarchical process memory: memory as an integral component of information processing. Trends Cogn. Sci. 19, 304–313 (2015). source
The section 'Process Memory and the Medial Temporal Lobe' states that the areas at the top of the hierarchy, whose responses are influenced by information accumulated over many minutes, 'overlap broadly with the default mode network (DMN)' (angular gyrus, retrosplenial cortex, precuneus, PCC, mPFC), which also covers the claim's 'hundreds of seconds' as the paper's minutes-long timescale. 'A hierarchical topography of process memory' adds that in long-TRW areas (TPJ, AG, mPFC) activity at each moment depended on information arriving over tens of seconds and responses were reliable only at the full-story and paragraph levels of scrambling, i.e. paragraph-into-narrative integration. The section 'Links between process memory and active long-term memory' supports the added clause about integration with distant/long-term knowledge, and the Concluding Remarks explicitly refer to 'the integration of information across multiple paragraphs of text'. (LLM, conf=0.82)
- Partly supportedcites [19] Lerner, Y., Honey, C. J., Silbert, L. J. & Hasson, U. Topographic mapping of a hierarchy of temporal receptive windows using a narrated story. J. Neurosci. 31, 2906–2915 (2011). sourceExperimental — accuracy not yet measured
Our copy supports the core: the apex of the TRW hierarchy consists of parietal and frontal areas (precuneus, mPFC, TPJ, pSTS, inferior frontal gyrus, pericingulate cortices) that 'extract the meaning and context embedded in both the intact paragraphs and the story as a whole' (Time scale of processing and functional specialization), and the Conclusion states high-order areas 'can accumulate information over long time scales (up to many minutes)', consistent with hundreds of seconds. Two attributed elements are not in this paper: it never labels these areas as the default mode network (no DMN passage in our copy), and it does not show integration drawing on relevant past information stored in long-term memory — the closest it comes is a discussion linking TRWs to Ericsson and Kintsch's (1995) 'long-term working memory' and to prior-context accumulation (TRWs in relation to working memory and predictive coding), which is conceptual framing rather than a demonstrated long-term-memory component. (LLM, conf=0.75)
- Partly supportedcites [20] Stephens, G. J., Honey, C. J. & Hasson, U. A place for time: the spatiotemporal structure of neural dynamics during natural audition. J. Neurophysiol. 110, 2019–2026 (2013). sourceExperimental — accuracy not yet measured
Our copy supports the hierarchy and timescale parts: the Abstract states that 'higher order regions integrate information (e.g., the meaning conveyed by a set of sentences or paragraphs) over seconds and minutes,' and the Discussion states that 'areas with the slowest dynamics (e.g., precuneus and medial prefrontal cortex) are coupled to the slowly varying stimulus dynamic (e.g., stimulus narrative),' with the Results/Discussion locating the slowest dynamics in precuneus, medial prefrontal cortex and angular gyrus (the canonical DMN hubs) and the long-TRW regions defined as those responding reliably only to intact paragraphs of the 7-min story. However, no passage in our copy uses the term 'default mode network' or attributes the narrative-integration role to the DMN as such, and the memory clause ('with relevant past information stored in long-term memories') is not stated — the closest text is the Discussion remark that slowest dynamics in precuneus and angular gyrus are 'consistent with their role as cortical network hubs ... and in episodic memory,' which credits Andrews-Hanna et al. 2010 and Ranganath and Ritchey 2012 rather than reporting a long-term-memory integration finding. The paragraph-to-narrative TRW hierarchy itself is also inherited from Lerner et al. 2011, whose data this paper re-analyses. (LLM, conf=0.72)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Rather, humans operate an online mechanism for accumulating information and integrating it into a broader contextual memory, which is changed and updated as the story unfolds17,18.”
§ Incremental accumulation of linguistic context in artificial and biological neural networks > Introduction
What the sources say
- Supportedcites [17] Yeshurun, Y., Nguyen, M. & Hasson, U. The default mode network: where the idiosyncratic self meets the shared social world. Nat. Rev. Neurosci. 22, 181–192 (2021). source
The Abstract and 'An active sense-making network' sections state that the DMN is an active, dynamic network integrating incoming extrinsic information with prior intrinsic information into rich, context-dependent models of situations as they unfold over time, with external information 'accumulated as events unfold'. 'Extrinsic responses of the DMN' describes moment-by-moment integration over seconds to minutes, 'The DMN is sensitive to long-term memories' describes accumulation into short- and long-term context, and 'The DMN represents acquired schemas' describes updating that context when events violate expectations. The only shift is wording: the source attributes the mechanism to the DMN while the citing sentence generalizes to 'humans', which the source's own framing ('making sense of the world is a dynamic, ever-changing process requiring the integration of incoming information with prior knowledge') supports. (LLM, conf=0.85)
- Supportedcites [18] Hasson, U., Chen, J. & Honey, C. J. Hierarchical process memory: memory as an integral component of information processing. Trends Cogn. Sci. 19, 304–313 (2015). source
The Process memory framework section states that 'virtually all cortical circuits have the ability to accumulate information over time,' with long temporal receptive windows (tens to hundreds of seconds) 'needed for the integration of sentences over time while comprehending a narrative,' and 'A method for mapping processing timescales' highlights 'the active and ongoing accumulation and integration of information that occurs during online processing' during minutes-long real-life stimuli such as stories and movies. 'Links between process memory and active long-term memory' adds that online processing integrates recent accumulated information with distant stored knowledge, matching the claim's 'broader contextual memory' that is continuously changed as the story unfolds. Only the citing wording reifies this slightly as a 'contextual memory' store, whereas the paper argues memory is intrinsic to processing circuits rather than a separate store; the attributed content is nonetheless stated. (LLM, conf=0.85)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Moreover, as illustrated in Fig. 2, this limitation was not only observed in temporal areas, which were previously linked to short timescales19,20, but also identified in higher-order areas at the DMN associated with longer timescales.”
§ Incremental accumulation of linguistic context in artificial and biological neural networks > Results > IC is processed in the brain through small context windows
What the sources say
- Partly supportedcites [19] Lerner, Y., Honey, C. J., Silbert, L. J. & Hasson, U. Topographic mapping of a hierarchy of temporal receptive windows using a narrated story. J. Neurosci. 31, 2906–2915 (2011). sourceExperimental — accuracy not yet measured
The Abstract and the section ‘The longest TRWs were observed in the precuneus and frontal areas’ describe a progression from short timescales in early auditory cortex to long timescales in higher-order regions, including the precuneus and mPFC. The paper measured response reliability to scrambled stories, not whether large context windows were limited in predicting brain activity; it also does not identify those regions as the DMN in the cited passages. (LLM, conf=0.83)
- Supportedcites [20] Stephens, G. J., Honey, C. J. & Hasson, U. A place for time: the spatiotemporal structure of neural dynamics during natural audition. J. Neurophysiol. 110, 2019–2026 (2013). source
The citation is attached to the statement that temporal areas were previously linked to short timescales, not to the citing paper’s observation about large context windows. The Discussion contrasts fast dynamics in early auditory areas with slower dynamics in higher-order regions, supporting the cited statement. (LLM, conf=0.89)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
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.
- It may state two separate claims as one. 12
- Probably not a substantial claim. 13
- The quote only partly states this claim. 1
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 cited source says what the claim attributes to it.
In the citing paper
“It does not retain persistent internal memory between separate inputs, meaning that each generation depends only on the tokens within the current context window.”
From Why large language models cannot possess consciousness: an integrated information theory perspective (2025) · cites this paper as [17]
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.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Incremental accumulation of linguistic context in artificial and biological neural networks > Methods > Correlating low-frequency and HFD with
“The range of the integral on LFP starts from Zero (the minimum decomposed frequency) to 0.02 Hz, which is equivalent to a window of 256 tokens.”
§ Incremental accumulation of linguistic context in artificial and biological neural networks > Methods > Correlating low-frequency and HFD with
“32 tokens takes ~12 s, and therefore, 0.08 Hz ≈ 1/12”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Incremental accumulation of linguistic context in artificial and biological neural networks > Results > A model-free spectral analysis of the frequency domain supports the model-based analyses
“a The cortical map of the (abbreviated as ) scores replicated from Fig. 3b but with no statistical thresholding.”
§ Incremental accumulation of linguistic context in artificial and biological neural networks > Results > A model-free spectral analysis of the frequency domain supports the model-based analyses
“In Fig. 5a we re-visualize this map, this time over the entire set of 9258 stimulus-locked voxels (i.e., voxels that yielded significant inter-subjects correlation scores, see Method), with no statistical threshold (as in Fig. 4b).”
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404
Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing
Benjamini, Y., & Hochberg, Y. (1995) · Journal of the Royal Statistical Society Series B: Statistical Methodology · doi:10.1111/j.2517-6161.1995.tb02031.x · Cited by 109,482
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · 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) · doi:10.18653/v1/n19-1423 · Cited by 54,325
Welch, P. (1967) · IEEE Transactions on Audio and Electroacoustics · doi:10.1109/tau.1967.1161901 · Cited by 12,277
Deep Contextualized Word Representations
Peters, M., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K., & Zettlemoyer, L. (2018) · Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) · doi:10.18653/v1/n18-1202 · Cited by 11,983
Fischl, B. (2012) · NeuroImage · doi:10.1016/j.neuroimage.2012.01.021 · Cited by 9,526
Hierarchical Attention Networks for Document Classification
Yang, Z., Yang, D., Dyer, C., He, X., Smola, A., & Hovy, E. (2016) · Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies · doi:10.18653/v1/n16-1174 · Cited by 4,712
Transformer-XL: Attentive language models beyond a fixed-length context
Dai, Z., Yang, Z., Yang, Y., Carbonell, J., Le, Q., & Salakhutdinov, R. (2019) · Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/p19-1285 · Cited by 3,169
Longformer: The Long-Document Transformer
Beltagy, I., Peters, M. E., & Cohan, A. (2020) · arXiv · doi:10.48550/arxiv.2004.05150 · Cited by 2,281
Intersubject synchronization of cortical activity during natural vision
Hasson, U., Nir, Y., Levy, I., Fuhrmann, G., & Malach, R. (2004) · Science · doi:10.1126/science.1089506 · Cited by 1,839
Self-citation
Research works printed but not linked: 3
- 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
- Radford, A. et al. Language models are unsupervised multitask learners. OpenAI Blog 1, 9 (2019).
- 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
Generative language reconstruction from brain recordings
Ye, Z., Ai, Q., Liu, Y., de Rijke, M., Zhang, M., Lioma, C., & Ruotsalo, T. (2025) · Communications Biology · doi:10.1038/s42003-025-07731-7 · Cited by 36
“Additionally, Tikochinski et al. 20 suggest that LLMs can process large contextual windows while the brain may preferentially focus on the content perceived most recently.”
Event segmentation applications in large language model enabled automated recall assessments
Panela, R. A., Barnett, A. J., Barense, M. D., & Herrmann, B. (2025) · Communications Psychology · doi:10.1038/s44271-025-00359-7 · Cited by 10
Lian, Z., & Kong, F. (2025) · Water Air & Soil Pollution · doi:10.1007/s11270-025-08055-2 · Cited by 9
“In addition, the complex interactions between microbial communities and environmental conditions makes it difficult to accurately predict the reactor performances (Tikochinski et al., 2025).”
Towards large language models with human-like episodic memory
Dong, C., Lu, Q., Norman, K. A., & Michelmann, S. (2025) · Trends in Cognitive Sciences · doi:10.1016/j.tics.2025.06.016 · Cited by 28
Uncovering a Timescale Hierarchy by Studying the Brain in a Natural Context
Hasson, U. (2025) · Journal of Neuroscience · doi:10.1523/jneurosci.2368-24.2025 · Cited by 13
Self-citation
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)
- PMC
- PMC11748659
- OpenAlex
- W4406513948
- Cited by
- 24
- References
- 35
- Citations by year
- 2026: 10 · 2025: 14
- FWCI
- 13.68
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by-nc-nd
- creativecommons.org/licenses/by-nc-nd/4.0 (tdm)
- creativecommons.org/licenses/by-nc-nd/4.0 (vor)
- 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.