CoChat
Sign Up
Cited by 345Neurobiology of Language and BilingualismNatural Language Processing TechniquesTopic Modeling

Dissociating language and thought in large language models

Open access
Mahowald, Kyle, Ivanova, Anna A., Blank, Idan A., Kanwisher, Nancy, Tenenbaum, Joshua B., Fedorenko, Evelina·2024·Trends in Cognitive Sciences
DOI 10.1016/j.tics.2024.01.011PMC PMC11416727
Open this paper in CoChat — free account

Abstract

Large Language Models (LLMs) have come closest among all models to date to mastering human language, yet opinions about their linguistic and cognitive capabilities remain split. Here, we evaluate LLMs using a distinction between formal linguistic competence—knowledge of linguistic rules and patterns—and functional linguistic competence—understanding and using language in the world. We ground this distinction in human neuroscience, which has shown that formal and functional competence rely on different neural mechanisms. Although LLMs are surprisingly good at formal competence, their performance on functional competence tasks remains spotty and often requires specialized fine-tuning and/or coupling with external modules. We posit that models that use language in humanlike ways would need to master both of these competence types, which, in turn, could require the emergence of mechanisms specialized for formal linguistic competence, distinct from functional competence.

Integrity signals

Registry

Crossref · 10.1016/j.tics.2024.01.011

Cites

Statistics

Nothing to recompute.

Claims the paper makes

0 of 155 cited sources checked · 0 supported · 0 flagged

  • Findingcitation

    LLMs show sensitivity to the surface form of the comparative correlative construction 'the better the syntax, the better the semantics'.

    “Models are also sensitive to the form of the comparative correlative “the better the syntax, the better the semantics” [107].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [107] Weissweiler L, Hofmann V, Köksal A, and Schütze H. The better your syntax, the better your semantics? probing pretrained language models for the English comparative correlative. In Goldberg Y, Kozareva Z, and Zhang Y, editors, Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10859– 10882, Abu Dhabi, United Arab Emirates, December 2022. Association for Computational Linguistics. source

      Not checked in the free run.

  • Findingcitation

    Inferences based on comparative correlative sentences can be challenging for LLMs.

    “Indeed, it appears that inferences based on these sentences can be a challenge”
    Quote in the paperSubstantial claim
  • Findingreasoning

    LLMs routinely generate false statements, informally known as hallucinations.

    “First, LLMs routinely generate false statements, informally known as “hallucinations”.”
    Quote in the paperSubstantial claim
  • Findingcitation

    Reasoning tests that break common co-occurrence patterns in the input or require multi-step operations lead to model failure.

    “Reasoning tests that break common co-occurrence patterns in the input or require multi-step operations also lead to model failure [144, 145].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [144] Valmeekam K, Olmo A, Sreedharan S, and Kambhampati S. Large language models still can’t plan (a benchmark for llms on planning and reasoning about change). In NeurIPS 2022 Foundation Models for Decision Making Workshop, 2022. source

      Not checked in the free run.

    • cites [145] Wu Z, Qiu L, Ross A, Akyürek E, Chen B, Wang B, Kim N, Andreas J, and Kim Y. Reasoning or reciting? exploring the capabilities and limitations of language models through counterfactual tasks. arXiv preprint arXiv:2307.02477, 2023. source

      Not checked in the free run.

  • Findingcitation

    LLM outputs are often inconsistent: the same prompt phrased in different ways can elicit different responses.

    “Second, LLM outputs are often inconsistent: the same prompt phrased in different ways can elicit different responses [156].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [156] Sclar M, Choi Y, Tsvetkov Y, and Suhr A. Quantifying language models’ sensitivity to spurious features in prompt design or: How I learned to start worrying about prompt formatting. arXiv preprint arXiv:2310.11324, 2023. source

      Not checked in the free run.

  • May combine two claims. 32
  • Classified as probably not a substantial claim. 49

  • May combine two claims.

    Findingnone given

    LLMs today generate highly coherent, grammatical texts that can be indistinguishable from human output.

    “LLMs today generate highly coherent, grammatical texts that can be indistinguishable from human output.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingcitation

    GPT-4 shows good performance on small-digit but not large-digit mathematical operations.

    “GPT-4 similarly shows good performance on small-digit but not large-digit mathematical operations [143].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [143] Dziri N, Lu X, Sclar M, Li XL, Jiang L, Lin BY, Welleck S, West P, Bhagavatula C, Bras RL, Hwang JD, Sanyal S, Ren X, Ettinger A, Harchaoui Z, and Choi Y. Faith and fate: Limits of transformers on compositionality. In Thirty-seventh Conference on Neural Information Processing Systems, 2023.

      Not checked in the free run.

  • May combine two claims.

    Findingreasoning

    LLMs are not perfect learners of abstract linguistic rules, but neither are humans.

    “These models are not perfect learners of abstract linguistic rules, but neither are humans.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    GPT-3 performs well on two-digit addition and subtraction but not on three-digit addition or two-digit multiplication.

    “GPT-3 performs well on two-digit addition and subtraction but not on more complex tasks, such as three-digit addition or two-digit multiplication [69].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [69] Brown TB, Mann B, Ryder N, Subbiah M, Kaplan J, Dhariwal P, Neelakantan A, Shyam P, Sastry G, Askell A, Agarwal S, Herbert-Voss A, Krueger G, Henighan T, Child R, Ramesh A, Ziegler DM, Wu J, Winter C, Hesse C, Chen M, Sigler E, Litwin M, Gray S, Chess B, Clark J, Berner C, McCandlish S, Radford A, Sutskever I, and Amodei D. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems, 2020. source

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    Such dissociated approaches exist but have not yet reached dominance, typically because of relatively low coverage of existing knowledge bases.

    “Such approaches exist [e.g., 161] but have not yet reached dominance in the field, typically because of relatively low coverage of existing knowledge bases.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [161] Borgeaud S, Mensch A, Hoffmann J, Cai T, Rutherford E, Millican K, Van Den Driessche GB, Lespiau J-B, Damoc B, Clark A, Improving language models by retrieving from trillions of tokens. In International conference on machine learning, pages 2206– 2240. PMLR, 2022.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Findingreasoning

    LLMs exhibit knowledge of hierarchical structure and linguistic abstractions.

    “they exhibit knowledge of hierarchical structure and linguistic abstractions”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    LLMs resemble human brain responses during language processing.

    “while resembling human brain responses during language processing.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    Modular models are more efficient, can be trained on smaller datasets, and have lower compute demands during inference.

    “are more efficient (i.e., can be trained on smaller datasets and have lower compute demands during inference)”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingcitation

    Codex can translate text input into meaningful structured probabilistic programs.

    “a version of GPT-3 fine-tuned to generate both natural language and code (Codex) can translate text input into meaningful structured probabilistic programs”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [149] Wong L, Grand G, Lew AK, Goodman ND, Mansinghka VK, Andreas J, and Tenenbaum JB. From word models to world models: Translating from natural language to the probabilistic language of thought. arXiv preprint arXiv:2306.12672, 2023. source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Findingcitation

    Inference in these programs can be used to reason over relational domains, grounded domains, and situations requiring planning and understanding the plans of others.

    “inference in these programs can be used to reason over relational domains (like kinship systems), grounded domains (like visual scenes), and situations that require planning and understanding the plans of others.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [149] Wong L, Grand G, Lew AK, Goodman ND, Mansinghka VK, Andreas J, and Tenenbaum JB. From word models to world models: Translating from natural language to the probabilistic language of thought. arXiv preprint arXiv:2306.12672, 2023. source

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    A claim that instruction-tuned LLMs have mastered theory-of-mind tasks was countered by a demonstration that including basic controls brought LLM performance to below-chance levels.

    “A bold claim that instruction-tuned LLMs have mastered theory-of-mind tasks [185] was quickly countered by a demonstration that including basic controls (such as character Y being told about the true object location) brought LLM performance to below-chance levels [186].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [185] Kosinski M. Theory of mind may have spontaneously emerged in large language models. arXiv preprint arXiv:2302.02083, 2023. source

      Not checked in the free run.

    • cites [186] Ullman T. Large language models fail on trivial alterations to theory-of-mind tasks. arXiv preprint arXiv:2302.08399, 2023. source

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    LLMs exhibit uneven performance across pragmatic domains: their ability to interpret sarcasm or complete jokes was limited even as their metaphor comprehension abilities soared.

    “LLMs exhibit uneven performance across pragmatic domains: their ability to interpret sarcasm or to complete jokes was limited even as their metaphor comprehension abilities soared [184].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [184] Hu J, Floyd S, Jouravlev O, Fedorenko E, and Gibson E. A fine-grained comparison of pragmatic language understanding in humans and language models. In Rogers A, Boyd-Graber J, and Okazaki N, editors, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4194– 4213, Toronto, Canada, July 2023. Association for Computational Linguistics.

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    Recent RLHF-trained models show strong performance in interpreting non-literal utterances, such as metaphors and polite deceits, suggesting they can reach human or near-human performance on at least some pragmatic tasks.

    “Recent models, trained with RLHF, have shown strong performance in interpreting non-literal utterances, such as metaphors and polite deceits, suggesting they can reach human or near-human performance on at least some pragmatic tasks [184].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [184] Hu J, Floyd S, Jouravlev O, Fedorenko E, and Gibson E. A fine-grained comparison of pragmatic language understanding in humans and language models. In Rogers A, Boyd-Graber J, and Okazaki N, editors, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4194– 4213, Toronto, Canada, July 2023. Association for Computational Linguistics.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    LLM performance is not yet fully human-like.

    “Although their performance is not yet fully human-like”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    People typically assume that a sentence was produced by a rational, thinking agent.

    “When we hear a sentence, we typically assume that it was produced by a rational, thinking agent (another person).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Large language models have come closer than any previous models to mastering human language.

    “Large Language Models (LLMs) have come closest among all models to date to mastering human language”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Mastery of both competence types could require mechanisms specialized for formal linguistic competence, distinct from functional competence.

    “which, in turn, could require the emergence of mechanisms specialized for formal linguistic competence, distinct from functional competence.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given

    Significant formal linguistic abilities arise in models on the scale of GPT-2 or BERT and seem to plateau at a high level in contemporary LLMs.

    “Significant formal linguistic abilities arise in models on the scale of GPT-2 or BERT and seem to plateau at a high level in contemporary LLMs (Box 1).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Despite nearly complete loss of linguistic abilities, individuals with severe aphasia can have intact non-linguistic cognitive abilities: they can play chess, solve arithmetic problems, leverage world knowledge, reason about cause and effect, and navigate complex social situations.

    “However, despite the nearly complete loss of linguistic abilities, individuals with severe aphasia can have intact non-linguistic cognitive abilities: they can play chess, solve arithmetic problems, leverage their world knowledge to perform diverse tasks, reason about cause and effect, and navigate complex social situations [49].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [49] Fedorenko E and Varley RA. Language and thought are not the same thing: evidence from neuroimaging and neurological patients: Language versus thought. Annals of the New York Academy of Sciences, 1369(1): 132– 153, April 2016. 27096882 PMC4874898 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    A formal language system in isolation is useless unless it can interface with the rest of perception, cognition, and action.

    “A formal language system in isolation is useless unless it can interface with the rest of perception, cognition, and action.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisenone given

    In humans, language is robustly dissociated from the rest of high-level cognition, as well as from perception and action.

    “In humans, language is robustly dissociated from the rest of high-level cognition, as well as from perception and action.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    For modern LLMs, formal competence in English is near human-level.

    “for modern LLMs, formal competence in English is near human-level”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    People's day-to-day sentences are based on their world knowledge, reasoning abilities, and goals.

    “The sentences that people generate in day-to-day conversations are based on their world knowledge (“Not all birds can fly.”), their reasoning abilities (“You’re 15, you can’t go to a bar.”), and their goals (“Would you give me a ride, please?”).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The capacities required to use language to do things in the world are distinct from formal competence and depend crucially on non-linguistic cognition.

    “The capacities required to use language to do things in the world are distinct from formal competence and depend crucially on non-linguistic cognition (Figure 1).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The Turing test uses language as an interface to cognition.

    “The Turing test uses language as an interface to cognition, allowing a human participant to probe the knowledge and reasoning capacities of two conversation partners to determine which of them is a human and which is a machine.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [1] Turing AM. Computing Machinery and Intelligence. Mind, 59(October): 433– 60, 1950. Publisher: Oxford University Press. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The Turing test has shaped how society today thinks about machine intelligence.

    “Although the utility of the Turing test has since been questioned, it has undoubtedly shaped the way society today thinks of machine intelligence [2].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [2] Chang Y, Wang X, Wang J, Wu Y, Yang L, Zhu K, Chen H, Yi X, Wang C, Wang Y, A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 2023.

      Not checked in the free run.

  • May combine two claims.

    Interpretationnone given

    The functional competence of modern LLMs remains patchy, with results depending on specific functional competence domains and on tasks within those domains.

    “whereas their functional competence remains patchy, with results depending on specific functional competence domains and on tasks within those domains.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    The conflation of language and thought is unsurprising because encountering an entity that generates fluent sentences without a human identity is novel and uncanny.

    “This conflation is unsurprising: it is still novel, and thus uncanny, to encounter an entity that generates fluent sentences despite lacking a human identity.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning

    This sharp dissociation suggests that, in examining language models' functionality, we should separate their linguistic abilities from their abstract knowledge and reasoning abilities, which can be probed through a linguistic interface but require more than formal linguistic competence.

    “This sharp dissociation suggests that, in examining language models’ functionality, we should separate their linguistic abilities from their abstract knowledge and reasoning abilities, which can be probed—–and perhaps even learned–—through a linguistic interface, but which require more than formal linguistic competence.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Premisereasoning

    People often use other people's statements as a window into their minds.

    “Thus, we often use other people’s statements as a window into their minds.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Premisenone given

    Much of human cognition can be conveyed through language and thus learned from language, even if these capacities are not themselves inherently linguistic.

    “in part because so much of human cognition (commonsense reasoning, scientific knowledge, everyday knowledge) can be conveyed through language and thus learned from language—even if these capacities are not themselves inherently linguistic.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given

    N-grams and word embedding models achieved some success in various natural language processing domains but never approached human-level performance on general language tasks like text generation.

    “N-grams and word embedding models achieved some success in various domains in natural language processing (e.g., spelling correction, spam classification, sentiment analysis). However, they never approached human-level performance on general language tasks like text generation”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    The 'good at language -> good at thought' fallacy leads people to mistakenly attribute intelligence and intentionality to even basic dialog systems.

    “The “good at language -> good at thought” fallacy makes it easy to confuse the two, leading people to mistakenly attribute intelligence and intentionality to even the most basic dialog systems”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Both formal and functional linguistic competence are essential components of human language use.

    “Both formal and functional linguistic competence are essential components of human language use: an effective communicator needs to both generate grammatical, meaningful utterances and strategically use those utterances to achieve diverse, context-dependent goals [12, 13].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [12] Grice H. Logic and conversation. In Cole P and Morgan JL, editors, Syntax and Semantics, Vol. 3, Speech Acts, pages 41– 58. Academic Press, New York, 1975.

      Not checked in the free run.

    • cites [13] Clark HH. Arenas of Language Use. University of Chicago Press, 1992.

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The language network supports both comprehension (spoken, written, and signed) and production.

    “This language network supports both comprehension (spoken, written, and signed) [30, 31, 32, 33] and production [34, 35]”
    Quote in the paperSubstantial claim0 of 6 sources checked
    • cites [30] Deniz F, Nunez-Elizalde AO, Huth AG, and Gallant JL. The Representation of Semantic Information Across Human Cerebral Cortex During Listening Versus Reading Is Invariant to Stimulus Modality. Journal of Neuroscience, 39(39): 7722– 7736, September 2019. Publisher: Society for Neuroscience Section: Research Articles. 31427396 PMC6764208 source

      Not checked in the free run.

    • cites [31] Fedorenko E, Hsieh P-J, Nieto-Castañón A, Whitfield-Gabrieli S, and Kanwisher N. New method for fMRI investigations of language: defining ROIs functionally in individual subjects. Journal of Neurophysiology, 104(2): 1177– 1194, August 2010. 20410363 PMC2934923 source

      Not checked in the free run.

    • cites [32] MacSweeney M, Woll B, Campbell R, McGuire PK, David AS, Williams SCR, Suckling J, Calvert GA, and Brammer MJ. Neural systems underlying British Sign Language and audio-visual English processing in native users. Brain, 125(7): 1583– 1593, July 2002. 12077007 source

      Not checked in the free run.

    • cites [33] Scott TL, Gallée J, and Fedorenko E. A new fun and robust version of an fMRI localizer for the frontotemporal language system. Cognitive Neuroscience, 8(3): 167– 176, 2017. 27386919 source

      Not checked in the free run.

    • cites [34] Menenti L, Gierhan SME, Segaert K, and Hagoort P. Shared language: overlap and segregation of the neuronal infrastructure for speaking and listening revealed by functional MRI. Psychological Science, 22(9): 1173– 1182, September 2011. 21841148 source

      Not checked in the free run.

    • cites [35] Hu J, Small H, Kean H, Takahashi A, Zekelman L, Kleinman D, Ryan E, Nieto-Castañón A, Ferreira V, and Fedorenko E. Precision fmri reveals that the language-selective network supports both phrase-structure building and lexical access during language production. Cerebral Cortex, 33(8): 4384– 4404, 2023. 36130104 PMC10110436 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    The pessimism about purely statistical approaches is now challenged by LLMs.

    “This pessimism is now challenged by LLMs.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    fMRI reveals that the language network responds reliably when people listen to, read, or generate sentences, but not when they perform arithmetic tasks, engage in logical reasoning, understand computer programs, listen to music, categorize objects or events, reason about people's mental states, or process non-verbal communicative information like facial expressions or gestures.

    “it responds reliably when people listen to, read, or generate sentences, but not when they perform arithmetic tasks, engage in logical reasoning, understand computer programs, listen to music, categorize objects or events, reason about people’s mental states, or process non-verbal communicative information like facial expressions or gestures [37, 41, 42, 43, 44, 45, 46, 47, 48].”
    Quote in the paperSubstantial claim0 of 9 sources checked
    • cites [37] Fedorenko E, Behr MK, and Kanwisher N. Functional specificity for high-level linguistic processing in the human brain. Proceedings of the National Academy of Sciences, 108(39): 16428– 16433, September 2011. PMC3182706 21885736 source

      Not checked in the free run.

    • cites [41] Amalric M and Dehaene S. Origins of the brain networks for advanced mathematics in expert mathematicians. Proceedings of the National Academy of Sciences of the United States of America, 113(18): 4909– 4917, May 2016. 27071124 PMC4983814 source

      Not checked in the free run.

    • cites [42] Benn Y, Ivanova AA, Clark O, Mineroff Z, Seikus C, Silva JS, Varley R, and Fedorenko E. The language network is not engaged in object categorization. Cerebral Cortex, 33(19): 10380– 10400, 2023. 37557910 PMC10545444 source

      Not checked in the free run.

    • cites [43] Chen X, Affourtit J, Ryskin R, Regev TI, Norman-Haignere S, Jouravlev O, Malik-Moraleda S, Kean H, Varley R, and Fedorenko E. The human language system, including its inferior frontal component in “Broca’s area,” does not support music perception. Cerebral Cortex, 33(12): 7904– 7929, 04 2023. 37005063 PMC10505454 source

      Not checked in the free run.

    • cites [44] Deen B, Koldewyn K, Kanwisher N, and Saxe R. Functional Organization of Social Perception and Cognition in the Superior Temporal Sulcus. Cerebral Cortex, 25(11): 4596– 4609, November 2015. 26048954 PMC4816802 source

      Not checked in the free run.

    • cites [45] Jouravlev O, Zheng D, Balewski Z, Pongos ALA, Levan Z, Goldin-Meadow S, and Fedorenko E. Speech-accompanying gestures are not processed by the language-processing mechanisms. Neuropsychologia, 132: 107132, September 2019. 31276684 PMC6708375 source

      Not checked in the free run.

    • cites [46] Liu Y-F, Kim J, Wilson C, and Bedny M. Computer code comprehension shares neural resources with formal logical inference in the fronto-parietal network. eLife, 9: e59340, dec 2020. 33319745 PMC7738180 source

      Not checked in the free run.

    • cites [47] Monti MM, Parsons LM, and Osherson DN. Thought beyond language: neural dissociation of algebra and natural language. Psychological Science, 23(8): 914– 922, August 2012. 22760883 source

      Not checked in the free run.

    • cites [48] Paunov AM, Blank IA, Jouravlev O, Mineroff Z, Gallée J, and Fedorenko E. Differential Tracking of Linguistic vs. Mental State Content in Naturalistic Stimuli by Language and Theory of Mind (ToM) Brain Networks. Neurobiology of Language, pages 1– 29, June 2022. 37216061 PMC10158571

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    These models have succeeded on tests of general language understanding developed by the NLP community, such as GLUE tasks, and on tests of linguistic competence in English and other languages with massive corpora available.

    “These models have succeeded not just on tests of general language understanding developed by the NLP community (e.g., GLUE tasks [75]), but, critically for our purposes, on tests of linguistic competence in English and other languages with massive corpora available”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [75] Wang A, Singh A, Michael J, Hill F, Levy O, and Bowman SR. GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353– 355, 2018.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Formal linguistic competence in LLMs improves drastically with the amount of training data.

    “whereas formal linguistic competence in LLMs improves drastically with the amount of training data”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Formal and functional linguistic competence are robustly dissociable in the human brain.

    “Our motivation for the formal/functional distinction comes from the human brain, where these skills are robustly dissociable.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Our heuristics for understanding what a language model is doing are broken.

    “Thus, our heuristics for understanding what a language model is doing—heuristics that emerged from our language experience with other humans—are broken.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Models achieve similarly impressive results on other linguistic benchmarks like SyntaxGym, and there are dozens of investigations of specific complex linguistic phenomena.

    “Models achieve similarly impressive results on other linguistic benchmarks like SyntaxGym [79], and there are now dozens of investigations of specific complex linguistic phenomena”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [79] Gauthier J, Hu J, Wilcox E, Qian P, and Levy R. SyntaxGym: An Online Platform for Targeted Evaluation of Language Models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 70– 76, Online, July 2020. Association for Computational Linguistics.

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    BERT succeeds partially at applying morphosyntactic rules to novel words: it learns to generalize, but only after repeated examples.

    “They conclude that BERT succeeds partially at this task: it does learn to generalize, but only after repeated examples”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [90] Kim N and Smolensky P. Testing for grammatical category abstraction in neural language models. In Ettinger A, Pavlick E, and Prickett B, editors, Proceedings of the Society for Computation in Linguistics 2021, pages 467– 470, Online, February 2021. Association for Computational Linguistics.

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    A bigram model could learn local subject-verb agreement but would not be able to learn agreement when the subject and verb are arbitrarily far apart.

    “But such a model would not be able to learn that the subject and verb agree even if arbitrarily far apart”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Neither LLMs nor the language network show clear spatial segregation for syntactic and semantic processing.

    “Neither system shows clear spatial segregation for syntactic and semantic processing (LLMs: [95, 115]; brain: [38, 114]), indicating that these processes are tightly functionally coupled in both.”
    Quote in the paperSubstantial claim0 of 4 sources checked
    • cites [95] Tenney I, Das D, and Pavlick E. BERT rediscovers the classical NLP pipeline. In Korhonen A, Traum D, and Màrquez L, editors, Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593– 4601, Florence, Italy, July 2019. Association for Computational Linguistics. source

      Not checked in the free run.

    • cites [115] Huang JY, Huang K-H, and Chang K-W. Disentangling semantics and syntax in sentence embeddings with pre-trained language models. In Toutanova K, Rumshisky A, Zettlemoyer L, Hakkani-Tur D, Beltagy I, Bethard S, Cotterell R, Chakraborty T, and Zhou Y, editors, Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1372– 1379, Online, June 2021. Association for Computational Linguistics. source

      Not checked in the free run.

    • cites [38] Fedorenko E, Blank IA, Siegelman M, and Mineroff Z. Lack of selectivity for syntax relative to word meanings throughout the language network. Cognition, 203: 104348, October 2020. 32569894 PMC7483589 source

      Not checked in the free run.

    • cites [114] Reddy AJ and Wehbe L. Can fMRI reveal the representation of syntactic structure in the brain? In Ranzato M, Beygelzimer A, Dauphin Y, Liang PS, and Vaughan JW, editors, Advances in Neural Information Processing Systems, volume 34, pages 9843– 9856. Curran Associates, Inc., 2021.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    In human languages, words are combined to make compositional meanings; individual words' meanings are not simply added linearly but can be combined hierarchically into tree-like structures.

    “In human languages, words are combined to make compositional meanings. In a multi-word sentence, the individual words’ meanings do not simply get added linearly one by one. Instead, they can be combined hierarchically into tree-like structures.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    A more human-like approach to world knowledge representation might require dissociating linguistic representation/processing and world knowledge storage/updates.

    “A more human-like approach to world knowledge representation might require dissociating linguistic representation/processing and world knowledge storage/updates.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisereasoning

    Four key capacities not specific to language but crucial for real-life language use are formal reasoning, world knowledge, situation modeling, and social reasoning.

    “We focus on four key capacities that are not language-specific but are nevertheless crucial for language use in real-life settings: i) formal reasoning—a host of abilities including logical and mathematical reasoning, computational thinking, and novel problem solving; ii) world knowledge—factual and commonsense knowledge about agents, objects, properties, actions, events, and ideas; iii) situation”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The four capacities (formal reasoning, world knowledge, situation modeling, social reasoning) are the key capacities for real-life language use.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Premisenone given

    Understanding a sentence, reasoning about its implications, and deciding how to respond rely on cognitive capacities beyond formal competence.

    “Understanding a sentence, reasoning about its implications, and deciding how to respond all rely on cognitive capacities that go beyond formal competence.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Functional competence of LLMs is uneven, often requiring specialized fine-tuning and/or lacking human-like robustness and generality.

    “We conclude that, unlike formal competence, functional competence of LLMs is uneven, often requiring specialized fine-tuning and/or lacking human-like robustness and generality.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation

    The most commonly cited reason for LLM formal reasoning failures is the failure of artificial neural nets to generalize to patterns outside their training distribution.

    “The most commonly cited reason for these failures is the failure of artificial neural nets to generalize to patterns outside their training distribution [145, 146].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [145] Wu Z, Qiu L, Ross A, Akyürek E, Chen B, Wang B, Kim N, Andreas J, and Kim Y. Reasoning or reciting? exploring the capabilities and limitations of language models through counterfactual tasks. arXiv preprint arXiv:2307.02477, 2023. source

      Not checked in the free run.

    • cites [146] Zhang H, Li LH, Meng T, Chang K-W, and Van den Broeck G. On the paradox of learning to reason from data. In Elkind E, editor, Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 3365– 3373. International Joint Conferences on Artificial Intelligence Organization, 8 2023. Main Track.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisecitation

    Individuals with language deficits may struggle to produce grammatical utterances and retrieve contextually appropriate words, but their ability to reason about objects and events presented non-linguistically often remains intact.

    “Individuals with language deficits may struggle to produce grammatical utterances and retrieve contextually appropriate words, but their ability to reason about objects and events presented non-linguistically often remains intact [42, 151].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [42] Benn Y, Ivanova AA, Clark O, Mineroff Z, Seikus C, Silva JS, Varley R, and Fedorenko E. The language network is not engaged in object categorization. Cerebral Cortex, 33(19): 10380– 10400, 2023. 37557910 PMC10545444 source

      Not checked in the free run.

    • cites [151] Ivanova AA, Mineroff Z, Zimmerer V, Kanwisher N, Varley R, and Fedorenko E. The Language Network is Recruited but Not Required for Nonverbal Event Semantics. Neurobiology of Language, pages 1– 26, January 2021. Publisher: MIT Press. PMC10158592 37216147 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    Evaluation issues have led to a particularly large amount of overclaiming about models' functional competence.

    “evaluation issues can occur in studies of either formal or functional competence, but we believe they have led to a particularly large amount of overclaiming about models’ functional competence.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Premisenone given

    Real-life language use is impossible without non-linguistic cognitive skills.

    “Real-life language use is impossible without non-linguistic cognitive skills.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    The multiple demand network is engaged in logic, mathematical reasoning, physical reasoning, and computer code comprehension.

    “these regions are engaged in many cognitively demanding tasks: logic [47], mathematical reasoning [41], physical reasoning [139], and computer code comprehension [46, 140].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [47] Monti MM, Parsons LM, and Osherson DN. Thought beyond language: neural dissociation of algebra and natural language. Psychological Science, 23(8): 914– 922, August 2012. 22760883 source

      Not checked in the free run.

    • cites [41] Amalric M and Dehaene S. Origins of the brain networks for advanced mathematics in expert mathematicians. Proceedings of the National Academy of Sciences of the United States of America, 113(18): 4909– 4917, May 2016. 27071124 PMC4983814 source

      Not checked in the free run.

    • cites [139] Fischer J, Mikhael JG, Tenenbaum JB, and Kanwisher N. Functional neuroanatomy of intuitive physical inference. Proceedings of the National Academy of Sciences, 113(34): E5072– E5081, August 2016. PMC5003259 27503892

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Formal and functional linguistic competence in the human brain draw on separate neural circuits.

    “which in the human brain draw on separate neural circuits”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationcitation

    Augmenting an LLM with a symbolic tracker of entity states and character beliefs is a solution to overcome theory of mind limitations.

    “One solution to overcome these limitations has been to augment an LLM with a symbolic tracker of entity states and character beliefs [191]”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [191] Sclar M, Kumar S, West P, Suhr A, Choi Y, and Tsvetkov Y. Minding language models’ (lack of) theory of mind: A plug-and-play multi-character belief tracker. In Rogers A, Boyd-Graber J, and Okazaki N, editors, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13960– 13980, Toronto, Canada, July 2023. Association for Computational Linguistics.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationcitation

    Emergent modularity may constitute an optimal solution to many complex tasks.

    “which suggests that emergent modularity may constitute an optimal solution to many complex tasks.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Situation modeling in LLMs faces two main challenges: extracting information from many sentences in a row, and integrating incoming inputs to update information about entities and their states.

    “Situation modeling in LLMs faces two main challenges: (1) extracting information from many sentences in a row; (2) integrating incoming inputs to appropriately update information about entities and their states.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Models that succeed at real-life language use will need to be modular, mimicking the division of labor between formal and functional competence in the human brain.

    “we suggest that models that succeed at real-life language use will need to be modular, mimicking the division of labor between formal and functional competence in the human brain.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Linguistic meaning radically depends on context.

    “Wittgenstein famously used single-word utterances like this to show that linguistic meaning radically depends on context.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    The theory of mind network is a set of brain regions engaged when a person attempts to infer somebody's mental state, with or without language.

    “The most relevant to our current discussion is the theory of mind network [178], a set of brain regions that are engaged when their owner is attempting to infer somebody’s mental state (with our without the use of language; [179, 180]).”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [178] Gopnik A and Wellman HM. Why the child’s theory of mind really is a theory. Mind and Language, 7(1–2): 145– 71, 1992. source

      Not checked in the free run.

    • cites [179] Saxe R and Kanwisher N. People thinking about thinking people. The role of the temporo-parietal junction in "theory of mind". NeuroImage, 19(4): 1835– 1842, August 2003. 12948738 source

      Not checked in the free run.

    • cites [180] Jacoby N, Bruneau E, Koster-Hale J, and Saxe R. Localizing Pain Matrix and Theory of Mind networks with both verbal and non-verbal stimuli. NeuroImage, 126: 39– 48, February 2016. 26589334 PMC4733571 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Theory of mind tasks require both social knowledge and the ability to maintain a situation model.

    “These tasks require both social knowledge and the ability to maintain a situation model.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    Different types of pragmatic reasoning can be classified either as formal or as functional competence.

    “Thus, different types of pragmatic reasoning can be classified either as formal or as functional competence.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationcitation

    Their approach demonstrates a promising avenue for integrating what LLMs succeed at (formal linguistic competence) with other cognitive modules that benefit from symbolic structure and abstraction.

    “Their approach demonstrates a promising avenue for integrating what LLMs succeed at (namely, formal linguistic competence) with other cognitive modules that benefit from symbolic structure and abstraction.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Future advances in developing intelligent systems will require combining language models with models that represent abstract knowledge and support complex reasoning, rather than expecting a single model trained with a single word prediction objective to do it all.

    “future advances in developing intelligent systems will require combining language models with models that represent abstract knowledge and support complex reasoning, rather than expecting a single model (trained with a single word prediction objective) to do it all.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The human mind and brain is a good example of generalized intelligence.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    More promising solutions will come in the form of modular architectures-built-in or emergent-that integrate language processing with additional systems that carry out perception, reasoning, and action.

    “more promising solutions will come in the form of modular architectures—built-in or emergent—that, like the human brain, integrate language processing with additional systems that carry out perception, reasoning, and action.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    LLMs are underused in linguistics as candidate models of human language processing.

    “As such, these LLMs are underused in linguistics as candidate models of human language processing.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    We need benchmarks that cleanly separate formal and functional linguistic competence.

    “we need benchmarks that cleanly separate formal and functional linguistic competence (Box 5).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    There is a strict separation of language and non-linguistic capabilities in the human mind.

    “given a strict separation of language and non-linguistic capabilities in the human mind”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Brain areas that support language processing in humans cannot do math, solve logical problems, or track the meaning of a story across sentences or paragraphs.

    “the brain areas that support language processing in humans also cannot do math, solve logical problems, or even track the meaning of a story across sentences or paragraphs.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    A modular model architecture is much better aligned with the brain's functional architecture for language, which includes separate components for formal and functional competence.

    “A modular model architecture is much better aligned with the brain’s functional architecture for language, which includes separate components for formal and functional competence.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Formally and rigorously evaluating functional competence in LLMs will be informative for both science and engineering.

    “Formally and rigorously evaluating functional competence in LLMs will be informative for both science and engineering”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    Explicitly disentangling formal and functional competence skills at the architectural level is perhaps the most fail-safe path toward ensuring that an AI model uses language in a humanlike way.

    “explicitly disentangling formal and functional competence skills at the architectural level is perhaps the most fail-safe path toward ensuring that an AI model uses language in a humanlike way.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Systems with different underlying architectures can, in principle, exhibit similar behaviors.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    These capabilities rely on brain machinery distinct from language processing machinery.

    “and rely on machinery in the human brain distinct from language processing machinery.”
    Quote in the paperNot a substantial claim
75 of 75

  • The paper relies on this without support.

    The four capacities (formal reasoning, world knowledge, situation modeling, social reasoning) are the key capacities for real-life language use.

    “We focus on four key capacities that are not language-specific but are nevertheless crucial for language use in real-life settings”

  • The paper relies on this without support.

    An average conversation requires all four capacities.

    “An average conversation requires the use of all these capacities”

  • The paper relies on this without support.

    None of these capacities are specific to language use.

    “yet none of them are specific to language use.”

  • The paper relies on this without support.

    Language inputs contain vast quantities of information about the world.

    “language inputs contain vast quantities of information about the world”

  • The paper relies on this without support.

    Language is both a crucial data source and representational substrate for much of people's world knowledge.

    “language is both a crucial data source and representational substrate for much of people’s world knowledge”

  • The paper relies on this without support.

    The human mind and brain is a good example of generalized intelligence.

    “the human mind and brain—a good example of generalized intelligence—as a guide”

  • The paper relies on this without support.

    Systems with different underlying architectures can, in principle, exhibit similar behaviors.

    “systems with different underlying architectures (e.g., modular vs. non-modular) can, in principle, exhibit similar behaviors.”

  • The paper relies on this without support.

    Much more input is required than a typical child gets exposed to.

    “even if, as of now, much more input is required than a typical child gets exposed to”
8 of 8

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?

  • LLMs' inner workings are fundamentally different to human cognition, and they are incapable of feeling or understanding emotion or language in any genuine sense.

    “it remains obviously true that LLMs' inner workings are fundamentally different to human cognition (63), and that they are incapable of feeling or even understanding emotion or language in any genuine sense (27, 64).”

    From The benefits and dangers of anthropomorphic conversational agents (2025) · cites this paper as [64]

    Citation (63) is attached to the claim about LLMs’ inner workings, not the separate claim about feeling or understanding, which cites (27, 64). In “LLMs are predictive of activity in the human language network” (page not provided), the paper says LLMs and the human language network have similar internal architectures and representations, though their correspondence is not one-to-one. That is contrary to the unqualified claim that their inner workings are fundamentally different. (LLM, conf=0.86)

    Checked by a reasoning model — may be wrong.

    Experimental — accuracy not yet measured

  • LLMs do not possess true empathy or social understanding.

    “without possessing any true empathy or social understanding”

    From The benefits and dangers of anthropomorphic conversational agents (2025) · cites this paper as [64]

    The Social reasoning section reports limitations in LLMs’ theory-of-mind performance, supporting a qualified concern about social understanding. But it also reports human or near-human performance on some pragmatic tasks; it does not establish that LLMs have no social understanding, and it does not address whether they possess true empathy. Page numbers are unavailable in our copy. (LLM, conf=0.86)

    Experimental — accuracy not yet measured

  • Other work reports failures in reasoning fidelity and intermediate consistency, where fluent outputs may mask unstable internal reasoning processes.

    “Other work reports failures in reasoning fidelity and intermediate consistency, where fluent outputs may mask unstable internal reasoning processes [49,50].”

    From Diagnosing the limits of large language models for structured scientific evidence extraction (2026) · cites this paper as [50]

    The Formal Reasoning section reports failures on multi-step reasoning tasks and says that generating intermediate computation steps does not make results foolproof. The language-thought conflation section also warns that coherent text can be mistaken for evidence of reasoning ability. Neither passage establishes failures of reasoning fidelity or intermediate consistency, or describes unstable internal reasoning processes; page numbers are unavailable in our copy. (LLM, conf=0.85)

    Experimental — accuracy not yet measured

  • GPT-3, in response to the prompt Get your sofa on the roof, suggested getting a very strong friend to help carry it up a ladder and breaking the windows to make room.

    “GPT-3's suggestions, in response to the prompt”

    From Large language models and the problem of rhetorical debt (2025) · cites this paper as Mahowald et al. 2024

  • A chatbot has formal linguistic competence without functional linguistic competence, an understanding of the physical-social-cultural space in which humans operate.

    “A chatbot has”

    From Large language models and the problem of rhetorical debt (2025) · cites this paper as Mahowald et al. 2024

The citation lists haven’t been fetched yet

References 237 · Cited by 345

Research works printed but not linked: 5

  1. Dalrymple, 2019, An amazing four doctoral dissertations, Argumentum, 15, 2019
  2. Bucholtz, 2004, Language and identity, Companion Linguist. Anthropol., 1, 369
  3. Kosinski, 2023, Theory of mind may have spontaneously emerged in large language models, arXiv
  4. Vaswani, 2017, Attention is all you need, Adv. Neural Inf. Proces. Syst., 30, 5998
  5. Radford, 2019, Language models are unsupervised multitask learners, OpenAI Blog, 1, 2019

Authors
Mahowald, K., Ivanova, A. A., Blank, I. A., Kanwisher, N., Tenenbaum, J. B., & Fedorenko, E.
Affiliations
  • The University of Texas at Austin (US)
  • Georgia Institute of Technology (US)
  • University of California, Los Angeles (US)
  • Massachusetts Institute of Technology (US)
Corresponding author
  • Mahowald, Kyle
  • Ivanova, Anna A.
  • Blank, Idan A.
  • Kanwisher, Nancy
  • Tenenbaum, Joshua B.
  • Fedorenko, Evelina
ORCID
Venue
Trends in Cognitive Sciences
Publisher
Elsevier BV
Published
2024-06
Volume
28
Issue
6
Pages
517-540
ISSN
1364-6613 (ISSN-L)
DOI
10.1016/j.tics.2024.01.011
PMC
PMC11416727
OpenAlex
W4392935324
Semantic Scholar
cf36c0c47e1f1a9bb5285c638bdd77244113bbae
Cited by
345
References
237
Citations by year
2026: 112 · 2025: 132 · 2024: 88 · 2023: 13
FWCI
21.35
Type
journal article
Language
en
Open access
Licence
Locations
  • Trends in Cognitive Sciences (acceptedVersion)
  • PubMed (publishedVersion)
  • PubMed Central (acceptedVersion)
Funders
  • NIH (R01-DC016607, U01-NS121471, R01-DC016950)
  • National Science Foundation (2104995)
  • McGovern Institute for Brain Research, Massachusetts Institute of Technology
  • Simons Foundation
  • Simons Center for the Social Brain, Massachusetts Institute of Technology
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Neurobiology of Language and Bilingualism
  • Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
Keywords
large language models, language and thought, language comprehension, linguistic rules, neural mechanisms of language, fine-tuning, external module
MeSH terms
Humans, Language, Linguistics, Thinking

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.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-06.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-06.

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