Evidence from formal logical reasoning reveals that the language of thought is not natural language — part 1/10

Cleaned-text code points [0, 13929). Bibliography omitted here; source.md includes it.

Evidence from Formal Logical Reasoning Reveals that the Language of Thought is not
Natural Language

Hope Kean1,2, Alexander Fung *,1,2, Paris Jaggers *,3, Jason Chen 1, Joshua S. Rule1,5, Yael Benn 4,
Joshua B. Tenenbaum1, Steven T. Piantadosi5, Rosemary A. Varley3, Evelina Fedorenko1,2

1Department of Brain & Cognitive Sciences, Massachusetts Institute of Technology
2McGovern Institute for Brain Research, Massachusetts Institute of Technology
3Department of Psychology and Language Sciences, University College London
4Department of Psychology, Manchester Metropolitan University
5Department of Psychology, UC Berkeley

Acknowledgements

We would like to acknowledge the Athinoula A. Martinos Imaging Center at the McGovern
Institute for Brain Research at MIT, including the technical team —Steve Shannon and Atsushi
Takahashi. We would also like to thank Anya Ivanova, Colton Casto, Ben Lipkin, Chiebuka
Ohams, and Ted Gibson for helpful comments on the manuscript. HK was supported by a Friends
of McGovern graduate fellowship and a graduate fellowship from the K. Lisa Yang Integrative
Computational Neuroscience (ICoN) Center. JBT was supported by research funds from the Air
Force Office of Scientific Research, the Office of Naval Research Scien ce of AI program, and a
Schmidt Sciences AI2050 Fellowship. JSR and STP were supported by the Division of Research
on Learning in Formal and Informal Settings (EDU/DRL) grant 2201843. RV was supported by a
Leverhume Research Fellowship (RF-2023-690\10). EF was supported by research funds from the
McGovern Institute for Brain Research, the Department of Brain and Cognitive Sciences, MIT’s
Quest for Intelligence, and a grant from the Simons Foundation to the Simons Center for the Social
Brain at MIT. We would like to dedicate this work to S.A., an extraordinary individual whose
time, effort, and remarkable mind and brain have made invaluable contributions to neuroscience
over the past three decades, since becoming aphasic in his mid-forties. We are indebted to him and
his family.

Keywords
Language of Thought, Language, Logical Reasoning, Aphasia, fMRI, Cognitive Neuroscience

.CC-BY 4.0 International licenseavailable under a
was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprint (whichthis version posted July 29, 2025. ; https://doi.org/10.1101/2025.07.26.666979doi: bioRxiv preprint

Abstract

Humans are endowed with a powerful capacity for both inductive and deductive logical thought:
we easily form generalizations based on a few examples and draw conclusions from known
premises. Humans also arguably have the most sophisticated communication system in the animal
kingdom: natural language allows us to express complex and structured meanings. Some have
therefore argued for a tight relationship between complex thought and language, postulating that
reasoning, including logical reasoning, relies on linguistic representations. We systematically
investigated the relationship between logical reasoning and language using two complementary
approaches. First, we used non -invasive brain imaging (fMRI) to examine neural activity as
healthy adults engaged in inductive and deductive logical reasoning tasks. And second, we
behaviorally evaluated logical abilities in individuals with extensive lesions to the language brain
areas and consequent severe linguistic impairment. Our findings reveal that the language sy stem
is not engaged during logical reasoning, and patients with severe aphasia exhibit intact
performance on logic tasks. Instead, inductive reasoning recruits the domain -general multiple
demand system implicated broadly in goal-directed behaviors, whereas deductive reasoning draws
on brain regions that are distinct from both the language and the multiple demand systems.
Together, these results indicate that linguistic representations are neither utilized nor required for
inductive or deductive logical reasoning.

Significance

Which brain areas allow humans to reason logically, to understand whether a conclusion follows
from the premises? Are they the same areas that allow the assembly of words into structured
representations? Scholars have debated for millennia whether logical reasoning is inextricably tied
to natural language, or instead relies on a distinct “language of thought” (LOT). Using fMRI in
healthy adults and evaluating logical ability in individuals with severe aphasia, we find that distinct
neural systems support language processing vs. logical (inductive and deductive) reasoning. These
results establish that language does not underpin logical inference and point to distinct
representational systems for the logical LOT. This work contributes to our understanding of the
division of cognitive labor in the human brain.

Introduction

In 350 BC, Aristotle wrote that “writing and speech are not the same for all people, but mental acts
themselves, of which words signify, are the same for all people” (tr. Cooke & Tredennick, 1938).
This reflection highlights a crucial distinction between linguistic symbols and the underlying
cognitive structures they aim to express: although natural languages differ across cultures, the
capacity for abstract thought and logical reasoning appears to be ubiquitously present across
human societies. Still, because all natural languages share critical features some have argued that
natural language is the underlying medium of complex thought (e.g., Chomsky, 1965; Carruthers,
2002).

One prominent proposal—the “language of thought” (LOT) hypothesis (Fodor, 1975) —has
emphasized the compositional and hierarchical nature of thoughts. According to this hypothesis,
thoughts are composed of smaller atomic pieces in a structured fo rmat with hierarchical relations
among the component elements, similar to how computer programs are built out of a small
collection of primitive operations, or sentences are built out of words. Building on this similarity
with language, some have then expl icitly argued that the language of thought is natural language
(e.g., Davidson, 1967, 1975; Dennett, 1991, 1996, 2017; Chomsky, 1993, 1995; Carruthers, 2002;
for earlier claims, see Wittgenstein, 1921; for counter-arguments, see Pinker, 1994; Fodor, 1998).
Many thoughts can certainly be cast into natural language: after all, an effective communication
system must allow for the transmission of internal mental representations. However, people also
express many ideas using non-linguistic symbols, mathematical expressions, visual schematics or
diagrams, and so on. Therefore, a degree of isomorphism between certain thoughts and linguistic
expressions need not entail that specifically linguistic representations are used for thinking, any
more than the fact that language shares hierarchical structures with music entails that language is
the substrate for music.

Indeed, empirical data have been accumulating that suggest that linguistic representations are
neither utilized nor necessary for thinking. Patients with acquired linguistic impairments (aphasia)
appear to engage in diverse forms of thought, as evidenced by their intact performance on tasks
requiring mathematical reasoning (Varley et al., 2005), causal reasoning (Varley & Siegal, 2000),
and Theory of Mind (Ap perley et al., 2004, 2006). A lso, the brain areas that support language
comprehension and production (see Fedorenko et al., 2024 for a review) are not engaged during
many cognitively demanding tasks. These tasks include solving arithmetic and algebraic problems
(Fedorenko et al., 2011; Monti et al., 2012; Amalric et al., 2019), understanding computer code
(Ivanova et al., 2020; Liu et al., 2020), engaging in social reasoning (Paunov et al., 2019, 2022;
Shain, Paunov, Chen et al., 2023), or performing executive function tasks (Fedorenko et al., 2011;
Malik-Moraleda, Ayyash et al., 2022; Hiersche et al., 2024), which require maintenance of
information in working memory and inhibiting irrelevant distractors —skills intimately linked to
fluid intelligence (Engle et al., 1999; Duncan, 2010).

However, the relationship between language and logical reasoning specifically has only received
limited attention in the prior neuroscience literature, in spite of the fact that logic is a hallmark
domain where the role of language has been emphasized (Kroger et al., 2008). One prior fMRI
study has examined the relatio nship between language and deductive reasoning by comparing
neural responses during a logic inference task (judging the validity of a conclusion given a
.CC-BY 4.0 International licenseavailable under a
was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprint (whichthis version posted July 29, 2025. ; https://doi.org/10.1101/2025.07.26.666979doi: bioRxiv preprint

premise) versus a linguistic inference task (judging the validity of thematic roles across two
sentences) relative to lower-level control conditions (judging expression well-formedness) (Monti
et al., 2009). A partial dissociation was observed: although both the logic and the language contrast
elicited left-lateralized responses in the frontal, temporal, an d parietal cortex, some brain areas—
in the vicinity of areas traditionally implicated in language processing (Geschwind, 1970;
Goodglass, 1993; Friederici et al., 2002; Hagoort, 2019; Fedorenko et al., 2024) —showed a
stronger response to linguistic inference, whereas other areas responded more strongly during logic
inference (see Coetzee et al., 2022 for concordant evidence from TMS). However, recent work
calls these findings into question. Research over the last decade has a) developed more robust ways
of identifying language brain areas (Fedorenko et al., 2010; Lipkin et al., 2022), and b) established
that the kind of linguistic paradigm employed by Monti et al. (2009) plausibly engages both
language areas but also brain areas sensitive to general cognitiv e effort (e.g., Diachek, Blank,
Siegelman et al., 2020), which makes previously observed dissociations more difficult to interpret.
Another past study examined logical, as well as mathematical, reasoning and argued that neither
type of reasoning engaged the language areas (Kroger et al., 2008), but that study did not include
a language task, which is necessary to make such a claim. Thus, these prior findings are intriguing,
but have not conclusively answered the question of whether logical reasoning engages linguistic
processing mechanisms.

Two other bodies of work bear relevance to the logic-language relationship. The first comes from
developmental psychology and suggests that the timelines of linguistic and logic development
diverge. Inductive reasoning c apacities appear to come online very early (e.g., Gopnik, 1982;
Goddu & Gopnik, 2025). The representations are argued to be sensorimotor, not linguistic, in
nature, and the main debate concerns their domain specificity (e.g., Carey, 2009; Gelman, 2003)
vs. generality (e.g., Gopnik et al., 2004; Gopnik & Wellman, 2012) (e.g., do infants have distinct
representations for reasoning about the physical world vs. about social agents?). For deductive
reasoning, the picture is less clear: some have argued that one -year-old preverbal infants can
already reason disjunctively (Cesana -Arlotti et al., 2018), which implies a separation of logical
and linguistic ability. However, others have reported failures in simple disjunctive inferences as
late as age 2.5 years (Mody & Carey, 2016). These late failures may suggest that a certain level of
linguistic competence is required for the development of these capacities, but they are also
consistent with the slow developmental trajectory of executive abilities (Tervo -Clemmens et al.,
2023; Best & Miller, 2010; Diamond, 2013; Zelazo & Carlson, 2020), which may be needed to
deal with the task demands in more complex paradigms.

The second body of work comes from artificial intelligence, where recent advances have provided
additional motivation for investigating the relationship between language and logic. In particular,
neural network language models (Devlin et al., 2019; Radford et al., 2018; Brown et al., 2020) not
only achieve human-level performance on diverse language tasks (Hendrycks et al., 2021; Achiam
et al., 2023; Rein et al., 2024), but also exhibit impressive successes on certain reasoning tasks
(e.g., ARC: Chollet et al., 2024; BIG -Bench: Srivastava et al., 2022; LogiQA: Liu et al., 2020;
MATH: Hendrycks et al., 2021; WildBench: Lin et al., 2024; Templates: Boix-Adserà et al., 2024).
These findings beg the question of whether linguistic competence inevitably leads to logical
ability, although many have pointed out that reasoning in large language models lacks robustness
and generalizability (e.g., Mahowald, Ivanova, et al., 2024; Qiu et al., 2024; Shojaee, Mirzadeh et
al., 2025; McCoy et al., 2023; Nezhurina et al., 2025).
.CC-BY 4.0 International licenseavailable under a
was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprint (whichthis version posted July 29, 2025. ; https://doi.org/10.1101/2025.07.26.666979doi: bioRxiv preprint