Evidence from formal logical reasoning reveals that the language of thought is not natural language — part 3/10
Cleaned-text code points [26448, 39576). Bibliography omitted here; source.md includes it.
results converged on a clear answer: linguistic representations are not engaged nor necessary for logical reasoning. Below we contextualize these findings with respect to the broader literature. Accumulating evidence for language selectivity. Our results showing that the left -lateralized fronto-temporal language network is not engaged in logical induction nor deduction adds to a growing body of evidence for the selectivity of the language network for linguistic computations. In particular, prior studies have evaluated diverse hypotheses about overlap between linguistic processing and different perceptual and cognitive tasks. Some have focused on the social - communicative function of language and argued for overlap with other social functions (Grice, 1968, 1975; Sperber & Wilson, 1986), but studies have shown that non -verbal communicative signals, such as facial expressions and gestures, are processed in brain area s distinct from the language network (e.g., Deen et al., 2015; Pritchett et al., 2018; Jouravlev et al., 2019). Others emphasized the hierarchical structure of language and argued for overlap with the processing of other structured stimuli, such as music ( Patel, 2003; Fitch & Martins, 2014), but studies have shown that music perception tasks do not engage the language areas (Fedorenko et al., 2011; Rogalsky et al., 2011; Chen et al., 2023). Yet others have argued that language processing requires reliance on domain-general executive resources (Thompson -Schill, 2005; Kaan & Swaab, 2002; Novick et al., 2005, 2014; January et al., 2009), but executive tasks do not engage the language system (Fedorenko et al., 2011; Malik-Moraleda, Ayyash et al., 2022; Hiersche et al., 2024), and demanding linguistic computations are processed within the language network (Blank et al., 2016; Shain et al., 2020 , 2022; Quillen et al., 2021; Wehbe et al., 2021; see Fedorenko & Shain, 2021 for a review). Finally, and of greatest relevance to the current investigation, some hypotheses have highlighted the similarity between linguistic structure and structure in some domains of abstract reasoning, such as mathematical or logical thinking (Chomsky, 1957; Marcus, 2001; Carruthers, 2002), or even reasoning in particular domains, such as intuitive physics ( McCarthy & Hayes, 1969; Kowalski & Sergot, 1986; Pinto & Reiter, 1993 ) or social reasoning (de Villiers & de Villiers, 2000). A number of studies have examined the relationship between language processing and these types of reasoning and found no overlap: e.g., mathematical reasoning (Fedorenko et al., 2011; Amalric & Dehaene, 2016; Amalric et al., 2019; for evidence from aphasia, see Varley et al., 2005), computer code comprehension (Ivano va et al., 2020; Liu et al., 2020), physical reasoning (Kean et al., 2025), or social reasoning (Paunov et al., 2019, 2022; Shain, Paunov, Chen et al., 2022; Du et al., 202 4; for evidence from aphasia, see Varley & Siegal, 2000; Apperley et al., 2006; Willems et al., 2011). Our study adds to this body of work, showing that formal logical reasoning—including both induction and deduction —does not recruit nor require linguistic representations. Why do language processing and logical reasoning dissociate? Despite the fact that language can be used to express complex ideas, the representations and computations that support linguistic processing (i.e., retrieving words from memory and combining them into structured representations) appear to be distinct from those that mediate formal reasoning abilities, such as mathematical and logical reasoning. This dissociation presumably stems from the distinct demands associated with linguistic communication vs. formal reasoning. One key difference may have t o do with the kinds of meanings that language vs. these other systems typically express: namely, natural languages tend to express meanings related to the external and internal world, but mathematics, logic, and computer code express mostly abstract, relat ional meanings that do not .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 bear a direct or necessary connection to the external world (see Malik -Moraleda et al., 2025 for further discussion). Another reason may be that the linguistic format is actually not well -suited for formal reasoning, in spite of superficial similarities in the structured nature of both linguistic expressions and formal logical expressions. In particular, linguistic representations are noisy and ambiguous in ways that make them unreliable for supporting formal inference. Natural language is riddled with referential vagueness, scope ambiguities, and under -specification, requiring pragmatic enrichment, all of which can obscure the logical structure of a sentence. Evidence from AI research shows boosts in LLM performance on reasoning tasks when linguistic inputs are converted into first -order logic, which can then be fed into an external theorem solver (e.g., the LINC system; Olausson, Gu, Lipkin, Zhang, et al., 2024; the SatLM system; Ye et al., 2023; the LogicGuide system; Poesia et al., 2023; the Logic -LM system; Pan et al., 2023; and many others, e.g. Nye et al., 2021; Borazjanizadeh and Piantadosi, 2024). This work suggests that natural language is not a reliable medium for inference, which requires context-independent, structurally explicit representations. Alternatives to linguistic representations. A central outstanding question concerns the form of non-linguistic representations, such as those supporting logical induction and deduction, in the human mind and brain. Multiple com peting theoretical accounts exist, which make distinct predictions about the format and neural implementation of the representations involved in reasoning. For example, a ccording to the mental model (or mental simulation) framework, reasoners construct analog simulations of the problem, exploiting visuo-spatial representations and working memory resources to model the possible consequences of premises (Johnson-Laird et al., 1983; Knauff, 2013). These accounts predict the engagement of brain areas that suppo rt visual imagery and spatial working memory (e.g., Knauff et al., 2002; Alfred et al., 2020). Alternatively, script- or schema-based approaches propose that reasoning draws on cached relational templates (structured representations of events or situations abstracted from prior experience) (Piaget, 1923; Bartlett, 1932; Schank & Abelson, 1977; Bower et al., 1979). These accounts suggest that reasoning is scaffolded by context - and domain -specific knowledge structures, drawing on multiple semantic networks s haped by prior experience (e.g., Gilboa & Marlatte, 2017; Masís - Obando et al., 2022). And the mental logic proposals treat abstract reasoning as the manipulation of propositional (sometimes, language-like) symbols according to syntactic inference rules (Rips, 1988, 1995, 2003). These accounts posit that reasoning operates over amodal, abst ract symbols, potentially implemented in fron tal and parietal circuits (e.g., Monti et al., 2009; Reverberi et al., 2007). Perhaps the most promising candidate from this c lass of proposals is the probabilistic language of thought (PLOT) hypothesis (Goodman et al., 2014; Ellis et al., 2020; Rule et al., 2020), where mental algorithms are construed as symbolic programs over concepts, which encode probabilistic knowledge, including both knowledge of particular domains and abstract relational knowledge. In this way, PLOT provides a flexible medium for both domain-specific and abstract, domain-general reasoning, and captures the graded nature of mental representations. Given that these different accounts make distinct implementation -level predictions, future brain imaging studies may help determine whether logical reasoning depends on visuo -spatial, fully abstract, or domain-specific (at least during early stages of development) representations. .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 Modularity of reasoning. Aside from the dissociation between language processing and logical reasoning, this study also highlights the fact that different kinds of reasoning dissoc iate from one another. Some forms of reasoning have been shown to be domain -specific, including social reasoning (Saxe & Kanwisher, 2003; Gallagher & Frith, 2003; Schurz et al., 2014) and intuitive physical reasoning (Fischer et al., 2016; Schwettmann et a l., 2019; Pramod et al., 2022; 2025 : Mitko & Fischer, 2024). Both of these systems are distinct from the Default network, implicated in episodic self-projection and constructing situation models (Hassabis & Maguire, 2007; Spreng et al., 2009; Baldassano et al., 2018; Buckner & DiNicola, 2019), and the Multiple Demand (MD) network, implicated in diverse goal -directed behaviors (Duncan, 2010; Assem et al., 2020a). We here show that inductive reasoning, similar to mathematical reasoning (e.g., simple arithmeti c), draws on the MD network, but deductive reasoning does not engage the MD network, instead recruiting a distinct set of frontal and parietal areas, in line with past work (Kroger et al., 2008; Coetzee & Monti, 2018). How these deductive -reasoning areas r elate to other known systems supporting complex cognition, why these different forms of reasoning dissociate, and whether there exist other specialized reasoning systems remains to be discovered. Limitations and open questions. The current study is limited in several ways . (1) The study is limited to native English speakers. Although we are not aware of hypotheses about cross-linguistic differences in whether language and logical reasoning overlap, it remains important to generalize these findings to other linguistic populations (Blasi et al., 2022; Malik-Moraleda et al., 2022). (2) The aphasia component is limited to two participants. This small number was dictated by the requirement of a profound linguistic impairment, which is rare. Most individuals with aphasia retain or recover linguistic abilities to some degree (Wilson et al., 2022). Evidence of intact logical abilities in such patients would be of limited utility because it would always be possible that they are relying on the remaining portions of t he language network to perform the reasoning tasks. Evidence of intact cogniti on in severely linguistically impaired individuals is precious and important, even if it comes from a limited number of participants. After all, at its inception, the field of co gnitive neuroscience gained some of its greatest insights from case studies (Harlow, 1848; 1869; Broca, 1861; Scoville & Milner, 1957). And (3): It would be useful to extend these findings to a broader array of logical reasoning tasks, including contextual ized paradigms (Cosmides, 1989; Cox & Griggs, 1982; Griggs & Cox, 1982; Gigerenzer & Hug, 1992) , and to develop non-verbal versions of purely deductive tasks for patients with aphasia. Our study also leaves a number of research questions open. For example: Is the language -logic dissociation already present during early childhood, or does it emerge over development? And is the dissociation between linguistic processing and logical reasoning, or among different types of reasoning, an inevitability given the computational demands of these different cognitive tasks, or is it merely an accident of biological evolution —a question that can now be tested in large reasoning models (Yang et al., 2019; Csordás et al., 2021; Lepori et al., 2023a, 2023b; AlKhamissi et al., 2025). Conclusions. Our results provide key evidence against the hypothesis that natural language serves as the medium of abstract reasoning and suggest that logical reasoning is underpinned by a distinct representational format. .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 .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