Evidence from formal logical reasoning reveals that the language of thought is not natural language — part 2/10

Cleaned-text code points [13929, 26448). Bibliography omitted here; source.md includes it.

To systematically examine the role of language in human logical reasoning, we adopted a two -
pronged approach. We first used functional MRI in healthy adults to test whether the language 
network (Fedorenko et al., 2024) would be engaged during logical reasoning tasks (Study 1). Next, 
we used behavioral experiments in individuals with severe aphasia resulting from extensive 
damage to left peri -Sylvian cortex, to test whether linguistic ability is necessary for logical 
reasoning (Study 2). In each study, we ex amined two paradigmatic forms of logical reasoning: 
inductive and deductive reasoning. The inductive reasoning paradigm (Figure 1, left panel) was 
adapted from Rule et al. (2024; for earlier, related paradigms, see Bruner et al., 1956; Bongard, 
1967; Gentn er, 1983; Hofstadter & Mitchell, 1994; Tenenbaum, 1999). Participants were 
presented with an input number list and an output list (e.g., [5, 7] → [7, 5, 7] ) and asked to 
infer the rule that governs the input-to-output transformation. They could then test their hypothesis 
on a new input list, and so on, until they guess the correct rule. The rules involve a combination of 
mathematical operations (e.g., f (l) = [ x + 2 ∣ x ∈ l ]), “add 2 to every number”), list operations 
(e.g., f (l) = [ xi ∣ xi ∈ l, ∀j < i, xi ≠ xj ], “leave only unique, non-repeated elements”), and structural 
operations (e.g., f (l) = l[3:] ^ l[0:3], “rotate the list by three elements”), each of which can be 
written as short computer programs. The fMRI version also included a control co ndition where 
participants were told the rule and asked to apply it to new input lists. 
 
For deductive reasoning, we used three paradigms. The first was introduced in Coetzee & Monti 
(2018) and was used for the fMRI component of the study ( Figure 1, center panel). Participants 
were presented with a classic syllogism consisting of two premises and a conclusion (e.g., i. If A 
is B, then A is also C. ii. A is not C. iii. Therefore, A is not B) and were asked to judge the validity 
of the conclusion. Half of the  syllogisms used real words (e.g., "If the block is large..."), and the 
other half used nonwords (e.g., If the tep is ag…”). Although deductive reasoning was required 
for all problems, the problems varied in the difficulty of the necessary deduction. The e asier 
problems, known as Modus Ponens, had the following forms: “If the block is large, then it is not 
yellow. The block is large. So, the block is not yellow.” (the conclusion follows from the premises), 
or “If the block is large, then is it not yellow. T he block is large. So, the block is yellow.” (the 
conclusion does not follow from the premises). The more complex problems, known as Modus 
Tollens, had the following forms: “If the block is large, then it is not yellow. The block is yellow. 
So, the block i s not large.” (the conclusion follows from the premises), or “If the block is large 
then it is not yellow. The block is not yellow. So, the ball is large.” (the conclusion does not follow 
from the premises). This is a classic paradigm in the study of logic al reasoning (Wason, 1966;  
Johnson-Laird & Wason, 1970; Wason & Johnson-Laird, 1972; Johnson-Laird, 1975; Rips, 1983) 
and the contrast between the Modus Tollens versus Modus Ponens problems elegantly isolates 
deductive reasoning demands. However, it is not suitable for patients with aphasia because it uses 
verbal stimuli. As a result, the second, complementary paradigm relied on non -verbal matrix 
reasoning (Cattell, 1949; for prior use in fMRI, see Woolgar et al., 2013). Participants were 
presented with a matrix of four abstract geometric patterns and were asked to decide which of the 
four was an outlier ( Figure 1 , right panel). Patients with aphasia performed a similar version 
(Wechsler, 2011), except that a matrix of abstract geometric patterns was missing  a pattern, and 
participants were asked to decide which option from a set of possible answers completes the matrix 
(Figure 1 , bottom right panel). This type of paradigm, commonly used in assessing the fluid 
reasoning capacity across diverse populations (Cattell, 1940; Raven, 2000; Wechsler, 2008; Primi 
et al., 2010; Flanagan & Harrison, 2012; Weiss et al., 2016; Roth et al., 2015; Bediou et al., 2018, 
.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

2023; Zorowtiz et al., 202 4), primarily engages deductive reasoning by requiring participants to 
evaluate l ogical constraints and eliminate incorrect options. However, participants may 
additionally employ inductive strategies to generate hypotheses about the rules that govern the 
given patterns, so this paradigm may tax both deductive and inductive reasoning. 
 
If linguistic representations underlie logical inference, then a) the language brain areas should be 
engaged during the induction task, the deduction tasks, or both, as measured with fMRI; and b) 
patients with profound aphasia should be unable to perform t hese tasks. To foreshadow our 
findings, we do not find support for the idea that language is the medium of logical reasoning. 
 
 
Results 
 
The language network is not engaged during inductive nor deductive reasoning. 
 
As expected based on a large body of prior work (e.g., see Fedorenko et al., 2024 for a review), 
the areas of the language network show a robust response to language processing, evidenced by a 
stronger response during the sentence condition compared to the nonword-list condition estimated 
in independent data ( p< 0.001 at the network level; Figure 2B, Table 1). These areas have been 
established in past work to support computations related to retrieving word meanings from 
memory, syntactic -structure building, and semant ic composition during both language 
comprehension and production (Fedorenko et al., 2010; Menenti et al., 2011; Pallier et al., 2011; 
Giglio et al., 2022; Shain, Kean et al., 2024). Especially important for the hypothesis that the 
language system mediates logical reasoning is the sensitivity of these brain areas to hierarchical 
linguistic structure (Just et al., 1996; Ben -Shachar et al., 2003; Bornkessel et al., 2005; Caplan et 
al., 2008; Pallier et al., 2011; Tyler et al., 2011; Blank et al., 2016; Ding et  al., 2016; Heilbron et 
al., 2022; Shain et al., 2022). 
 
Critically, the language -processing brain areas show little or no response to the inductive and 
deductive reasoning contrasts ( Figure 2A -B, Table 1 ). The two deductive contrasts elicit no 
significant response (ps>0.1 at the network level; see SI Figure 1a and SI Table 1 for evidence 
that the results are similar for the five language areas separately), although as expected, the 
responses are strong to both conditions of the verbal deductive paradigm ( SI Figure 1b ). The 
inductive contrast elicits a reliable response at the network level ( p<0.05), but the effect is small, 
with the language contrast being over four times stronger, and the critical induction condition 
eliciting a response that is at or belo w the level of  the control condition of the language task 
(reading nonword lists) (SI Figure 1b). 
 
Patients with profound aphasia can nevertheless reason logically. 
 
Two individuals with profound aphasia, S.A. and G.S., showed preserved logical reasoning 
abilities. Both individuals have sustained extensive damage to their left peri -Sylvian cortex 
(Figure 2C) and, consequently, exhibit severe linguistic impairments (Figure 2D, SI Table 3). In 
particular, critical to the hypothesis that linguistic syntactic s tructures mediate logical reasoning, 
these patients show severe grammatical impairments in both comprehension and production, with 
at or near -chance performance on multiple syntactic assessments.  For example, both exhibited 
.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

marked impairment in assigning thematic roles in understanding reversible sentences, such as The 
diver splashed the dolphin . Nevertheless, despite their profound aphasia, both patients show 
typical-like performance on the inductive and the deductive reasoning tasks. In the inductive 
reasoning task (Figure 1, left panel), they solved the vast majority of the rules presented to them 
(19/25 and 39/40, respectively). This performance level is comparable to a normative sample of 
control participants; neither patient significantly differs from the control sample (Crawford -
Howell single-case test ps>0.499). Similarly, in the deductive matrix reasoning task, both patients 
solved the vast majority of the problems (25/30 and 26/30), which puts them +2.3 and +1.8 
standard deviations above the mean ba sed on the age -matched normative data available for this 
task (WASI-II; Wechsler, 2011) (these raw scores correspond to T-scores of 73 and 68; normative 
T = 50, SD = 10). 
Inductive and deductive reasoning elicit strong responses outside the language network. 
Given that our logic tasks did not engage the language system, we wanted to ensure that they elicit 
a response somewhere in the brain. We first examined neural responses in one plausible candidate 
system—the Multiple Demand (MD) network (Duncan & Owen, 2000; Duncan, 2010; Duncan et 
al., 2020). This network comprises a set of bilateral frontal and parietal brain areas that show strong 
activity during diverse cognitively demanding tasks, such as standard executive function tasks 
(Duncan & Owen, 2001;  Niendam et al., 2012; Hugdahl et al., 2015; Shashidhara et al., 2019; 
Assem et al., 2020a). Damage to the MD network is associated with decreases in fluid intelligence 
(Woolgar et al., 2010, 2018). In addition, and most relevantly, certain forms of reason ing appear 
to rely on the MD network, including mathematical reasoning (Monti et al., 2012; Fedorenko et 
al., 2013; Amalric et al., 2019) and the kind of reasoning necessary to understand computer code 
(Ivanova et al., 2020; Liu et al., 2020). Replicating much prior work (e.g., Fedorenko et al., 2013; 
Assem et al., 2020b), the areas of the MD network show a robust response to a spatial working 
memory task, including a stronger response during the more demanding condition (p<0.001 at the 
network level; SI Figure 2 and SI Table 2). The inductive reasoning task elicited a robust response 
across the MD network, evidenced by a stronger response during the rule induction condition 
compared to the control, rule application condition (p<0.001 at the network level; SI Figure 2, SI 
Table 2). However, the deductive reasoning task did not engage the MD network ( p>.48 at the 
network level). This finding aligns with Coetzee & Monti (2018), who reported a dissociation 
between brain areas sensitive to deductive reasoning and  those sensitive to general task 
difficulty—a key signature of the MD network (see also Kroger et al., 2008 for concordant data). 
A whole-brain search revealed several frontal and parietal areas that showed a stronger response 
to the Modus Tollens condition compared to the Modus Ponens condition estimated in independent 
data (p<0.001 across the set of areas; SI Figure 2). 
 
Discussion 
 
To shed light on the long -standing debate on the role of linguistic syntactic representations in 
formal logical reasoning, w e examined the responses of the language brain areas —robustly 
sensitive to syntactic structure ( Pallier et al., 2011; Giglio et al., 2022; Shain et al., 2022; Shain, 
Kean et al., 2024 )—to inductive and deductive reasoning. In a complementary approach, we 
examined logical reasoning abilities in individuals with severe damage to the language areas. The 
.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