Highlights
-
• The LIFG has a causal role in domain-general cognitive control, not language control.
-
• The LMTG specializes in language processing, not language or cognitive control.
-
• Causal evidence shows distinct neural mechanisms for language and cognitive control.
1. Introduction
Bilingual speakers seem to switch effortlessly between languages. Research has shown that both languages are activated simultaneously during speech production, resulting in cross-language competition (e.g., Abutalebi & Green, Reference Abutalebi and Green2007; Costa & Caramazza, Reference Costa and Caramazza1999; Green, Reference Green1998; Kroll et al., Reference Kroll, Bobb and Wodniecka2006; Levelt, Reference Levelt2001). Consequently, bilinguals employ mechanisms to manage this competition in response to contextual demands, which are referred to as bilingual language control (hereafter referred to as language control; Crinion et al., Reference Crinion, Turner, Grogan, Hanakawa, Noppeney, Devlin and Price2006; Green, Reference Green1998). Specifically, Green’s inhibitory control model (ICM) (1998) proposed that bilingual speakers produce the target language by inhibiting the interference from the nontarget language. A central and unresolved question in bilingualism is the relationship between language control and domain-general cognitive control (hereafter referred to as cognitive control), which involves managing thoughts and actions according to goals and is typically evaluated through tasks like the Simon task and nonverbal switching tasks (e.g., Braver, Reference Braver2012; Miller, Reference Miller2000; Wu et al., Reference Wu, Yang, Chen, Li, Zhang, Kang and Guo2019). To address this question, the present study examined bilinguals’ performance in language and nonverbal switching tasks after applying transcranial magnetic stimulation (TMS) to two functionally distinct brain areas: one potentially supporting both language control and cognitive control and another specific to language processing and possibly language control, namely, the left inferior frontal gyrus (LIFG) and the left middle temporal gyrus (LMTG).
The language switching paradigm has been commonly used to verify the ICM (e.g., Meuter & Allport, Reference Meuter and Allport1999). In this paradigm, participants named pictures or digits in either their first language (L1) or second language (L2) within the dual-language context, based on cues indicating the target language for each trial. These cues led to two trial types: switch trials, which required a different language from the previous trial (i.e., L1 → L2, L2 → L1), and repeat trials, which used the same language (i.e., L1 → L1, L2 → L2). Research has shown significantly longer reaction times (RTs) and higher error rates (ERs) on switch trials compared to repeat trials (Christoffels et al., Reference Christoffels, Firk and Schiller2007; Jylkkä et al., Reference Jylkkä, Lehtonen, Lindholm, Kuusakoski and Laine2018, Reference Jylkkä, Soveri, Laine and Lehtonen2020; Meuter and Allport, Reference Meuter and Allport1999; Timmer et al., Reference Timmer, Calabria, Branzi, Baus and Costa2018) – a phenomenon known as the switching cost (i.e., switching cost = switch – repeat). In addition, researchers have included single-language contexts, in which bilinguals name items exclusively in either L1 or L2 throughout an entire block (e.g., Calabria et al., Reference Calabria, Costa, Green and Abutalebi2018; Christoffels et al., Reference Christoffels, Firk and Schiller2007; de Bruin & Xu, Reference de Bruin and Xu2023; Ma et al., Reference Ma, Li and Guo2016; Wang et al., Reference Wang, Wu, Ji, Yan and Wu2022). Repeat trials in the dual-language context typically elicit longer RTs and higher ERs than trials in the single-language context – a phenomenon known as the mixing cost (i.e., mixing costs = repeat – single). Notably, switching and mixing costs can be situated within the dual mechanisms of control framework (Braver Reference Braver2012), which posits that cognitive control operates either proactively by anticipating demands or, reactively, by responding to interference as it arises. This perspective is further supported by empirical findings from language switching research (Ma et al., Reference Ma, Li and Guo2016) and has been widely adopted in subsequent studies (e.g., Snijders et al., Reference Snijders, Van Witteloostuijn, Boerma, Timmermeister and Blom2025; Jiao, Timmer et al., Reference Jiao, Timmer, Liu and Chen2022; Timmer, Christoffels et al., Reference Timmer, Christoffels and Costa2019; Wu et al., Reference Wu, Ji, Qu, Zuo, Liang, Su and Ding2025; Yang et al., Reference Yang, Cai, Lin and Wang2024). Specifically, switching costs are interpreted as reflecting transient, stimulus-driven processes that resolve interference after it arises, operating on a local, trial-by-trial basis (i.e., reactive control or local control). In contrast, mixing costs reflect sustained, anticipatory processes that prevent interference before it occurs, operating on a global, block-level basis (i.e., proactive control or global control).
The literature presents two broad perspectives on the relationship between bilingual language control and domain-general cognitive control. Overlap accounts – such as the inhibitory control model (Green, Reference Green1998) and the adaptive control hypothesis (Green & Abutalebi, Reference Green and Abutalebi2013) – propose that both domains recruit a shared network of frontal–parietal and subcortical regions. Dissociation accounts (e.g., Calabria et al., Reference Calabria, Hernández, Branzi and Costa2012, Reference Calabria, Suades, Juncadella, Ortiz-Gil, Ugas, Sala and Lleo2025; Branzi et al., Reference Branzi, Calabria, Boscarino and Costa2016), by contrast, posit at least partially distinct resources, supported by evidence of weak cross-domain correlations and neuroimaging findings of domain-specific activation patterns (e.g., Blanco-Elorrieta & Pylkkänen, Reference Blanco-Elorrieta and Pylkkänen2016). Clarifying whether these mechanisms overlap or dissociate is critical for refining theoretical models of bilingual control and for informing the design of targeted assessment and training protocols. Against this backdrop, the following paragraphs review behavioral, neuroimaging and neuromodulation evidence from both perspectives to situate the present study within this debate.
Several behavioral studies have reported significant correlations between bilinguals’ performance in language switching tasks and nonverbal switching tasks, suggesting overlapping mechanisms underlying these two processes (e.g., Declerck et al., Reference Declerck, Grainger, Koch and Philipp2017; Liu et al., Reference Liu, Liang, Dunlap, Fan and Chen2016; Prior & Gollan, Reference Prior and Gollan2011; Timmer et al., Reference Timmer, Calabria, Branzi, Baus and Costa2018; Timmer, Calabria et al., Reference Timmer, Calabria and Costa2019; Verreyt et al., Reference Verreyt, Woumans, Vandelanotte, Szmalec and Duyck2016; Woumans et al., Reference Woumans, Ceuleers, Van der Linden, Szmalec and Duyck2015). For instance, Timmer et al. (Reference Timmer, Calabria, Branzi, Baus and Costa2018) and Timmer, Calabria et al. (Reference Timmer, Calabria and Costa2019) identified a significant positive correlation between switching costs across the two tasks in trilinguals and further demonstrated that bilinguals trained in a language-switching context exhibited reduced nonverbal switching costs compared to those trained in a single-language context. These studies suggest the existence of shared mechanisms underlying language control and cognitive control. However, this view has been challenged, as other studies have reported conflicting results (e.g., Branzi et al., Reference Branzi, Calabria, Boscarino and Costa2016; Calabria et al., Reference Calabria, Hernández, Branzi and Costa2012, Reference Calabria, Branzi, Marne, Hernández and Costa2015; Cattaneo et al., Reference Cattaneo, Calabria, Marne, Gironell, Abutalebi and Costa2015). For instance, Calabria et al. (Reference Calabria, Hernández, Branzi and Costa2012) found no significant correlation between language and nonverbal switching performance in 28 Catalan-Spanish bilinguals, and this result was later replicated with a larger sample of 60 bilinguals, revealing age effects on nonverbal but not language switching costs (Calabria et al., Reference Calabria, Branzi, Marne, Hernández and Costa2015). Thus, the current body of behavioral evidence presents an inconsistent and controversial view on the extent of overlap between language control and cognitive control.
Yet, brain imaging techniques can and have enhanced our understanding of the neural mechanisms underlying language control and cognitive control. For example, Abutalebi and Green (Reference Abutalebi and Green2008, Reference Abutalebi and Green2016) proposed a neural model of language control based on a comprehensive review of neuroimaging studies on language control, positing that language control requires the activation of a neural network involving the prefrontal cortex, caudate nucleus, anterior cingulate cortex and supramarginal gyrus (for a more intricate overview, see Abutalebi & Green, Reference Abutalebi and Green2016). Notably, the recruitment of areas such as the anterior cingulate cortex and the prefrontal cortex is not exclusive to language control, as they also play a role in cognitive control (Aarts et al., Reference Aarts, Roelofs and Van Turennout2008; Cools, Reference Cools2011; Frank, Reference Frank2011; Kerns et al., Reference Kerns, Cohen, MacDonald, Cho, Stenger and Carter2004). Evidence supports that language control and cognitive control share overlapping neural mechanisms across widely distributed brain regions (e.g., Anderson et al., Reference Anderson, Chung-Fat-Yim, Bellana, Luk and Bialystok2018; De Baene et al., Reference De Baene, Duyck, Brass and Carreiras2015; Weissberger et al., Reference Weissberger, Gollan, Bondi, Clark and Wierenga2015; Wu et al., Reference Wu, Yang, Chen, Li, Zhang, Kang and Guo2019; Jiao, Meng et al., Reference Jiao, Meng, Wang, Schwieter and Liu2022). For instance, Jiao, Meng et al. (Reference Jiao, Meng, Wang, Schwieter and Liu2022) conducted a meta-analysis of neuroimaging studies, identifying a prominent shared activation pattern in the left dorsolateral prefrontal cortex, pre-supplementary motor area/dorsal anterior cingulate cortex and left precuneus during both language switching tasks and nonverbal switching tasks. However, some studies challenge this view, suggesting that the mechanisms involved in both types of switching tasks may be primarily distinct (Blanco-Elorrieta & Pylkkänen, Reference Blanco-Elorrieta and Pylkkänen2016; Magezi et al., Reference Magezi, Khateb, Mouthon, Spierer and Annoni2012). For example, Blanco-Elorrieta and Pylkkänen (Reference Blanco-Elorrieta and Pylkkänen2016) found that language control primarily activates the anterior cingulate cortex, while cognitive control involves the left dorsolateral prefrontal cortex. As neuroimaging techniques are inherently correlational rather than indicative of causality (e.g., Pestalozzi et al., Reference Pestalozzi, Annoni, Müri and Jost2020; Wu, Ji et al., Reference Wu, Ji, Cai, Pu, Wang, Yan and Wang2024; Wu et al., Reference Wu, Ji, Qu, Zuo, Liang, Su and Ding2025), brain activity triggered by stimuli does not guarantee that all activated regions are responsible for the task; as a result, some studies may report epiphenomenally activated regions that are absent in others. For instance, the LIFG has been shown to activate during both language switching and nonverbal switching tasks in functional magnetic resonance imaging (fMRI) studies (e.g., De Baene et al., Reference De Baene, Duyck, Brass and Carreiras2015; Jiao, Meng et al., Reference Jiao, Meng, Wang, Schwieter and Liu2022). However, the causal mechanisms underlying the LIFG in these processes remain inadequately explored.
To complement brain imaging techniques, noninvasive brain stimulation techniques, such as TMS and transcranial direct current stimulation (tDCS), can modulate cortical activities and influence behaviors linked to specific brain regions, providing causal insights into brain functions (Siddiqi et al., Reference Siddiqi, Kording, Parvizi and Fox2022; Silvanto & Pascual-Leone, Reference Silvanto and Pascual-Leone2012). These techniques provide a valuable opportunity to investigate the relationship between language control and cognitive control from a causal perspective. For instance, although the right inferior frontal gyrus (RIFG) is typically associated with cognitive control (e.g., Aron et al., Reference Aron, Robbins and Poldrack2004; Jahfari et al., Reference Jahfari, Waldorp, van den Wildenberg, Scholte, Ridderinkhof and Forstmann2011), Wu et al. (Reference Wu, Ji, Qu, Zuo, Liang, Su and Ding2025) applied offline TMS to this region to examine its causal role in language control. They found that stimulating the RIFG facilitated performance on language switching tasks, indicating that this region plays an important role in suppressing the nontarget language during language switching. However, the absence of a nonverbal switching task makes it unclear whether the effects of neural modulation in language control are comparable to those in cognitive control. Recently, Vaughn et al. (Reference Vaughn, Watlington, Linares Abrego, Tamber-Rosenau and Hernandez2021) instructed language switching tasks and nonverbal switching tasks after applying tDCS to the bilateral dorsolateral prefrontal cortex in bilinguals. Their findings revealed that stimulation of the left dorsolateral prefrontal cortex selectively impaired nonverbal switching, whereas stimulation of the right dorsolateral prefrontal cortex did not influence switching costs in either task. These results suggest that the left dorsolateral prefrontal cortex primarily facilitates cognitive control rather than language control. However, the limited precision of tDCS raises concerns, as it may inadvertently stimulate a broader area of the prefrontal cortex, including the middle frontal gyrus and the inferior frontal gyrus, rather than the intended target.
Specifically, the LIFG, a subregion of the prefrontal cortex, is regarded as a key neural substrate at the intersection of language processing and cognitive control, warranting further investigation into its dual role in both domains. Notably, the LIFG has been shown to activate during both language and nonverbal switching tasks in fMRI studies (e.g., De Baene et al., Reference De Baene, Duyck, Brass and Carreiras2015; Jiao, Meng et al., Reference Jiao, Meng, Wang, Schwieter and Liu2022). Consequently, we propose that the significant influence observed in Vaughn et al. (Reference Vaughn, Watlington, Linares Abrego, Tamber-Rosenau and Hernandez2021) may be attributed to the LIFG. Furthermore, the correlational nature of fMRI data complicates the determination of whether activation in the LIFG is merely epiphenomenal or causally involved in both language and nonverbal switching tasks. To our knowledge, no studies have yet employed TMS to investigate the role of the LIFG in language control and cognitive control, despite TMS providing superior spatial precision in stimulation compared to tDCS (Keeser et al., Reference Keeser, Meindl, Bor, Palm, Pogarell, Mulert and Padberg2011; Klaus & Schutter, Reference Klaus and Schutter2018; Zhou et al., Reference Zhou, Qiu, He and Zhang2023). Therefore, the primary objective of the present study is to investigate whether the LIFG functions in the same way in both language control and cognitive control.
In addition, the left temporal lobe is well established as a core region for language processing (e.g., neuromodulation studies; Binder et al., Reference Binder, Frost, Hammeke, Rao and Cox1996; Choi et al., Reference Choi, Park and Paik2015; Friederici & Kotz, Reference Friederici and Kotz2003; Meyer et al., Reference Meyer, Mecklinger, Grunwald, Fell, Elger and Friederici2005; Powell et al., Reference Powell, Parker, Alexander, Symms, Boulby, Wheeler-Kingshott and Duncan2007), and emerging evidence suggests that such processing is not only governed by language control mechanisms but may also exert influence on them in return (Wu, Ji et al., Reference Wu, Ji, Cai, Pu, Wang, Yan and Wang2024; Wu, Zhao et al., Reference Wu, Zhao, Wu, Liu, Su, Ji and Wang2025). For instance, TMS to the left superior temporal gyrus (LSTG) – a region critical for phonological processing and speech production (Hickok et al., Reference Hickok, Erhard, Kassubek, Helms-Tillery, Naeve-Velguth, Strupp and Ugurbil2000; Vigneau et al., Reference Vigneau, Beaucousin, Hervé, Duffau, Crivello, Houde and Tzourio-Mazoyer2006) – significantly reduced language switching costs (Wu, Zhao et al., Reference Wu, Zhao, Wu, Liu, Su, Ji and Wang2025), a key index of language control. Compared to the LSTG, the LMTG is also involved in language processing, but is more specifically associated with lexical–semantic processing (e.g., Ashtari et al., Reference Ashtari, Lencz, Zuffante, Bilder, Clarke, Diamond, Kane and Szeszko2004; Cui et al., Reference Cui, Liu, Song, Lipnicki, Li, Xie, Chen, Li, Lu, Lv, Wang, Yan, Yan, Zhang, Zhang and Jiang2018; Morese et al., Reference Morese, Brasso, Stanziano, Parola, Valentini, Bosco and Rocca2022). However, it remains unclear whether the LMTG contributes to language control through its role in lexical–semantic processing, as the LSTG does through phonological processing. Addressing this question is theoretically significant, as it can help determine whether the interface between language processing and language control is confined to phonological processing or extends to lexical–semantic processing as well. Therefore, a secondary aim of the present study is to examine whether modulating a region involved in lexical–semantic processing (namely, the LMTG) affects language control, while cognitive control – where no effect is expected – serves as a comparison.
To address these research objectives, the present study employed navigated TMS to transiently modulate the neural activity in the LIFG and LMTG of bilingual participants, examining their subsequent performance on language switching tasks and nonverbal switching tasks. Specifically, the TMS protocol employed in the present study was continuous theta burst stimulation (cTBS), typically intended to reduce neural excitability and produce inhibitory effects (e.g., Huang et al., Reference Huang, Edwards, Rounis, Bhatia and Rothwell2005; but see Hamada et al., Reference Hamada, Murase, Hasan, Balaratnam and Rothwell2013 for a more intricate overview). The vertex served as a control stimulation site, allowing us to evaluate the specific contributions of both regions. We hypothesize that language control and cognitive control share the same mechanisms in the LIFG. If this hypothesis holds true, we predict similar patterns of TMS-induced worse performance, as indexed by switching and/or mixing costs for both language switching and nonverbal switching tasks. Alternatively, we anticipate divergent patterns of TMS-induced changes between the two tasks. Furthermore, we hypothesize that the language processing in the LMTG exclusively affects language control. If this holds, we anticipate worse performance, as indexed by increased switching and/or mixing costs, following LMTG stimulation in the language switching task only, with no effect in the nonverbal task. Alternatively, these costs may remain intact in both tasks, or similar patterns of TMS-induced changes may be observed in both.
2. Methods
2.1. Participants
In the current study, we recruited 35 unbalanced Chinese-English bilinguals, all of whom were native speakers of Chinese. Two participants withdrew from the experiment midway due to personal reasons. Therefore, the remaining 33 participants (19 females; age: 20.83 ± 1.13 [mean ± standard deviation] years, ranging from 19 to 24) were included in the analyses. All participants were right-handed, with normal or corrected-to-normal vision. None had any psychological disorders (e.g., claustrophobia) or neurological diseases (e.g., epilepsy) that would violate magnetic resonance imaging (MRI) and TMS guidelines (Rossi et al., Reference Rossi, Hallett, Rossini and Pascual-Leone2009, Reference Rossi, Antal, Bestmann, Bikson, Brewer, Brockmöller and Hallett2021). A language history questionnaire (LHQ) was used to assess the linguistic background of participants (Li et al., Reference Li, Sepanski and Zhao2006). Participants’ age of acquisition (AOA) for English was 7.67 ± 3.42 years. All participants rated their proficiency of L1 and L2 (10-point scale: 1 = “very unskilled”; 10 = “very skilled”). Participants averagely scored 8.35 ± 1.41 and 5.28 ± 1.64 for L1 and L2 proficiency, respectively. A paired t-test showed a significantly higher proficiency for Chinese than that for English (t = 10.01, p < 0.001), indicating that unbalanced bilinguals were recruited in the present study. All participants were fully briefed on the experimental requirements and signed informed consent forms before the experiment. This study met the guidelines of the Declaration of Helsinki and was approved by the Tianjin Normal University Ethics Committee.
2.2. Materials and design
Each participant performed both language switching and nonverbal switching tasks following a TMS session. Both tasks consist of four single-task blocks and one mixed block, presented in the specified order. In the language switching task, participants named pictures exclusively in L1 or L2 during single-task blocks, and in either L1 or L2 based on unpredictable cues in the mixed block. This design enables the extraction of switching and mixing costs associated with language control. In the nonverbal switching task, participants judge solely the shape or the color of the stimuli in single blocks, and either the shape or the color based on unpredictable cues in the mixed block. This framework facilitates the assessment of switching and mixing costs related to cognitive control. Further details can be found in Section 2.3.
In the language switching task, 24 standardized black and white line drawings were selected as stimuli from Snodgrass and Vanderwart (Reference Snodgrass and Vanderwart1980), with 3 extra images designated for the practice phase or fillers. According to the norms from Snodgrass and Vanderwart (Reference Snodgrass and Vanderwart1980), the familiarity ratings for the English names of all images were 3.48 ± 0.95, image agreement was 3.63 ± 0.56 and visual complexity was 2.91 ± 0.89. For the Chinese names, according to the database (Zhang & Yang, Reference Zhang and Yang2003), the familiarity ratings were 4.56 ± 0.39, image agreement was 3.60 ± 0.49, and visual complexity was 2.43 ± 0.71. The familiarity, visual complexity and image agreement were rated using a 5-point scale (1 for very unfamiliar/simple or low agreement and 5 for very familiar/complex or high agreement). Given that previous studies have employed face cues in language switching tasks and still observed robust switching costs (e.g., Zhu et al., Reference Zhu, Seymour, Szakay and Sowman2020, Reference Zhu, Blanco-Elorrieta, Sun, Szakay and Sowman2022), and that ecologically natural face cues can facilitate naming when the sociocultural identity of the face aligns with the language used for naming (e.g., Blanco-Elorrieta & Pylkkänen, Reference Blanco-Elorrieta and Pylkkänen2017; Li et al., Reference Li, Yang, Scherf and Li2013), we chose to use four different face images as cues in the language switching task: a male Asian face, a female Asian face, a male Caucasian face and a female Caucasian face (Figure 1A). These cue images were artificially generated from https://www.seeprettyface.com/index.html. Participants were instructed to name pictures in L1 given Asian faces and in L2 given Caucasian faces. Face cues in the language switching task were presented at a size of 3.9 × 2.9 cm, while line-drawing stimuli were 14 × 7.8 cm.
Figure 1. Cue images. (A) Cue images of language switching task: Displayed from left to right are the images of a male Asian face, a female Asian face, a male Caucasian face and a female Caucasian face. (B) Cue images of nonverbal switching task: a color gradient bar (left) and a row of small black shapes (right).
In the nonverbal switching task, both stimulus images and cue images were included. Specifically, the stimulus images consisted of pairs of shapes (i.e., a triangle and a circle) and colors (i.e., red and green), resulting in four images: a red triangle, a green triangle, a red circle and a green circle. The two cue images included one depicting a color gradient bar and the other containing a row of small black shapes (Figure 1B). The sizes are 9.76 × 3.90 cm for the cues and 11.71 × 7.81 cm for the stimuli in the nonverbal switching task. Different shapes indicate that participants should respond to the shape of the stimulus, whereas different colors indicate a response based on color. Participants report the shape or color of the stimuli by pressing the left key or the right key.
2.3. Procedures
As shown in Figure 2, this study comprises three TMS sessions to investigate the roles of the LIFG and the LMTG in bilingual control and cognitive control. The three sessions were conducted approximately one week apart (Figure 2A), with the order counterbalanced across participants by Latin square design. These sessions targeted the LIFG, the LMTG and the vertex (served as baseline), respectively. Each participant completed a language switching task and a nonverbal switching task following each cTBS intervention (Figure 2B). Both tasks consist of four single-task naming blocks and one mixed block, presented in the specified order (Figure 2C). At the end of the third session, participants filled out a modified version of the language history questionnaire (Li et al., Reference Li, Sepanski and Zhao2006).
Figure 2. Main experimental procedure. (A) Experimental flow: T1 image acquisition is conducted, followed by the cTBS experiment. Each stimulation session is spaced about a week apart, and the order of TMS sessions is counterbalanced among participants. (B) cTBS experimental procedure: The cTBS intervention is performed first, followed by a language switching task (D) and a nonverbal switching task (E). (C) Within-task block order: Both tasks consist of four single-task blocks and one mixed block, presented in the specified order. (D) One example of a trial sequence for the language switching task: Participants name the pictures either in L1 or L2 based on the cues. Here, a Caucasian face indicates naming in L2, while an Asian face indicates naming in L1. (E) One example of a trial sequence for the nonverbal switching task: Participants judge the shape or color of the stimuli based on the cues by pressing the left key or the right key. Here, different shapes indicate that participants should respond to the shape of the stimulus, whereas different colors indicate a response based on color.
Note: cTBS, continuous theta burst stimulation; LIFG, the left inferior frontal gyrus; LMTG, the left middle temporal gyrus; vertex, the baseline.
Prior to testing, participants were instructed to familiarize themselves with the task instructions, the pictures and their names in both Chinese and English (See details in Wang et al., Reference Wang, Wu, Ji, Yan and Wu2022). Subsequently, they underwent navigated TMS (See Section 2.4 for the TMS procedure). Participants then completed brief practice blocks and proceeded to complete the formal tasks including the language switching task and nonverbal switching task, with the order of the tasks counterbalanced across participants.
In the language switching task, participants were instructed to name the stimulus pictures as quickly and accurately as possible based on the cues. That is, they named the stimulus picture in L1 when given an Asian face and in L2 when given a Caucasian face. The task consisted of five blocks (Figure 2C) that alternately presented different language contexts (e.g., Da Baene et al., Reference De Baene, Duyck, Brass and Carreiras2015; Wu, Ji et al., Reference Wu, Ji, Cai, Pu, Wang, Yan and Wang2024), with the third block being a mixed block (i.e., a dual-language context) that included 192 trials: 48 trials for each of the four conditions (i.e., L1 repeat, L1 switch, L2 repeat, L2 switch). The other four blocks were single blocks (i.e., a single-language context), each consisting of 24 trials. This design enables continuous monitoring of changes in language control as participants transition between single-language and mixed-language contexts, while minimizing potential experimental confounds such as order effects, practice effects and fatigue associated with the mixed block, and is consistent with established paradigms in bilingual research (e.g., Casado et al., Reference Casado, Szewczyk, Wolna and Wodniecka2022; Da Baene et al., Reference De Baene, Duyck, Brass and Carreiras2015; Wang et al., Reference Wang, Wu, Ji, Yan and Wu2022). In addition, in the mixed block, the switch and repeat trials were pseudo-randomized to prevent the appearance of identical pictures in consecutive trials and to ensure that no more than 4 consecutive trials were the same trial types (i.e., switch and repeat). In the single blocks, participants named the stimulus pictures only in their L1 or L2 in each block, with the order of the languages reversed before and after the dual-language block and counterbalanced across participants. Each trial in both single-task and mixed blocks began with a fixation cross in the center of the screen for 300 ms (Figure 2D), followed by a blank screen for 200 ms, after which a picture stimulus was displayed for 1000 ms. The cue appeared simultaneously with the stimulus picture and was positioned above it. Finally, a blank screen was presented for 1200 ms before the next trial.
The nonverbal switching task was identical to the language switching task in stimulus presentation and trial designs. In this task, participants were required to make quick and accurate judgments about the shapes or colors of the stimuli based on cues by pressing the left or right key. Participants responded to the shape of the stimulus picture when the cue was a row of small black shapes, but to the color of the same picture when the cue was a color gradient. The key-pressing requirements for the left and right keys were balanced among participants. On each trial, the cue image and the stimulus picture appeared simultaneously, with the cue positioned above the stimulus. Similar to the language switching tasks, the nonverbal switching task consisted of five blocks (Figure 2C). The third block was a mixed block containing 192 trials: 48 trials for each of the four conditions (i.e., shape repeat, shape switch, color repeat, color switch). The other four blocks were single-task blocks, each consisting of 24 trials. In the mixed block, participants responded to the color or shape of the stimulus pictures according to unpredictable cues. Again, the switch and repeat trials were pseudo-randomized to ensure that no more than four consecutive trials were of the same trial type (i.e., switch and repeat). In the single-task blocks, participants responded to only the shape or color in each block according to the cue. As shown in Figure 2E, each trial began with a fixation cross for 300 ms, followed by a blank screen for 200 ms, after which a picture stimulus was displayed for 1000 ms. The cue was presented simultaneously with the stimulus picture and positioned above it. Finally, a blank screen was presented for up to 1200 ms, disappearing immediately once participants pressed the keys.
The procedures for the second and third sessions remained identical to those of the first session. In all behavioral tasks, the presentation of all trials was programmed, and responses were recorded using PsychoPy version 2021.2.3 (Peirce et al., Reference Peirce, Gray, Simpson, MacAskill, Höchenberger, Sogo and Lindeløv2019).
2.4. Transcranial magnetic stimulation
We employed cTBS to target the LIFG and LMTG in the current study. The Montreal Neurological Institute (MNI) coordinates of stimulation were as follows: LIFG (x = −52, y = 12, z = 16) and LMTG (x = −62, y = −48, z = 6) (Zaccarella et al., Reference Zaccarella, Schell and Friederici2017). Additionally, cTBS was administered to the vertex (x = 0, y = 0, z = 75) to establish a baseline (Jung et al., Reference Jung, Bungert, Bowtell and Jackson2016).
To facilitate TMS navigation guided by imaging, each participant underwent high-resolution T1-weighted anatomical MRI scans at the MRI Center of Tianjin Normal University, utilizing a Siemens Prisma 3-T MRI Scanner. The images were acquired for co-registration with the following parameters: repetition time (TR) = 2530 ms, echo time (TE) = 2.98 ms, flip angle = 7°, field of view = 256 × 256 mm, matrix size = 256 × 256, in-slice resolution = 1.0 × 1.0 mm, slice thickness = 1.0 mm and voxel size = 1 × 1 × 1 mm3.
A frameless stereotaxic localization system (Localite GmbH, Bonn, Germany) was employed to establish a real-time navigation monitoring system. The MRI T1-weighted images of each participant were loaded into the navigation system and manually registered by identifying key anatomical landmarks, including the anterior commissures, posterior commissures and a specific point on the falx. This process ensured precise stimulation of the target region. The participant-specific target region was defined using trajectory markers based on the Montreal Neurological Institute (MNI) coordinate system. Before each TMS session, an MRI co-registration procedure was performed to align the 3D model with the participant’s head in real space. Participants wore a headband equipped with reflective spherical markers, which were tracked by the navigation system to facilitate accurate placement of the coil over the designated target location.
In this study, a TMS stimulator (MagPro X100, MagVenture) equipped with a standard 70-mm-figure-eight coil (MagVenture MCF-B65) was utilized to generate pulsed stimulation. Biphasic pulses were used with the default TMS stimulator settings to generate an anterior-to-posterior current in the brain during the first phase and a posterior-to-anterior current during the second phase, which is opposite to the direction of current flow in the coil. To determine the resting motor threshold (RMT) for each participant, single-pulse TMS was applied to the motor cortex hand area (M1) of the left hemisphere to obtain motor-evoked potentials (MEPs) from the first dorsal interosseous muscle of the right hand via electromyography (Ware et al., Reference Ware, Lum and Kirkovski2021; Timofeeva et al., Reference Timofeeva, Finisguerra, D’Argenio, García, Carreiras, Quiñones and Amoruso2024). Specifically, the coil was positioned in a belly-tendon montage, with the ground electrode on the right wrist. Initially, the stimulation intensity was set to 35% of the maximal stimulator output (MSO). This intensity was gradually increased in 5% MSO increments until MEPs with peak-to-peak amplitudes exceeding 50 μV were consistently elicited. Subsequently, the intensity was systematically decreased in 1% MSO increments until five out of ten consecutive stimuli elicited MEPs exceeding 50 μV, thereby defining this intensity as the RMT (e.g., Rossini et al., Reference Rossini, Barker, Berardelli, Caramia, Caruso, Cracco and Tomberg1994, Reference Rossini, Burke, Chen, Cohen, Daskalakis, Di Iorio and Ziemann2015; Steel et al., Reference Steel, Song, Bageac, Knutson, Keisler, Saad and Wilkinson2016). The measured RMT for all participants ranged from 42% to 72%, with a mean of 58.48% ± 6.62%. Subsequently, following the protocol established by Huang et al. (Reference Huang, Edwards, Rounis, Bhatia and Rothwell2005), cTBS was delivered as bursts of three pulses at 50 Hz, repeated every 200 ms (5 Hz), for a total of 600 pulses over 40 seconds. The stimulation intensity was set at 80% of the RMT (Jung & Ralph, Reference Jung and Ralph2021; Steel et al., Reference Steel, Song, Bageac, Knutson, Keisler, Saad and Wilkinson2016).
During the three TMS sessions, participants were comfortably seated with their chins resting on the chin support, ensuring head stability while keeping their eyes closed throughout the procedure. Participants wore soundproof earplugs during stimulation to minimize auditory discomfort. Immediately after stimulation, participants began the behavioral task to ensure the entire experiment occurred within the neuromodulatory effect period. The stimulation coil was positioned tangentially on the scalp above the LIFG, the LMTG and the vertex (as the baseline) at an approximate angle of 45° to the midsagittal plane, with the handle oriented laterally and posteriorly. The angle of the coil was adjusted conservatively to minimize the risk of inadvertent facial nerve stimulation for each participant. A navigation system was employed to ensure the precise placement of the coil at the predetermined target location (see above).
2.5. Data analysis
2.5.1. Data preprocessing
Prior to the formal data analysis, we extracted the onset times of the verbal recordings obtained from the language switching task using Praat (Boersma & Weenink, Reference Boersma and Weenink2018). We first emphasized the sound, and by using a 100-ms time step to calculate the intensity and applying a minimum pitch setting of 0 Hz, we converted each recording into an intensity contour. Each recording, representing each trial/word production, was subsequently segmented into periods of silence and speech sound by setting the TextGrid parameters, which included a silence threshold of −25 dB, a minimum duration of 100 ms for a silence period and a minimum duration of 50 ms for a speech sound period. The RTs for the language switching task were defined as the onset times of the production of speech sounds when participants named the stimuli.
We also preprocessed the data of our 33 participants for both language switching tasks and nonverbal switching tasks. Specifically, for the language switching tasks, trials were excluded from the subsequent statistical analyses on RTs in any of the following circumstances: no response (0.68%), responses in the incorrect language (4.09%), responses following the naming trial of the incorrect language (3.22%), incorrect words from the correct language (0.34%), hesitations (0.35%), absolute outliers (below 300 ms or above 2200 ms, 0.08%) and relative outliers (2.5 standard deviations above or below the individual mean, 2.14%). For the ER data, trials following a language error trial were excluded (3.61%), while all other correct and incorrect trials were retained for further analyses. Additionally, the audio recording for the vertex level condition was missing for one participant in the language switching task; therefore, the vertex level data for this participant were excluded from subsequent analyses.
For the nonverbal switching tasks, we preprocessed the RT data by excluding trials with keypress errors or missing (8.20%), trials immediately following a keypress error trial (6.86%), trials with RTs identified as absolute outliers (below 300 ms or above 2200 ms; 3.89%) and trials with RTs identified as relative outliers (those beyond the mean ± 2.5 standard deviations; 1.49%). For the ER data, only trials following an error trial were excluded (8.20%), while all other correct and incorrect trials were retained for further analyses.
To mitigate the influence of extreme values on the results, data exceeding 2.5 standard deviations for both switching and mixing costs were excluded from language switching tasks and nonverbal switching tasks at the group level. For instance, the switching costs for all participants at three sessions were calculated. Participants whose switching costs fell beyond 2.5 standard deviations from the mean were identified. Then, switch and repeat trials for these participants at the corresponding TMS sessions (accounting for 2.02%) were excluded before analyzing switching costs in RTs for all participants. This approach was applied to the analysis of mixing costs in RTs for language switching tasks, as well as the switching and mixing costs for nonverbal switching tasks, with respective proportions of 1.01%, 1.01% and 3.03% trials. Similarly, this approach was employed in the analysis of ERs, where the switching and mixing costs for language switching tasks and nonverbal switching tasks accounted for 4.04%, 3.03%, 2.02% and 4.04% trials, respectively. After these exclusions, the remaining data were analyzed further to fit models.
2.5.2. Statistical analysis
The RT data were analyzed using linear mixed-effects models (LMEMs). We selected LMEMs over traditional analysis of variance because they enable the inclusion of control variables and allow for both by-participant (F1) and by-item (F2) analyses within a unified framework. This approach enhances the appropriateness of data modeling and increases the generalizability of the results to other participants and items (e.g., Baayen et al., Reference Baayen, Davidson and Bates2008). The analyses of ER data were conducted using mixed-effects logistic regression models (MELRMs) since the ER data were binary.
We began with comprehensive models that included fixed effects and maximal random-effects structures for LMEMs (Barr et al., Reference Barr, Levy, Scheepers and Tily2013) and for MELRMs (Harris, Reference Harris2021). Specifically, the fixed effects encompassed trial type (for switching costs: switch versus repeat; for mixing costs: repeat versus single), TMS session (LIFG versus vertex; LMTG versus vertex) and their interactions. The random intercepts for both participants and items and the random slopes for all the fixed effects were included as random effects. Our full models were built as follows: RT or ER ~ TMS session × trial type + (1 + TMS session × trial type | participant) + (1 + TMS session × trial type | item) (model−1). RT was transformed using ordered quantile normalization with the bestNormalize package (Peterson, Reference Peterson2021). For this model selection, we utilized a likelihood ratio test with a significance threshold set at 0.2. In the LMEMs and MELRMs analyses, dummy coding was applied to the TMS session, designating the vertex as the baseline ([0 0]), with LIFG coded as [1 0] and LMTG as [0 1]. Contrast coding was employed to represent trial types, defining switching costs as (switch trials = 0.5, repeat trials = −0.5) and mixing costs as (repeat trials = 0.5, single trials = −0.5). When model convergence was not achieved, we implemented a stepwise reduction approach to simplify the random-effects structures. This process began by eliminating the random slope interactions that explained the smallest amount of variance. When necessary, we also excluded the main effects from the random structure (Barr et al., Reference Barr, Levy, Scheepers and Tily2013). Following the method outlined by Matuschek et al. (Reference Matuschek, Kliegl, Vasishth, Baayen and Bates2017), we continued the simplification process until further reductions would significantly compromise the model fit.
2.5.3. Modeling for switching costs
Ultimately, for the language switching task, the final model for switching costs in RT was specified as follows: RT ~ TMS session × trial type + (1 + TMS session + trial type | participant) + (1 | item) (model-2-LS-sc-RT). For the nonverbal switching task, the final model of switching costs in RT was as follows: RT ~ TMS session × trial type + (1 + TMS session + trial type | participant) + (1 | item) (model-3-NS-sc-RT).
Regarding the ER data, for the language switching task, the final model for switching costs was represented as follows: ER ~ TMS session × trial type + (1 | participant) + (1 | item) (model-4-LS-sc-ER). For the nonverbal switching task, the final model for switching costs was specified as follows: ER ~ TMS session × trial type + (1 + TMS session + trial type | participant) + (1 | item) (model-5-NS-sc-ER).
2.5.4. Modeling for mixing costs
Regarding the RT data, for the language switching task, the final model for mixing costs was specified as follows: RT ~ TMS session × trial type + (1 + TMS session + trial type | participant) + (1 | item) (model-6-LS-mc-RT). For the nonverbal switching task, the final model of mixing costs: RT ~ TMS session × trial type + (1 + TMS session × trial type | participant) + (1 | item) (model-7-NS-mc-RT).
Regarding the ER data, for the language switching task, the final model for mixing costs was specified as follows: ER ~ TMS session × trial type + (1 | participant) + (1 | item) (model-8-LS-mc-ER). For the nonverbal switching task, the final model for mixing costs was specified as follows: ER ~ TMS session × trial type + (1 + trial type | participant) + (1 + trial type | item) (model-9-NS-mc-ER).
Given our research questions, we focused on the fixed effects of trial type to estimate switching costs or mixing costs at baseline (i.e., when the vertex was stimulated). Additionally, the interaction between trial type and TMS session was examined as our main interest. This analysis aimed to determine whether stimulating the target brain regions (LIFG or LMTG) significantly affects switching costs and mixing costs. A summary of all models, together with their respective dependent variables and analysis types, was provided in Table S1 in the supplementary materials.
In this study, the analyses of LMEMs and MELRMs were conducted with the lme4 package (Bates et al., Reference Bates, Mächler, Bolker and Walker2015). All the statistical analyses were implemented in the R statistical computing environment (R Core Team, 2013).
3. Results
The descriptive statistics for RTs and ERs for language switching tasks and nonverbal switching tasks across TMS sessions are presented in Table 1.
Table 1. Descriptive statistical results of RTs and ERs
Notes: RT refers to reaction time, and ER indicates error rate. LIFG, LMTG and vertex, which are the stimulated brain regions during the TMS sessions, refer to the left inferior frontal gyrus, the left middle temporal gyrus and the vertex (baseline), respectively. Switching cost is calculated as the mean of switch trials minus the mean of repeat trials, and mixing cost is calculated as the mean of repeat trials minus the mean of single trials. The values in parentheses indicate standard deviation. The bolded data in LIFG and LMTG marked with an asterisk (*) indicate a significant difference (p < 0.05) when compared to the vertex.
3.1. Switching costs
The results of the switching costs in RTs are shown in Table 2 and Figure 3. The switching costs in RTs from the language switching task were evaluated using model-2-LS-sc-RT. Accordingly, the effect of trial type [switch versus repeat] was significant at the baseline (t = 10.89, p < 0.001), indicating that RTs were significantly longer on switch trials than on repeat trials (1142 ms and 1083 ms, respectively) after the stimulation of the vertex. This difference yielded the switching costs of 59 ms. The effect of TMS session [LIFG versus vertex] was not significant (t = −1.05, p = 0.292), suggesting that stimulation of the LIFG did not affect RTs in the dual-language context relative to baseline. In contrast, the effect of TMS session [LMTG versus vertex] was significant (t = −2.60, p = 0.009), demonstrating a reduction in RTs in the dual-language context following stimulation over the LMTG compared to baseline. There was no significant interaction between TMS session [LIFG versus vertex; LMTG versus vertex] and trial type [switch versus repeat] (LIFG: t = 0.17, p = 0.864; LMTG: t = −0.07, p = 0.940). This suggests that, compared to the baseline, neither stimulation over the LIFG nor the LMTG significantly affected switching costs (LIFG: switching costs = 60 ms; LMTG: switching costs = 55 ms).
Table 2. Estimated coefficients (RTs) from LMEMs for examining the switching costs
Notes: RT, reaction time; b, raw (unstandardized) coefficient. RT was transformed using ordered quantile normalization with the bestNormalize package (Peterson, Reference Peterson2021). Contrast coding was applied to trial type (switch = 0.5, repeat = −0.5). TMS session was dummy coded, with the vertex set as the baseline ([0 0]), while the LIFG was coded as [1 0] and the LMTG as [0 1]. TMS, transcranial magnetic stimulation; LIFG, the left inferior frontal gyrus; LMTG, the left temporal gyrus; vertex, the baseline. The bold values represent significant differences (p < 0.05).
Figure 3. The effects of cTBS on the switching costs in RTs for the language switching task (left) and nonverbal switching task (right) after different TMS sessions. The red dot represents the mean, and the numbers on the right indicate the mean values.
Note: cTBS, continuous theta burst stimulation; LIFG, the left inferior frontal gyrus; LMTG, the left middle temporal gyrus; vertex, the baseline.
The switching costs in RTs from the nonverbal switching task were analyzed using model-3-NS-sc-RT. The effect of trial type [switch versus repeat] was significant at the baseline (t = 7.45, p < 0.001), indicating that RTs were significantly longer for switch trials (761 ms) compared to repeat trials (708 ms). This difference resulted in significant switching costs of 53 ms following stimulation of the vertex. The effect of TMS session [LIFG versus vertex; LMTG versus vertex] was not significant (LIFG: t = −0.25, p = 0.799; LMTG: t = −0.88, p = 0.381), indicating that stimulation of the LIFG nor the LMTG failed to significantly influence RTs compared to baseline. A significant interaction was observed between the TMS session [LIFG versus vertex] and trial type [switch versus repeat] (t = 2.07, p = 0.038), suggesting that stimulation of the LIFG resulted in a notable increase in switching costs (67 ms) relative to the vertex condition. This result suggests that stimulation of the LIFG had a significant impact on local cognitive control in RT (Figure 3). However, the interaction between TMS session [LMTG versus vertex] and trial type [switch versus repeat] was not significant (t = 1.17, p = 0.242), indicating that the increase in switching costs (58 ms) caused by the stimulation over the LMTG was not significant compared to that of the vertex.
The results of the switching costs in ERs are presented in Figure S1 and Table S2 as supplementary materials. The switching costs in ERs from the language switching task were analyzed using model-4-LS-sc-ER. The effect of trial type [switch versus repeat] was significant at the baseline (z = 5.42, p < 0.001), indicating that ERs were significantly larger for switch trials (7.34%) compared to repeat trials (4.06%) after the stimulation of the vertex. This difference resulted in significant switching costs of 3.29% at the baseline. The effect of TMS session [LIFG versus vertex; LMTG versus vertex] was not significant (LIFG: z = −0.21, p = 0.830; LMTG: z = 1.30, p = 0.194), indicating that stimulation of the LIFG nor the LMTG failed to produce a significant effect on ERs in the dual-language context compared to baseline. There was no significant interaction between TMS session [LIFG versus vertex; LMTG versus vertex] and trial type [switch versus repeat] (LIFG: z = 0.38, p = 0.705; LMTG: z = 0.15, p = 0.878). This indicates that, compared to the baseline, no significant effects were observed on the switching costs after stimulating the LIFG (switching costs = 3.45%) or LMTG (switching costs = 3.76%).
The switching costs in ERs from the nonverbal switching task were assessed using model-5-NS-sc-ER. The effect of trial type [switch versus repeat] was significant at the baseline (z = 4.55, p < 0.001), showing that ERs were significantly larger for switch trials (11.25%) compared to repeat trials (6.93%). This resulted in the switching costs of 4.32% after the stimulation of the vertex. The effect of TMS session [LIFG versus vertex; LMTG versus vertex] was not significant (LIFG: z = 0.23, p = 0.819; LMTG: z = 0.13, p = 0.900), suggesting that stimulating the LIFG or the LMTG did not influence the ERs compared to the baseline. No significant interaction was detected between TMS session and trial type (LIFG: z = −1.52, p = 0.128; LMTG: z = −1.03, p = 0.305). That is, relative to the baseline, stimulation over the LIFG and LMTG did not yield significant effects on the switching costs (LIFG: switching costs = 3.01%; LMTG: switching costs = 3.43%).
3.2. Mixing costs
The results regarding the mixing costs in RTs are presented in Table 3 and Figure 4. The mixing costs in RTs from the language switching task were assessed using model-6-LS-mc-RT. The effect of trial type [repeat versus single] was significant at the baseline (t = 17.36, p < 0.001), indicating that RTs were significantly longer on repeat trials (1083 ms) than that of single trials (930 ms) following stimulation over the vertex. This difference resulted in significant mixing costs of 153 ms. A significant effect of TMS session [LIFG versus vertex] was not found (t = −0.77, p = 0.443), suggesting that stimulating the LIFG did not significantly influence the RTs compared to the baseline. In contrast, the effect of TMS session [LMTG versus vertex] was significant (t = −2.55, p = 0.011), indicating that the RTs significantly reduced after stimulation of the LMTG compared to the baseline. There was no significant interaction between TMS session [LIFG versus vertex; LMTG versus vertex] and trial type [switch versus repeat] (LIFG: t = −0.38, p = 0.705; LMTG: t = −0.91, p = 0.363). This suggests that relative to the baseline, neither stimulation of the LIFG nor the LMTG significantly influenced mixing costs (LIFG: mixing costs = 145 ms; LMTG: mixing costs = 140 ms). These findings indicate that the modulation of these two target brain regions did not significantly impact the global control of bilingual language processing in RTs.
Table 3. Estimated coefficients (RTs) from the LMEMs for examining the mixing costs
Notes: RT, reaction time; b, raw (unstandardized) coefficient. RT was transformed using ordered quantile normalization with the bestNormalize package (Peterson, Reference Peterson2021). Contrast coding was applied to trial type (repeat = 0.5, single = −0.5). TMS session was dummy coded, with the vertex set as the baseline ([0 0]), while the LIFG was coded as [1 0] and the LMTG as [0 1]. TMS, transcranial magnetic stimulation; LIFG, the left inferior frontal gyrus; LMTG, the left temporal gyrus; vertex, the baseline. Bold values denote significant differences (p < 0.05).
Figure 4. The effects of cTBS on the mixing costs in RTs for the language switching task (left) and nonverbal switching task (right) in different TMS sessions. The red dot represents the mean, and the numbers on the right indicate the mean values.
Note: cTBS, continuous theta burst stimulation; LIFG, the left inferior frontal gyrus; LMTG, the left middle temporal gyrus; vertex, the baseline.
The mixing costs in RTs from the nonverbal switching task were analyzed using model-7-NS-mc-RT. The effect of trial type [repeat versus single] was significant at the baseline (t = 22.88, p < 0.001), indicating that RTs were significantly longer for repeat trials (705 ms) compared to single trials (456 ms). This difference resulted in significant mixing costs of 248 ms following the stimulation of the vertex. The effect of TMS session [LIFG versus vertex; LMTG versus vertex] was not significant (LIFG: t = −0.57, p = 0.568; LMTG: t = −1.14, p = 0.253), suggesting that stimulating the LIFG nor the LMTG did not significantly influence the RTs compared to the baseline. There was no significant interaction between TMS session [LIFG versus vertex; LMTG versus vertex] and trial type [switch versus repeat] (LIFG: t = −0.55, p = 0.580; LMTG: t = −0.82, p = 0.412). This suggests that relative to the baseline, neither the stimulation of the LIFG nor the LMTG significantly influenced mixing costs (LIFG: mixing costs = 234 ms; LMTG: mixing costs = 223 ms).
The results of the mixing costs in ERs are presented in Figure S2 and Table S3 as supplementary materials. The mixing costs in ERs from the language switching task were evaluated using model-8-LS-mc-ER. The effect of trial type [repeat versus single] was significant at the baseline (z = 6.53, p < 0.001), indicating that ERs were significantly larger for repeat trials (3.23%) compared to single trials (0.58%) following the stimulation of the vertex. This difference resulted in significant mixing costs of 2.66%. The effect of TMS session [LIFG versus vertex; LMTG versus vertex] was not significant (LIFG: z = 1.74, p = 0.082; LMTG: z = 1.39, p = 0.165), suggesting that stimulating the LIFG nor the LMTG did not significantly affect the ERs compared to the baseline. There was no significant interaction between TMS session [LIFG versus vertex; LMTG versus vertex] and trial type (LIFG: z = −1.13, p = 0.259; LMTG: z = 0.15, p = 0.877). This indicates that relative to the baseline, no significant effects were observed on the mixing costs following the stimulation of the LIFG (mixing costs = 2.65%) or the LMTG (mixing costs = 3.59%).
The mixing costs in ERs from the nonverbal switching task were evaluated using model-9-NS-mc-ER. The effect of trial type [repeat versus single] was not significant at the baseline (z = 1.12, p = 0.264), showing that ERs for repeat trials (6.74%) and single trials (4.02%) were not significantly different after the stimulation of the vertex, resulting in nonsignificant mixing costs of 2.72%. The effect of TMS session [LIFG versus vertex; LMTG versus vertex] was not significant (LIFG: z = 1.08, p = 0.280; LMTG: z = 0.52, p = 0.600), indicating that stimulation of the LIFG nor the LMTG did not significantly affect ERs compared to the baseline. No significant interaction was detected between TMS session [LIFG versus vertex; LMTG versus vertex] and trial type [switch versus repeat] (LIFG: z = 0.72, p = 0.470; LMTG: z = −0.57, p = 0.569). That is, relative to the baseline, stimulation of the LIFG and the LMTG did not produce significant effects on the mixing costs (LIFG: mixing costs = 3.47%; LMTG: mixing costs = 2.44%).
In summary, our results show classic switching costs and mixing costs in both language switching and nonverbal switching tasks across all three TMS sessions, consistent with previous work investigating language control and cognitive control (e.g., Jylkkä et al., Reference Jylkkä, Lehtonen, Lindholm, Kuusakoski and Laine2018, Reference Jylkkä, Soveri, Laine and Lehtonen2020; Meuter & Allport, Reference Meuter and Allport1999; Wang et al., Reference Wang, Wu, Ji, Yan and Wu2022; Wu et al., Reference Wu, Yang, Chen, Li, Zhang, Kang and Guo2019, Reference Wu, Ji, Qu, Zuo, Liang, Su and Ding2025). Importantly, our TMS sessions only showed an effect on LIFG in nonverbal switching tasks, as reflected in the increased switching costs. In addition, TMS session targeting at the LMTG appeared to facilitate picture naming in general (as in both single and mixed blocks), compared to the baseline.
4. Discussion
To our knowledge, this study provides the first causal evidence on the roles of the LIFG and LMTG in both language control and cognitive control, using TMS in unbalanced Chinese-English bilinguals. Situated within the broader debate between overlap accounts (e.g., Green, Reference Green1998; Green & Abutalebi, Reference Green and Abutalebi2013) and dissociation accounts (e.g., Calabria et al., Reference Calabria, Hernández, Branzi and Costa2012, Reference Calabria, Suades, Juncadella, Ortiz-Gil, Ugas, Sala and Lleo2025; Branzi et al., Reference Branzi, Calabria, Boscarino and Costa2016), our findings lend stronger support to the latter. LIFG stimulation selectively increased nonverbal switching costs, without affecting language switching or mixing costs, nor nonverbal mixing costs, providing causal evidence that this region is engaged in cognitive control but not critically involved in language control. This dissociation contrasts with the predictions of overlap accounts, which anticipate parallel effects across linguistic and nonlinguistic tasks. By comparison, LMTG stimulation produced no significant changes in switching or mixing costs in either domain, but significantly reduced RTs in the language switching task, consistent with a facilitative role in lexical–semantic processing rather than in control per se. Overall, these results strengthen the view that language control and cognitive control are at least partially dissociable at the neural level, with the LIFG subserving domain-general control processes and the LMTG supporting language-specific processing.
With regard to the effects of LIFG stimulation, our results revealed a significant increase in nonverbal switching costs, without a significant effect on language switching costs. This finding contradicts previous fMRI studies (e.g., De Baene et al., Reference De Baene, Duyck, Brass and Carreiras2015; Jiao, Meng et al., Reference Jiao, Meng, Wang, Schwieter and Liu2022), as well as our initial hypothesis that language control and cognitive control share common neural mechanisms as in the LIFG. Prior research has investigated the neural substrates in bilinguals underlying language control and cognitive control by employing both language switching tasks and nonverbal switching tasks (e.g., De Baene et al., Reference De Baene, Duyck, Brass and Carreiras2015; Jiao, Meng et al., Reference Jiao, Meng, Wang, Schwieter and Liu2022). For instance, a meta-analysis by Jiao, Meng et al. (Reference Jiao, Meng, Wang, Schwieter and Liu2022) reported significant LIFG activation in both task types, which contrasts with the current results. Other studies have shown the neural differences between bilinguals and monolinguals in nonverbal switching tasks (e.g., Anderson et al., Reference Anderson, Chung-Fat-Yim, Bellana, Luk and Bialystok2018; Garbin et al., Reference Garbin, Sanjuan, Forn, Bustamante, Rodríguez-Pujadas, Belloch and Ávila2010). For example, Garbin et al. (Reference Garbin, Sanjuan, Forn, Bustamante, Rodríguez-Pujadas, Belloch and Ávila2010) found that bilingual participants exhibited reduced switching costs and increased activation in the LIFG, while monolinguals did not show these effects. They attributed these effects to bilinguals’ early training in switching between languages, which leads to the recruitment of brain regions, such as the LIFG involved in language control, and benefits their performance in nonverbal cognitive tasks as a bilingual advantage. However, this reasoning warrants reconsideration, as the LIFG is a critical region frequently associated with cognitive control (Berkman et al., Reference Berkman, Kahn and Merchant2014; Philipp et al., Reference Philipp, Weidner, Koch and Fink2013; see Irlbacher et al., Reference Irlbacher, Kraft, Kehrer and Brandt2014, for a review). Based on our findings, we propose that the LIFG activation observed in bilingual neuroimaging studies may be epiphenomenal, with its causal role limited to cognitive control.
Regarding mixing costs, we found that LIFG stimulation had no significant impact on either language switching tasks or nonverbal switching tasks, consistent with previous research (e.g., Branzi et al., Reference Branzi, Humphreys, Hoffman and Ralph2020; Braver et al., Reference Braver, Reynolds and Donaldson2003; Guo et al., Reference Guo, Liu, Misra and Kroll2011). As outlined in the Introduction, mixing costs are thought to reflect sustained, proactive mechanisms that prevent interference at the global level, with switching costs instead associated with more transient, reactive mechanisms that resolve interference at the local level (Ma et al., Reference Ma, Li and Guo2016; Timmer, Christoffels et al., Reference Timmer, Christoffels and Costa2019; Wu et al., Reference Wu, Ji, Qu, Zuo, Liang, Su and Ding2025). In the domain of cognitive control, Braver et al. (Reference Braver, Reynolds and Donaldson2003) investigated the neural substrates of mixing costs using nonverbal switching tasks and found that the right anterior prefrontal cortex is selectively engaged. In the domain of language control, Guo et al. (Reference Guo, Liu, Misra and Kroll2011) compared the neural mechanisms underlying local versus global language control, finding distinct brain region activations: the anterior cingulate cortex and the pre-supplementary motor area were associated with local control, whereas the left dorsolateral prefrontal cortex and parietal cortex were critical for global control during bilingual word production. Notably, these studies did not report direct involvement of the LIFG. Thus, in the present study, we reason that the LIFG may not play a causal role in the proactive and global aspects of either cognitive control or language control.
In contrast, stimulation of the LMTG produced little effect on switching and mixing costs in both language switching and nonverbal switching tasks. These results contradict our initial hypothesis that the LMTG plays divergent roles in language control and cognitive control. The LMTG, encompassing the left posterior middle and superior temporal gyri, is largely engaged in language processing, such as semantic and phonological operations (De Bleser et al., Reference De Bleser, Dupont, Postler, Bormans, Speelman, Mortelmans and Debrock2003; Binder et al., Reference Binder, Westbury, McKiernan, Possing and Medler2005; Lau et al., Reference Lau, Phillips and Poeppel2008; Bemis & Pylkkänen, Reference Bemis and Pylkkänen2011). Therefore, while its lack of involvement in cognitive control is somewhat expected, the absence of an effect on language control is surprising and appears to contradict previous studies (e.g., Bemis & Pylkkänen, Reference Bemis and Pylkkänen2011). In particular, Wu, Ji et al. (Reference Wu, Ji, Cai, Pu, Wang, Yan and Wang2024) found that applying online TMS to the LSTG (MNI coordinates: x = −54, y = −12, z = 6), a region engaged in language processing, significantly reduced switching costs in language switching tasks at certain time windows upon picture onsets. Their results further suggest that the LSTG may act as an interface between language control and language processing, namely, a language processing region but interacting with language control. Like the LSTG in Wu, Ji et al. (Reference Wu, Ji, Cai, Pu, Wang, Yan and Wang2024), the LMTG examined in the current study is also a classic area responsible for language processing, located approximately 36 mm from the LSTG investigated in Wu, Ji et al. (Reference Wu, Ji, Cai, Pu, Wang, Yan and Wang2024). However, we did not observe significant effects on either switching costs or mixing costs upon the stimulation of the LMTG. These results suggest that the LMTG plays a different role from the LSTG, less crucial for language control but more specific to language processing.
Importantly, despite the null effects of the LMTG stimulation on both language and cognitive control, the notable effect was on the facilitation of picture naming in language switching tasks. These results indicate that the LMTG is crucially engaged in language processing, as reflected in facilitation rather than inhibition upon stimulation. Given that the existing literature largely emphasizes the LMTG’s role in language processing (De Bleser et al., Reference De Bleser, Dupont, Postler, Bormans, Speelman, Mortelmans and Debrock2003; Binder et al., Reference Binder, Westbury, McKiernan, Possing and Medler2005; Bemis & Pylkkänen, Reference Bemis and Pylkkänen2011), it is not surprising to observe enhanced language processing following stimulation of the LMTG. Our findings are also in line with previous studies using similar methods (e.g., Lau et al., Reference Lau, Phillips and Poeppel2008; Choi et al., Reference Choi, Park and Paik2015; Leff et al., Reference Leff, Crinion, Scott, Turkheimer, Howard and Wise2002). For instance, Choi et al. (Reference Choi, Park and Paik2015) reported that participants receiving repetitive TMS to a subregion of the LMTG, namely, the posterior portion of the middle temporal gyrus, experienced disruptions in lexical decision, a language processing task. This highlights the critical role of the LMTG in lexical processing. Moreover, Leff et al. (Reference Leff, Crinion, Scott, Turkheimer, Howard and Wise2002) found that patients with LMTG infarctions exhibited significant deficits in word comprehension, further underscoring the LMTG’s integral function in language processing. Thus, consistent with these studies, our findings suggest that the LMTG is not causally involved in language control but is crucial for language processing.
Of note, several considerations and limitations of the present study warrant acknowledgment. One important issue is the use of different cue types across tasks, which may have influenced RTs and affected task comparability. Specifically, RTs in the language-switching tasks with face cues in the present study (approximately 928 ms) were longer than those with color cues in our previous study (approximately 754 ms; Wu et al, Reference Wu, Ji, Qu, Zuo, Liang, Su and Ding2025). We confirmed that participants could clearly perceive both cue types (visual angle >3.73°, accuracy >91.18%), ruling out perceptual difficulty as a cause of the longer RTs. Nevertheless, the use of different cues may raise concerns about the comparability between language and nonlanguage tasks. In fact, the literature identifies two ideal contrasts: tasks that are identical except for one key difference, allowing any observed effect to be attributed to that difference (e.g., De Baene et al., Reference De Baene, Duyck, Brass and Carreiras2015), and tasks that differ in all but one shared feature, where consistent patterns across tasks suggest a common underlying mechanism (e.g., de Bruin et al., Reference de Bruin, Roelofs, Dijkstra and FitzPatrick2014, Calabria et al., Reference Calabria, Costa, Green and Abutalebi2018, Yahya et al., Reference Yahya and Özkan Ceylan2022, comparing language switching with Simon or Stroop tasks). Most studies, including ours, fall somewhere between these two extremes. It is common to compare tasks such as cued language switching and color–shape switching that differ in response modality (verbal vs manual: Vaughn et al., Reference Vaughn, Watlington, Linares Abrego, Tamber-Rosenau and Hernandez2021; Weissberger et al., Reference Weissberger, Gollan, Bondi, Clark and Wierenga2015; Wu et al., Reference Wu, Yang, Chen, Li, Zhang, Kang and Guo2019) or cue type (national flag vs shape-color: Anderson et al., Reference Anderson, Chung-Fat-Yim, Bellana, Luk and Bialystok2018; Prior & Gollan, Reference Prior and Gollan2013; Segal et al., Reference Segal, Stasenko and Gollan2019; text vs non-text: Cattaneo et al., Reference Cattaneo, Costa, Gironell and Calabria2020). This variation is partly justified by the way switching and mixing costs are computed, as contrasting otherwise identical conditions within each task minimizes the impact of task-specific features. Accordingly, the cue-type variation introduced in the present design remains well within the range of contrasts commonly used in previous studies. Future research may systematically examine how highly matched versus minimally matched tasks influence investigations of the relationship between language control and cognitive control.
Another key issue is the absence of a double dissociation in the present study, which would have more clearly demonstrated the divergence between cognitive and language control. Instead, we observed a single dissociation: TMS to the LIFG selectively affected switching costs in nonverbal tasks but had no effect on language switching, suggesting a primary role in cognitive control. Conversely, TMS to the LMTG reduced RTs in the language-switching task without influencing nonverbal performance, indicating functional specificity for language-related processing. Although this does not constitute a full double dissociation, these objective findings provide initial causal evidence that advances prior research and offers new insights into the relationship between language control and cognitive control. Further research targeting additional brain regions may help establish a double dissociation and provide more definitive evidence regarding the neural substrates underlying these two forms of control. In sum, these considerations underscore the complexity of disentangling cognitive and language control at the neural level and highlight the need for continued methodological refinement in this area.
Taken together, our findings refine the understanding of the causal roles of the LIFG and LMTG in language and cognitive control. The LIFG appears to be selectively engaged in cognitive control, with no critical involvement in language control, challenging models that posit substantial overlap between the two domains. The LMTG, while not directly implicated in control processes, may facilitate language processing. This pattern supports a partial dissociation between language and cognitive control mechanisms, particularly with respect to the LIFG. Theoretically, these results advance models of bilingual control by providing causal evidence against full-overlap accounts. Practically, they underscore the need for domain-specific assessments and targeted interventions in bilingual populations.
5. Conclusions
In the present study, we examined the causal roles of the LIFG and LMTG in language control and cognitive control using TMS to determine whether these two brain regions share the same cognitive-neural functions. Our findings show that stimulation of the LIFG only significantly increases nonverbal switching costs, suggesting its involvement in domain-general cognitive processes rather than language control. In contrast, the LMTG stimulation did not affect switching costs but reduced RTs during language switching tasks, suggesting its specialization in language processing. Overall, our study offers insights into the diverging neural mechanisms in the LIFG that may underlie language control and cognitive control.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S1366728925100783.
Data availability statement
All data and R scripts are publicly available via the GitHub: https://github.com/AnnnnnieKe/TMS-Reveals-Divergent-Roles-of-Left-IFG-in-BLC-and-Domain-general-Cognitive-Control.
Acknowledgements
We would like to thank Dr. Yannan Ji and Dr. Qiping Wang for their inspirational discussions. We also thank Ms. Yao Cheng and Ms. Hongfu Qu for their assistance in data collection, Dr. Luyao Chen for his contributions to experimental design, Dr. Qiang Wang for collecting the MRI data and Dr. Da Li for guiding the operation of the TMS stimulator.
Funding statement
This work was supported by the Humanities and Social Sciences Research Project of the Ministry of Education (Grant No. 23YJCZH041 to YD), the National Nature Science Foundation of China (Grant No. 32100854 to JW), the China Scholarship Council (Grant No. 202308120071 to JW), the Key Research Project on Economic and Social Development in Heilongjiang Province in 2025 (Grant No. WY2025006 to JW), the China Scholarship Council (to JW), the Australia Research Council (Grant No. DP 210102789 to XW), the Shenzhen Higher Institution Stability Support Plan (Grant No. 20231123133926001 to HL), and the Guangdong Basic and Applied Basic Research Foundation (Grant No. 2024A1515011572 to HL).
Competing interests
The authors declare none.
Open access
