The white matter of Aha! moments

Abstract

Insights, or "Aha!" moments, are a crucial aspect of idea generation in creative cognition. While functional neuroimaging studies have identified brain regions involved in these insights, their white matter substrate remains unexplored. This study employed diffusion tensor imaging (DTI) to investigate how white matter microstructure—measured by fractional anisotropy (FA) and mean diffusivity (MD)—relates to individuals’ tendency to solve Compound Remote Associates problems through insight versus step-by-step analytical reasoning. After controlling for age and gender, left-hemisphere omnibus tests (Stouffer’s Z and FDR) showed significant FA associations for left dorsal tracts composites (i.e., Arcuate Fasciculus, Posterior Arcuate Fasciculus, and Superior Longitudinal Fasciculus III), while MD tracts composites trended but were not FDR-significant (p= 0.032 q= 0.081). Findings point to a left-lateralized dorsal substrate of insight. These findings suggest that insight may benefit from more diffuse connectivity patterns, allowing for broader semantic activation and cognitive flexibility. Our study provides novel evidence for distinct structural connectivity patterns associated with different idea-generation approaches, contributing to a more comprehensive understanding of the neural architecture supporting creative cognition.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40359-025-03593-0.

Introduction

Idea generation plays a crucial role in driving human innovation, from scientific discoveries to artistic breakthroughs. Scientists have identified two main ways in which people generate creative ideas and solve problems: through sudden insights or through a continuous step-by-step “analytical” process [1–3]. Insights are characterized by an unexpected discovery or transformative idea [4, 5], which emerges into awareness suddenly, in a discontinuous manner, often interrupting one’s train of thought [6, 7]. In contrast, analytical ideas are yielded by a deliberate and controlled process. Insights are accompanied by a subjective "Aha!" experience, and they entail a conceptual restructuring that results in a novel, nonobvious interpretation of people’s mindset, which is often identified as a form of creativity [6, 8]. Insightful ideas have been demonstrated to be more accurate and creative than deliberate step-by-step solutions since they rely on information that may appear distantly related to the original problem and on the retrieval of uncommon interpretations of problem elements [2, 9, 10]. This is partly because insight entails below-awareness recombination of information, allowing for the formation of novel associations that emerge into consciousness suddenly and often without warning. This subjective quality of suddenness is thought to be distinct from the more accumulative results that emerge from deliberative reasoning [7, 9–14]. While both insight and analytically derived ideas play a role in creative cognition and problem solving, their phenomenology and underlying brain circuitries are different [1, 2, 6, 15–17]. Functional neuroimaging research has elucidated the brain mechanisms underlying insight-based idea generation (for comprehensive reviews, see Kounios & Beeman [2], Salvi [6], Chesebrough et al. [15], Salvi & Bowden [16]), and a growing body of research indicates that structural brain connectivity patterns may correlate uniquely with different cognitive functions [3]. Nonetheless, research investigating the link between white matter connectivity and individual differences in insight propensity remains elusive. Diffusion tensor imaging (DTI) offers a unique opportunity to address this knowledge gap by providing detailed information about white matter microstructure and connectivity patterns in the brain [18, 19].

DTI is a magnetic resonance imaging technique that measures the diffusion of water molecules in biological tissues, particularly in white matter tracts ​[20, 21]. This method allows researchers to visualize and quantify the organization and integrity of white matter fibers, providing insights into the structural connectivity between different brain regions [22, 23]. By analyzing the direction and magnitude of water diffusion, DTI can reveal the orientation and properties of white matter pathways, offering a noninvasive means to study brain structure in vivo [24]. DTI metrics such as fractional anisotropy (FA) and mean diffusivity (MD) can be used to quantify individual differences in white matter integrity and provide information about the average molecular motion of water in brain tissue [25, 26]. In DTI analysis, converging findings from FA and MD provide validity to any underlying white matter microstructural differences. FA is a measure of the directional coherence of water diffusion, reflecting the degree of myelination and axonal integrity within a white matter tract. The MD, on the other hand, reflects the overall magnitude of water diffusion, providing information about tissue density and membrane permeability. While findings just on FA are scientifically relevant, when the effects for FA and MD are in opposing direction, it strengthens any interpretations linking structural connectivity patterns to cognitive processes rather than those being driven by nonspecific factors.

The application of DTI to investigate insight-related individual differences is crucial for several reasons. First, it can provide complementary structural information to existing functional studies, revealing the underlying white matter pathways that facilitate communication between regions activated during insight [27]. Second, DTI allows for the examination of stable, trait-like structural characteristics that may predispose individuals to experience insights more frequently [28]. By mapping white matter tracts, DTI can elucidate how different brain regions involved in insight are structurally connected, potentially revealing integrated networks supporting this cognitive process [29]. Thus, a DTI study investigating individual differences in insight problem solving would significantly advance our understanding of the structural neural substrates underlying this critical aspect of creative cognition, complementing existing functional neuroimaging findings and providing a more comprehensive picture of the neural basis of insight.

DTI literature on convergent and divergent thinking

Traditionally, scientific literature in the field of creativity distinguishes between convergent and divergent thinking [30]. Convergent thinking involves focusing on a single correct solution to a problem, whereas divergent thinking is characterized by the generation of multiple, diverse, and original solutions to open-ended problems [31]. While both of these processes contribute to creativity, they involve different cognitive mechanisms and neural pathways (e.g., Fink et al. [32]). The application of DTI to the study of creativity has yielded inconsistent findings, reflecting the complexity of creative cognition and the challenges in its measurement. Some studies have reported lower FA across several white matter tracts [33–35], whereas others have reported the opposite relationship [36–41]. These inconsistencies may be attributed to variations in task administration and the potential conflation of divergent and convergent thinking processes. The main results are outlined below.

Divergent thinking

The most common divergent thinking measure is the alternate uses task (AUT), which involves the generation of unusual uses for everyday objects [42]. Seminal investigations into the structural brain underpinnings of divergent thinking revealed intriguing patterns. Jung and colleagues [33] reported a negative correlation between FA within the left inferior frontal white matter and performance on the AUT.

Wertz and colleagues [35] extended these findings, reporting negative correlations between FA and divergent thinking abilities across a series of predominantly left-lateralized tracts spanning both frontal and temporal regions. This study employed a composite score derived from multiple creativity tasks, not limited to the AUT, which may account for the broader range of implicated tracts.

In contrast, other studies reported positive associations between FA across several bilateral white matter tracts and divergent thinking abilities [36, 39, 41]. ​​For example, Takeuchi and colleagues [39] reported positive correlations between FA and divergent thinking abilities in regions including the bilateral prefrontal cortices, corpus callosum (CC), bilateral basal ganglia, bilateral temporoparietal junctions (TPJs), and right inferior parietal lobule, suggesting that enhanced structural connectivity in these areas may support creative cognitive processes.

The inconsistencies in these results could reflect variations in the way authors measured divergent thinking, DTI, or both. First, the studies did not follow a single, standardized DTI preprocessing pipeline, and different pipelines can yield different results [43]. Second, few studies measured divergent thinking using the same task (or collection of tasks), there was also variation in how “creativity” scores were then derived from task scores. Third, this methodological variation is likely compounded by demographic variation across the samples employed in these studies, with sample features such as gender balance and linguistic and cultural background already known to mediate divergent thinking [35].

Convergent thinking

One study reported a positive correlation between performance on the German version of the remote associates task1 and FA across the right corticostriatal pathway [36]. Another study reported significant positive correlations between convergent thinking ability and FA across the left inferior longitudinal fasciculus (ILF) and the left frontal‒occipital fasciculus (FOF) as well as the CC [37, 38]. In this latter case, researchers employed a Japanese variant of the CRA (JRAT), which varies from the English and German versions, given the linguistic differences between the two languages and the variability in the task instructions. For example, the JRAT requires participants to replace linguistic units of writing (i.e., kanji) from prompt words, whereas the English and German versions do not involve replacement. Crucially, none of these studies investigates idea generation via insight. Thus, drawing reliable conclusions on individual differences in white matter structure related to insight is impossible.

In summary, while past findings collectively suggest a relationship between white matter structure and creative thinking, these findings still lack a clear and reliable explanation of the precise nature of this relationship. Additionally, they do not directly address the specific aspect of creativity that our study focuses on: the role of Aha! Moments in creative problem solving.

Insight and convergent thinking

Insight is often studied via convergent thinking tasks such as remote associates (RATs or compound remote associates—CRAs) because of their methodological advantages and statistical power [6, 44, 45]. Remote associates are quick to solve, compact, and elicit both insight and step-by-step analytical solutions, allowing for direct comparisons between problem-solving modes and efficient experimental designs. Their reliance on semantic processing makes them well-suited for neuroscientific studies of the neural correlates of insight [1, 6, 46, 47].

Neural correlates of insight

The pioneering study that delved into the neural basis of insight employed both fMRI and high-density EEG in separate experiments with a consistent methodology [1]. Among several results (see Kounios & Beeman [2] for a review), researchers have reported specific localized neural activity associated with the Aha! Moment over the right temporal cortex. The EEG results revealed a sudden burst of 40-Hertz gamma-band activity over this brain region, which occurred 300 ms before the participants pressed a button to signal their insight. Imaging results pinpointed this activity to the medial aspect of the right superior temporal gyrus (STG). This brain area is known for its role in semantic integration of distantly related associations needed for achieving global coherence in discourse processing and overall reasoning [48], as well as understanding metaphors, implicit comprehension, and humor [49–53].

Researchers have argued that the right STG supports the connection of distantly related information during insight problem solving, enabling participants to perceive associations that would otherwise be missed [1, 44]. Subsequent studies using various brain stimulation techniques, such as transcranial alternating current stimulation and transcranial direct current stimulation, provided causal evidence for the role of the right temporal lobe during insight problem solving [47, 54–57]. Conversely, the left temporal lobe appears to support finer semantic coding, characterized by more focused neural activity leading to one or a few dominant interpretations or alternative meanings [58–60]. This region tends to be “chronically inhibited” in individuals who solve problems via insight, perhaps promoting the emergence of weakly activated information processed in the right temporal lobe [61].

Early studies on neuroscience revealed that gamma-band oscillation over the right temporal lobe is preceded by alpha-band activity over the right occipital cortex, which is thought to reflect the inhibition of visual inputs [1, 62]. More recently, Yu and colleagues [63] reported that more time spent in a state with widespread alpha-band synchronization was associated with the generation of more novel metaphors. This aligns with studies showing increased alpha power during tasks requiring the generation of novel uses for common objects [64–67], novel names for abbreviations [68], and explanations for unusual situations [32].

Alpha-band synchronization is thought to facilitate selective inhibition [69] and is sensitive to internal processing demands such as creative cognition. In the context of insight problem solving, inhibition plays a crucial role in suppressing dominant but unhelpful interpretations, allowing less obvious associations to surface. Alpha oscillations, particularly in parietal‒occipital and prefrontal regions, have been linked to this inhibitory process, helping individuals disengage from irrelevant cognitive processes and shift toward novel, insightful solutions [69, 70]. This mechanism aligns with findings that insight often requires overcoming habitual thought patterns and permitting weakly activated remote associations—processed in the right hemisphere—to gain prominence [2, 6].

FA has been found to correlate with the alpha rhythm, suggesting a structural basis for individual differences in oscillatory dynamics [71]. Specifically, research indicates that FA, which reflects white matter integrity, fiber density, and myelination, is significantly related to the alpha peak frequency, with stronger correlations observed in posterior commissural fibers and thalamocortical pathway dynamics [71]​. This relationship likely stems from the role of white matter in facilitating communication between distant brain regions. Given that alpha-band activity is thought to support inhibition and top-down control processes, variations in FA could influence the efficiency of these mechanisms. For example, higher FA in the superior and posterior corona radiata, which are implicated in thalamocortical interactions, has been linked to increased alpha frequency, potentially reflecting more effective inhibition of irrelevant sensory information​. Conversely, lower FA in certain regions may contribute to weaker inhibitory control, leading to reduced alpha synchronization [71].

In summary, past work illustrates the role of lateralized functional patterns in supporting insight problem solving (for a review, see Salvi [6]), whereas emerging research highlights the importance of selective inhibition in facilitating the cognitive flexibility needed to reach insightful solutions. While relationships between white matter integrity and alpha band activity provide some initial evidence for the role of brain structures in insight, to our knowledge, no work has investigated the possible link between white matter and insightful problem solving.

Present research

Given the literature on insight and creative cognition, as well as the inconsistent findings in DTI studies in the field, we decided to investigate the structural connectivity patterns associated with individual differences in insight propensity. We thus employed DTI to examine white matter microstructure and connectivity patterns in relation to insight and analytical problem solving in the CRA.

On the basis of previous functional neuroimaging and brain stimulation studies, we generated compatible hypotheses: (1) Individuals with greater insight propensity have lower FA values in left hemisphere white matter tracts, particularly those connected to the left temporal lobe. This prediction is based on the work of Jung-Beeman and colleagues [1] and Erickson et al. [61], who suggested that inhibition of the left temporal lobe may promote the emergence of weakly activated information processed in the right temporal lobe, facilitating insight. (2) Convergent thinking is linked to higher FA values in the CC, similar to the findings of Takeuchi et al. [37–39]. However, our study will extend these findings by specifically examining whether these structural differences are associated with insight or step-by-step analytical solutions to the CRA problem, which could confound prior results. Unlike previous studies that did not differentiate between problem-solving modes, our approach allows us to determine whether the increased FA values are specifically related to insight processing or if they reflect more general problem-solving abilities. This distinction is crucial for understanding the unique structural correlates of insight as opposed to non-insight solving.

By testing these hypotheses, our study aims to bridge the gap between functional and structural neuroimaging findings in insight research, potentially revealing the underlying white matter architecture that supports individual differences in insight propensity. Furthermore, we employ a preprocessing DTI pipeline via the open-source brainlife.io platform, which ensures straightforward replicability of our methodology without the need for custom code. This investigation will contribute to a more comprehensive understanding of the neural basis of insight, complementing existing functional studies and shedding light on the stable, trait-like characteristics that may predispose individuals to experience insights more frequently.

Methods

Subjects

Forty-four right-handed, native American English speakers were recruited for the study. Participants qualified for inclusion in the study if they met the following conditions: (1) no prior history of neurological or psychiatric disorders; (2) no use of drugs affecting the central nervous system, mood, or attention (e.g., antidepressants, amphetamines, or anxiolytics); and (3) no history of traumatic brain injury or intracranial metal implants. Participants older than 45 years (3 participants) and those who solved fewer than 10 problems (2 participants) or no problems because they misunderstood the instructions (1 participant) were excluded from the analysis. Thus, a final sample of 38 participants (25 self-identified as females; 13 self-identified as males; M age = 24.55; SD = 5.2) was used for the data analysis. The participants’ level of education corresponded to an average of 15.8 years (SD = 3.1). The participants were paid for completing the study, and each experimental session lasted approximately 1.5 h.

Behavioral measures of problem solving

Participants were presented with 60 compound remote associates (CRA) problems, balanced for difficulty and randomly ordered, selected from the item set developed by Bowden and Jung-Beeman [44]. CRA problems featured words displayed in 28-point Times New Roman font, presented in black text on a white background, and were centrally aligned. During the instructional phase, participants were trained to distinguish between insight and step-by-step analytical problem-solving approaches for solving a problem. Each trial featured three stimulus words (e.g., crab, pine, sauce), and participants were instructed to identify a fourth word (e.g., apple) that could form a common compound word or familiar phrase with each of the three. They were given 15 s to respond. Once participants thought they had a solution, they had to press any button on the mouse, report the solution to the experimenter, who would record its accuracy on a keyboard as well as on a paper spreadsheet. Then participants had to report if the solution came to mind as an insight or in a step-by-step fashion by pressing the left or the right button of the mouse (counterbalanced order across participants). When participants were ready, the experimenter would manually press a key on the keyboard to move on to the next trial. This procedure allowed participants to take breaks if needed between problems. If participants did not find a solution within the 15 s, the problem would disappear from the screen, and the experimenter would move forward to the next trial by pressing a key on the keyboard. The reliability and validity of such self-reports of insight-based problem solving have been supported by extensive behavioral and neuroimaging research (e.g., [1, 47, 72]). Over the past twenty years, CRA problems have become a widely accepted method for studying insight-based problem solving and its neural correlates, as they retain the essential features of traditional insight tasks while offering greater reliability and statistical power [1, 6, 73]. Furthermore, performance on CRA problems has been shown to positively correlate with success in solving classic insight problems [11, 14, 74].

The experiment was conducted via E-Prime 2.10 software (Psychology Software Tools, Inc.), which was presented on a 24-inch Dell screen with participants viewing from approximately 60 cm away.

Image acquisition

The participants underwent MR imaging via a Siemens 3 T Prisma Fit (sw version VE11C) scanner with a 64-channel head coil at the Center for Translational Imaging, Northwestern University. A navigated, multi-echo MPRAGE 3D T1-weighted sagittal volume was collected (TR/TE1/TE2/TE3 = 2170/1.69/3.55/5.41 ms, TI = 1160 ms, flip angle = 7°, FOV = 256 × 256 mm2, voxel size = 1 × 1x1 mm3, GRAPPA inplane acceleration = 2, 176 sagittal slices). A root mean square volume was generated from the 3 TE volumes to maximize image quality. A multi-shell, multi-band high-resolution DTI sequence (TR = 3000 ms, TE = 72.4 ms, flip angle = 90°, FOV = 222 × 222 mm2, voxel size = 1.5 × 1.5x1.5 mm3, multiband factor = 4, GRAPPA inplane acceleration = 2, 96 interleaved slices, phase encoding direction A > P) with 64 gradient directions per shell (b values = 1000, 2000s/mm2, and 1 non-diffusion weighted (b = 0) volume was acquired in the axial plane.

Data preprocessing

Anatomical processing. T1-weighted anatomical images were preprocessed and aligned to the anterior commissure–posterior commissure (ACPC) plane via brainlife.app.273. Following alignment, the ACPC-aligned T1w anatomical scans for each participant were segmented to generate 5-tissue type (5tt) masks via functionality provided by MRTrix3 [75], which was implemented as brainlife.app.239. The resulting 5tt masks were then used as seed masks for subsequent white matter tractography. Additionally, the aligned anatomical T1W images were used to segment and generate surfaces via Freesurfer 7.1.1’s recon-all function [76] (brainlife.app.462).

Diffusion (dMRI) data processing

Preprocessing & model fitting

Diffusion MRI (dMRI) data were preprocessed according to the protocol outlined in [77] using brainlife.app.68. The preprocessing pipeline began with denoising and removing Gibbs ringing artifacts via MRTrix3 functions before being corrected for susceptibility, motion, and eddy distortions and artifacts via FSL’s top-up and eddy functions [78, 79]. Eddy-current and motion correction were performed using FSL’s eddy_cuda8.0, incorporating the outlier slice replacement option (i.e., repol) [80–83]. To address potential misaligned gradient vectors, MRTrix3’s dwigradcheck functionality was used [84]. Further steps involved debiasing via ANT’s n4 functionality [85] and background noise removal via MRTrix3.0’s dwidenoise functionality [86]. Finally, the preprocessed dMRI images were registered to the anatomical (T1w) image via FSL epi_reg functionality [87–89]. After preprocessing, the diffusion tensor (DTI) model [90] model was fit to the preprocessed dMRI images for each subject via the Brainlife app brainlife.app.319 for DTI model fitting.

Whole-brain tractography

Whole-brain tractography was performed via anatomically constrained probabilistic tractography (ACT; [91]) as implemented in MRtrix3 via the Brainlifeio app (brainlife.app.319). Fiber orientation distributions were estimated following model fitting, and tractography was performed throughout the white matter. The tracking parameters were set to a step size of 0.2 mm, with minimum and maximum streamline lengths of 20 mm and 220 mm, respectively. The maximum curvature angle between successive steps was limited to 35°. Anatomical constraints were applied via 5TT segmentation to ensure that streamlines were biologically plausible, initiating and terminating in the gray matter and remaining within white matter pathways.

Segmentation of white matter tracts

After whole-brain tractography was performed, 61 major white matter tracts were segmented for each run via a customized version of the white matter query language [92] implemented as the Brainlife.io app brainlife.app.188. The outlier streamlines were subsequently removed via the functionality provided by Vistasoft, which was implemented as a Brainlife.io app brainlife.app.195.

Tract profile

After data cleaning, tract profiles consisting of 200 nodes were generated for all DTI measures across the 61 tracts for each participant and each test–retest condition, using functionality provided by Vistasoft implemented as the Brainlife.io app brainlife.app.361. To avoid partial-voluming effects and cleanly separate the bundle from gray matter, we removed the first and last 10 nodes from the tract profiles via brainlife.app.685. The average FA and MD values were then computed for each tract via brainlife.app.706 along the 180 nodes and used for further analysis (see Supplementary Material 1 for brainlife.app details.)

Multiple comparisons

To address multiplicity across 61 tracts, we controlled the false discovery rate using the Benjamini–Hochberg procedure at q = 0.05 within each a-priori family (Insight–FA, Insight–MD, Step-by-step–FA, Step-by-step–MD). We report both raw p and FDR-adjusted q values and label p < 0.05 & q ≥ 0.05 as “nominal.”

Omnibus test was performed across a priori tract sets; we combined tract-level two-tailed p-values using Stouffer’s Z with an empirical inter-tract correlation matrix (dependence-aware Stouffer). As a complementary, theory-driven analysis, we summarize effects in the left dorsal language network and in the Left Perisylvian Language network composites as a semi-exploratory analysis. We used equal weights, preserved effect direction via the sign of the tract-level t, and obtained a composite two-tailed p; multiplicity across composites was controlled with BH–FDR within metric.

Sample size and sensitivity

Because DTI tractography with tract profiling across 61 white-matter bundles is technically intensive (limited scanner access and long per-participant sessions), our final N = 38 reflects realistic logistical constraints for high-quality data in this design.

To quantify sensitivity, we conducted a post-hoc power analysis. With N = 38, two-tailed α = 0.05, and power = 0.80, the minimum detectable zero-order correlation is |r|≈0.44. In multiple-regression models testing a single predictor while controlling age and sex (df = 34), the minimum detectable partial correlation is |r|≈0.45.

Results

Behavioral data analysis was performed via JASP version 0.18.1 (JASP Team—2023), and the significance level was set to p < 0.05. The data were tested for normality (Kolmogorov–Smirnov test) and homogeneity of variance (Levene’s test). The data were normally distributed, and assumptions for the use of analysis of variance were not violated.

Problem solving

For 60 CRA problems, participants correctly solved an average of 23.42 problems (SD = 7.2) per person.2 Among the 60 problems administered, an average of 12.8 (SD = 5.9) problems per person were correctly solved via insight, and an average of 10.6 (SD = 10.7) problems per person were solved correctly via step-by-step analysis. Of the 60 administered, an average of 6.5 (SD = 7.2) were commission errors, 2.8 (SD = 4.2) by insights, and 3.65 (SD = 4.1) by step-by-step. These solution averages are consistent with prior findings [46, 72, 93–97].

Insight results

An overall nominal analysis of 61 tracts (for the list of tracts, see Bullock et al. [98]; Hanekamp et al. [99]) was performed; significant differences across the specific brain tracts are reported in Table 1.

Table 1.

Tracts in which insight is significantly inversely related to FA and positively related to MD within regions in the left and right hemispheres. In each linear regression, age and sex were entered as covariates

Insight				- FA				+ MD
Tract	MNI X	MNI Y	MNI Z	βstd	partial r	p	q	βstd	partial r	p	q
Left Posterior Arcuate Fasciculus (pArc)	−37	−49	+12	−0.399	−0.429	0.009*	0.341	0.315	0.331	0.048*	0.294
Right Posterior Arcuate Fasciculus (pArc)	+37	−49	+12	−0.395	−0.407	0.014*	0.341	0.084	0.087	0.615	0.952
Left Superior Longitudinal Fasciculus III (SLF III)	−3.8	−15.9	+19.9	−0.347	−0.36	0.031*	0.377	0.355	0.357	0.032*	0.276
Right Superior Longitudinal Fasciculus III (SLF III)	+3	−14	+2	−0.361	−0.38	0.022*	0.341	0.34	0.35	0.036*	0.276

Coordinates X,Y, Z are representative MNI152 waypoints for anatomical orientation; tracts span volumes

p = unadjusted two-tailed p-value; q = Benjamini–Hochberg FDR within family (q =.05) across 61 tracts (Insight–FA and Insight–MD)

*indicates q<.05 or p<.05

Left-hemisphere omnibus tests

To address multiplicity while respecting inter-tract dependence, we combined tract-wise insight associations within a priori left-hemisphere sets using Stouffer’s Z with an empirical inter-tract correlation matrix (age/sex controlled) and controlled the resulting composite tests with BH–FDR.

For FA, both the Left Dorsal (Arc + pArc + SLF III +) and Left Perisylvian Language (Arc + pArc + SLF III + IFOF + ILF) composites were significant after FDR (ps = 0.016–0.017; q = 0.041). For MD, composite effects were directionally consistent but did not survive FDR (best: p = 0.032, q = 0.081). Together, these omnibus results support a left-lateralized dorsal/perisylvian pattern linking white-matter microstructure to insight, convergent with tract-wise trends but more robust to multiple comparisons (see Table 2).

Table 2.

Rows list the tracts included in each composite with their X/Y/Z coordinates

Composite	Tracts					- FA			+ MD
		X	Y	Z	n	z	p	q (FDR)	z	p	q (FDR)
Left Dorsal	Arcuate Fasciculus (Arc)	−36	−38	+27	38	−2.415	0.016*	0.041*	2.14	0.032*	0.081
	Posterior Arcuate Fasciculus (pArc)	−37	−49	+12
	Superior Longitudinal Fasciculus III (SLF III)	−3.8	−15.9	+19.9
Left Perisylvian Language	Arcuate Fasciculus (Arc)	−36	−38	+27	38	−2.397	0.017*	0.041*	1.575	0.115	0.144
	Posterior Arcuate Fasciculus (pArc)	−37	−49	+12
	Superior Longitudinal Fasciculus III (SLF III)	−3.8	−15.9	+19.9
	Inferior fronto-occipital fasciculus (IFOF)	−53	+27	+20
	Inferior longitudinal Fasciculus (ILF)	−42	−70	−10

Coordinates are representative MNI152 waypoints for anatomical orientation; tracts span volumes. Composites statistics (n, z, p, q) are shown on the first tract row of each composite/metric

q = BH–FDR within metric across composite definitions indicates q<.05 (for FA) or p<.05 (for MD

*indicates q<.05 or p<.05

Step-by-step results

An overall nominal analysis of 61 tracts (for the list of tracts see Bullock et al. [98]; Hanekamp et al. [99]) was performed; significant differences across the specific brain tracts are reported in Table 3.

Table 3.

Tracts in which step-by-step is positively related to FA and negatively related to MD. In each linear regression, age and sex were entered as covariates

Step by Step				+ FA				- MD
Tract	MNI X	MNI Y	MNI Z	βstd	partial r	p	q	βstd	partial r	p	q
Right Parieto-Cerebellar	+6	−36	−22	−0.163	−0.166	0.332	0.787	0.389	0.398	0.016*	0.955
Forceps Minor	+2	+30	+6	0.343	0.359	0.031*	0.695	0.015	0.015	0.931	0.955
Left Vertical Occipital Fasciculus (VOF)	−26	−72	+4	0.342	0.353	0.035*	0.695	−0.34	−0.337	0.044	0.955
Left Middle Longitudinal Fasciculus Connection to the Angular Gyrus (MDLFang)	−50	−70	+30	0.333	0.341	0.042*	0.695	0.191	0.199	0.244	0.955

p = unadjusted two-tailed p-value; q = BH–FDR within family (Step-by-step–FA and Step-by-step–MD across 61 tracts. Coordinates X,Y,Z are representative MNI152 waypoints for anatomical orientation; tracts span volumes

*indicates q<.05 or p<.05

Using step-by-step as the predictor (age/sex covariates), several nominal tract-wise associations emerged, but none survived BH–FDR within family (Step-by-step –FA or Step-by-step –MD). Thus, these effects should be interpreted cautiously and are reported for completeness.

Omnibus tests

To address multiplicity while respecting inter-tract dependence, we combined tract-wise step-by-step associations within a priori left and right hemisphere sets using Stouffer’s Z with an empirical inter-tract correlation matrix (age/sex controlled) and controlled the resulting composite tests with BH–FDR. None of these tests was significant.

Lateralization analyses

To quantify hemispheric asymmetry, we computed lateralization indices (LI3 = (Left − Right)/(Left + Right)) for paired tracts (Arc, pArc, SLF III, IFOF, ILF) and for a Dorsal composite LI (Arc + pArc + SLF III), separately for FA and MD (age/sex covariates). Positive LI indicates Left > Right (for FA or MD). For insight, the Dorsal FA LI showed a nominal association (β_std = 0.323, partial r = 0.351, p = 0.036), but this effect did not survive BH–FDR within the FA LI family (q = 0.216). All other LI associations (including MD and Step-by-step) were non-significant after FDR. These analyses suggest leftward dorsal asymmetry may be present at a trend level, consistent with tract-wise and omnibus findings (Fig. 1).

Fig. 1.

A From left to right: 3D lateral projection of the arcuate fasciculus (Arc) and the left posterior arcuate fasciculus (pArc), and in green overlaid on the semitransparent MNI pial surface. Arc and pArc overlaid in directional color coding on T1-weighted images. B From left to right: 3D lateral projection of the Superior Longitudinal Fasciculus III (SLF III) in green overlaid on the semitransparent MNI pial surface. Left and right SLF III in blue overlaid in directional color coding on T1-weighted images. C From left to right: 3D lateral projection of the Inferior fronto-occipital fasciculus (IFOF) overlaid on the semitransparent MNI pial surface. Left and right IFOF overlaid in directional color coding on T1-weighted images. D From left to right: 3D lateral projection of the Inferior Longitudinal Fasciculus (ILF) overlaid on the semitransparent MNI pial surface. Left and right ILF overlaid in directional color coding on T1-weighted images. A and B correspond to the Dorsal Composite Stream; A, B, C, and D correspond to the Perisylvian Language Composite Stream. Atlas taken from [100]

Age and education

We performed an overall correlation analysis between FA, MD, age, and years of education. The results are reported in (see Table 4 in Supplementary Material 1). In line with prior studies, FA exhibited an age-related inverse correlation due to myelin degradation, axonal loss, and increased extracellular water, leading to less restricted and more random water diffusion in white matter. As the brain ages, structural integrity declines, causing water molecules to diffuse in multiple directions rather than along well-organized pathways, which reduces FA values [101]. Higher education levels are often associated with higher FAs, particularly in white matter regions related to cognitive function, such as the corpus callosum and prefrontal pathways. While FA naturally decreases with age, individuals with more education tend to show slower declines, possibly due to better-maintained brain networks, which may explain why we find a milder negative correlation between FA and education in our dataset analysis [102, 103].

Discussion

Insight problem solving is a distinct cognitive process essential for creative cognition. This study investigated the relationships between white matter microstructure and individual differences via insight and step-by-step analytical problem solving via diffusion tensor imaging (DTI). Our findings reveal distinct white matter correlates for Aha! Moments, providing novel evidence for the structural basis of these cognitive processes. While prior research has focused primarily on functional imaging, EEG, and brain stimulation [1, 2, 17, 47, 48, 104], our results extend this understanding by revealing structural foundations supporting insight into problem solving. This bridges the gap between functional and structural results on idea generation via insight, extending our understanding on creative cognition.

Insight problem-solving findings

The results obtained using the left-hemisphere omnibus tests indicate that insight propensity correlates with lower FA in the left dorsal tracts’ composite. This cluster of tracts includes a) the left posterior arcuate fasciculus, which connects the left superior temporal gyrus (encompassing the auditory cortex, angular gyrus, and Wernicke’s area) to parietal regions. This pathway supports language processing, semantic integration, and phonological working memory—functions that may compete with the cognitive flexibility required for insight ([105]; b) The left arcuate fasciculus serves as a major white matter pathway bridging Broca’s area in the frontal lobe and Wernicke’s area in the temporal lobe, supporting key language processes such as speech production, word learning, and speech comprehension ([104, 106–109]; c) the left superior longitudinal fasciculus III (SLF III). This tract links the temporoparietal junction (intersecting Wernicke’s area) with the inferior frontal gyrus (IFG, Broca’s area in the left hemisphere). Notably, the left SLF III supports speech processing, whereas the right SLF III facilitates visuospatial functions [110].

Together, the finding that lower FA in the left dorsal pathway suggests that integrated language-related networks may inhibit insight. This aligns with findings from Shamay-Tsoory et al. [111], who reported that patients with left temporoparietal and inferior frontal lesions presented greater originality on divergent thinking tasks, implying a "releasing effect" on creativity. Similar effects have been observed in stroke [112] and frontotemporal dementia patients [113–115], where lesions in the left frontotemporal regions have been shown to increase artistic creativity. Taken together, these findings imply that left hemispheric regions play a regulatory role in creativity and that their disruption lifts this constraint, thus promoting novel ideas.

Research on hemispheric differences in semantic processing further supports this interpretation. Beeman and colleagues [116] proposed that the left hemisphere engages in fine semantic coding—generating focused, context-specific semantic fields—while the right hemisphere performs coarse coding, integrating broader, loosely connected concepts [59]. This broad integration facilitates unconventional associations that are critical for insight [2]. The sudden emergence of insight likely reflects a buildup of weakly activated solution-related information, reaching a threshold before bursting into consciousness — the classic Aha! Moment. Thus, reduced left-hemisphere structural connectivity supports theories suggesting that insight thrives on less constrained semantic processing and distant information integration. FA variations in these pathways may influence alpha-band oscillatory dynamics, modulating selective inhibition mechanisms crucial for insight. This aligns with prior evidence linking alpha waves to insight’s unconscious buildup phase, where suppressed dominant strategies allow novel connections between concepts to emerge in memory [46]. FA, a marker of white matter integrity, may thus explain individual differences in insight performance, linking structural properties to cognitive flexibility and creative ideation.

The results obtained on the omnibus test for the left perisylvian language lead in the same direction. This tract’s composite includes the arcuate fasciculus, posterior arcuate fasciculus, SLF III, IFOF, and ILF, forming a network that is also critical for language processing, semantic integration, and multimodal association. The inclusion of the IFOF and ILF extends the prior findings beyond the dorsal language system, including ventral pathways that connect occipital and temporal regions with frontal and anterior temporal semantic hubs. The IFOF, indeed, has been found to be involved in top-down modulation of perceptual and linguistic information [117, 118], reduced FA in this tract may reflect a loosening of top-down constraints, enabling a broader and probably more flexible semantic search space conducive to insight. Similarly, ILF links occipital regions with anterior temporal areas, and has been found to be crucial for semantic memory and concept integration. [119, 120], reduced FA in this tract may promote the retrieval of more remote or weakly activated associations, consistent with the conceptual restructuring characteristic of Aha! Moments. Although the MD effects for this composite were directionally consistent with FA findings, they did not survive FDR correction, warranting cautious interpretation. Nevertheless, the convergence of FA findings across both the dorsal and perisylvian composites suggests that left-hemisphere language-related white matter pathways—including both dorsal phonological/working memory circuits and ventral semantic-associative pathways—play a modulatory role in insight problem solving. These results are in line with models positing that reduced left-hemisphere connectivity facilitates a shift toward right-hemisphere coarse semantic coding [60], ultimately enabling the emergence of novel solutions.

Step-by-step problem-solving findings

In contrast to insight, our omnibus test did not yield any FDR-significant associations between step-by-step analytical problem solving and white matter microstructure. Several nominal tract-wise effects emerged, including positive associations between FA in the forceps minor, left vertical occipital fasciculus (VOF), and left middle longitudinal fasciculus connection to the angular gyrus, as well as a negative association between MD in the right parieto-cerebellar tract. However, all of these effects failed to survive BH–FDR correction and should therefore be interpreted cautiously.

Although exploratory, the pattern of nominal effects is consistent with the notion that step-by-step analytical problem solving relies more on occipital and interhemispheric connectivity supporting executive control, working memory, and top-down attention [37, 38, 121–123]. The forceps minor is the key pathway for frontal interhemispheric communication, and higher FA in these tracts may facilitate the controlled, deliberative integration of information characteristic of analytical reasoning. The left VOF, which links dorsal and ventral visual streams, has been implicated in grapheme–phoneme conversion and lexical access [124, 125], stronger connectivity here could support the stepwise mapping of perceptual and semantic features required by the CRA task.

Nevertheless, these associations did not meet corrected significance thresholds; therefore, they must be considered only as preliminary and hypothesis-generating rather than conclusive. It is possible that structural correlates of step-by-step problem solving are more complex and thus require either a larger sample or targeted region-of-interest approaches to detect reliably. Alternatively, step-by-step problem solving may depend more on dynamic functional recruitment of cognitive control networks than on stable individual differences in white matter microstructure. Therefore, future studies, using multimodal designs combining DTI with functional connectivity analyses, could help clarify the structural–functional interplay underlying analytical problem solving.

Lateralization analysis

Our lateralization index (LI) analyses revealed a nominally significant association between insight propensity and leftward FA asymmetry in the dorsal composite (β_std = 0.323, p = 0.036), however this effect did not survive FDR correction (q = 0.216). This trend is consistent with our omnibus and tract-wise findings, which point toward a left-lateralized dorsal and perisylvian substrate associated with insight. The observation that greater insight was linked to relatively lower FA in left-hemisphere tracts (reflected in a positive LI) supports prior models proposing that reduced left-hemisphere dominance facilitates the broader, right-hemisphere–driven semantic search processes underlying insight [60, 126]. Future work could examine whether such structural asymmetries correspond to lateralized functional activation patterns during insight tasks.

We are aware that although these lateralization effects must be interpreted cautiously given their lack of FDR significance, they, however, converge with the main FA findings to suggest that hemispheric balance — and potentially the relative "release" from left-hemisphere language control — may play a role in enabling the sudden restructuring of problem representations characteristic of Aha! Moments. Future studies with larger samples could help clarify whether such lateralization effects are reliable and whether they interact with individual differences in cognitive style or functional activation patterns during insight problem solving.

Comparison with previous DTI studies on creativity

Our findings partially align with, yet also diverge from, previous DTI studies on creative cognition. For example, prior studies [33, 35] reported lower levels of FA within the left inferior frontal tract (overlapping the uncinate fasciculus and anterior thalamic radiation) and divergent thinking (measured by a creative composite index). In our study, we found a similar inverse relationship between FA and insight within the SLF III that projects to the left inferior frontal lobe, corroborating the idea that left IFG damage may produce a "releasing effect" on creativity, allowing for more novel and unconventional ideas that might rise as sudden insight. This relationship thus points to an overlap between divergent thinking and insight rather than analytical problem solving. Nonetheless, insightfulness is not usually measured in divergent thinking tasks, leaving this last conclusion thus far just a speculation. In contrast, Takeuchi and colleagues [37–39] reported positive correlations between FA and both divergent and convergent thinking abilities across several bilateral white matter tracts (convergent thinking—left ILF and left frontal‒occipital fasciculus, divergent thinking – corpus callosum, the bilateral basal ganglia, the bilateral temporoparietal junction, and the right inferior parietal lobe). These findings broadly align with our observation that analytical problem solving is positively associated with FA across the VOF. As such, it is possible that the findings of Takeuchi and colleagues [39] were driven by a higher rate of problems being solved via step-by-step analysis. However, the work of Takeuchi and colleagues [39] did not assess how participants generated ideas (via insight or step-by-step analysis), thus leaving this question unanswered.

Past studies linking FA and creative cognition have only measured task performance in terms of either the originality of ideas (in divergent thinking) or the number of problems that were solved correctly (in convergent thinking tasks). Insight problem solving can potentially occur not only in convergent tasks such as the CRA but also during divergent thinking; however, this topic has rarely been investigated in the literature. Consequently, given the general lack of research on the neuroscience of insight for divergent thinking tasks, it remains unclear whether our conclusions can be directly extended to divergent thinking performance.

In summary, our study addresses a significant gap in the literature by directly comparing the structural correlates of insight and analytical problem solving. The present investigation enables us to disentangle the specific neural architecture supporting the distinct cognitive processes of insight and step-by-step analysis, a differentiation that was not addressed in past studies. Our findings support and extend previous research on the neural basis of creative cognition. The lower FA associated with insight in several tracts is consistent with some previous studies on divergent thinking [33, 35], suggesting potential overlaps between insight and broader aspects of creative cognition.

The discrepancies between our findings and those of previous studies underscore the complexity of creative cognition and the importance of specificity and standardized methodological approaches when investigating idea generation in neuroimaging research. As such, future studies should continue to refine the operational definitions and measurements of unique creative processes to better understand their unique and shared neural substrates.

Conclusion

Our study provides novel evidence for distinct structural connectivity patterns associated with insight and analytical problem solving. The findings suggest that insight is associated with lower FA across left-hemisphere dorsal and perisylvian language tracts, with convergent but subthreshold trends observed for MD. These results complement the established literature documenting relationships between FA and cognitive functions [19, 28]. Specifically, lower FA across left hemisphere tracts reflects a more diffuse connectivity pattern that allows for broader semantic activation and the cognitive flexibility necessary for insight [2, 126]. The involvement of the SLF III, which connects frontal, temporal, and parietal regions [107, 127], suggests that insight may rely on more distributed neural activation rather than highly focused connections [1, 128]. This pattern of structural connectivity aligns with functional neuroimaging studies that have highlighted the importance of widespread activation in insight problem solving [1].

In contrast, we did not find any FDR-significant associations between step-by-step analytical problem solving and white matter microstructure. Several nominal tract-wise effects suggested a positive relationship between FA in the forceps minor, left vertical occipital fasciculus, and left middle longitudinal fasciculus connection to the angular gyrus, but these findings should be interpreted as preliminary.

The dissociation between insight and analytical problem solving is particularly noteworthy, highlighting the distinct neural architectures supporting these two modes of problem solving. These results contribute to a more nuanced understanding of the structural basis underlying different aspects of idea generation in creative cognition. Our findings pave the way for future investigations into how variations in white matter microstructure may influence an individual's propensity for insight versus analytical thinking and how these structural differences relate to functional activation patterns observed during creative cognition.

Limitations and future directions

While our study provides valuable insights into the structural correlates of insight and step-by-step analytical problem solving, some limitations should be noted. First, the cross-sectional nature of our study precludes causal inferences about the relationship between white matter structure and problem-solving ability. Longitudinal studies or training interventions could help clarify the directionality of these relationships. Similarly, integrating DTI data with functional neuroimaging and electrophysiological measures would offer a more complete understanding of how structural connectivity supports the dynamic neural processes involved in insight and analytical problem solving. As such, future work should extend the present study by employing a more multimodal methodology, thus providing a more complete picture of the neuroscience of insight.

Notably, although we controlled for age and sex in our analysis, the demographic characteristics of our sample, such as socioeconomic status and education, might have affected the observed results. Indeed, past work indicated that negative correlations between FA and divergent thinking across several tracts were observed only for female participants, whereas males exhibited the opposite effect [34, 129]. This work outlines the importance of considering sample characteristics such as gender when investigating links between individual differences in creative cognition and patterns in white matter integrity. Nevertheless, it is worth noting that the present sample was roughly balanced between male and female participants. While few studies to date have reported interactions between creative abilities and age, socioeconomic status, or educational level, all of these variables have been noted to influence white matter integrity [102, 130–133].

Our sample (N = 38), typical for technically intensive DTI protocols, may have limited sensitivity to small effects, especially in tract-wise analyses. We therefore controlled the false-discovery rate and emphasized convergent FA/MD patterns rather than isolated nominal findings. We believe that future work with larger cohorts and preregistered ROI-focused tests will be well-positioned to corroborate the present pattern and to achieve adequate power under stricter multiplicity control.

Notably, our findings may not apply to all other convergent thinking tasks. Indeed, solving anagrams via insight has been shown to elicit left-lateralized brain activity during insight [134]. This finding has been attributed to the unique task demands of completing anagrams compared with other insight problems, such as the CRA. While anagram completion relies heavily on grapheme feature integration, the CRA uniquely requires the integration of semantically distant concepts. Our decision to employ the CRA in the present study was guided by current best practices in insight research and in coherence with prior imaging studies (i.e., Jung-Bemman et al. [135]).

The study of insight has undergone a significant methodological shift, transitioning from classic problems [136, 137] to newer paradigms such as the CRA [44, 138]. This change was driven by the need for increased statistical power and neuroimaging compatibility, which is afforded by the CRA [11, 47, 72, 139]. Traditional insight problems, such as nine-dot or two-string problems [136, 137], while directly tapping into insight processing, present limitations in terms of reliability and adaptability to neuroscientific methods. However, this methodological shift has led to a continued focus on convergent thinking tasks to measure insight, potentially overlooking how insightful ideas emerge during divergent thinking processes. Insight, which is fundamentally about idea generation, may manifest in both convergent and divergent thinking scenarios, suggesting the need for a more comprehensive approach to studying this phenomenon. As such, future work is needed to determine whether the present findings extend to insights into other convergent thinking tasks or even divergent thinking studies.

Abbreviations

ACT

Anatomically constrained probabilistic tractography

ACPC

Anterior commissure–posterior commissure

AUT

Alternate uses task

Arc

Arcuate fasciculus

CRA

Compound remote associates

CC

Corpus callosum

dMRI

Diffusion MRI

DTI

Diffusion tensor imaging

EEG

Electroencephalogram

FA

Fractional anisotropy

FOF

Frontal–occipital fasciculus

fMRI

Functional magnetic resonance imaging

IFOF

Inferior fronto-occipital fasciculus

ILF

Inferior longitudinal fasciculus

JRAT

Japanese variant of the CRA

MD

Mean diffusivity

MDLFang

Middle longitudinal fasciculus connection to the angular gyrus

pArc

Posterior arcuate fasciculus

SLF III

Superior longitudinal fasciculus III

STG

Superior temporal gyrus

TPJs

Temporoparietal junctions

VOF

Vertical occipital fasciculus

Authors’ contributions

Carola Salvi conceived the study, collected and analyzed the data, and led the writing of the manuscript. Simone A. Luchini contributed to the writing of the manuscript and data interpretation. Franco Pestilli developed the diffusion tensor imaging (DTI) pipeline and assisted with data interpretation. Sandra Hanekamp was responsible for data preprocessing. Todd Parrish contributed to experiment design, provided imaging support, and assisted with data collection. Mark Beeman and Jordan Grafman contributed to study ideation and data collection.

Funding

This research was supported by the United States Air Force Research Laboratory FA8650-15-2-5518 to MB and by the Smart Family Foundation of New York to JG. CS was supported in part by NIH training grant T32 NS047987.

Data availability

The datasets generated and/or analyzed during the current study are available from Brainlife.app. (see Supplementary Material 1).

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the ethical standards outlined in the Declaration of Helsinki. Ethical approval was obtained from the Northwestern University Institutional Review Board (IRB #: STU00202210-MOD0009 approved by NU IRB), and informed consent was obtained from all participants prior to their inclusion in the study.

All participants provided written informed consent prior to participation in the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.