Narrative ‘twist’ shifts within-individual neural representations of dissociable story features

Abstract

Given the same input, one’s understanding of that input can differ based on contextual knowledge. Where and how does the brain represent latent mental frameworks that interact with incoming sensory information to shape subjective interpretations? In this study, participants listened to the same auditory narrative twice; the narrative had a plot twist in the middle that dramatically shifted interpretations of the story. Using a robust within-subject whole-brain approach, we leveraged shifts in neural activity between the two listens to identify where latent representations are updated and, by extension, where interpretations are instantiated in the brain. We considered the story in terms of the broader narrative model and two specific components, episodes and characters. Neural activity patterns varied with participants’ latent understanding of these three elements in overlapping but partially distinct sets of temporal, parietal, and prefrontal regions. Results suggest that even when the sensory information and the individual are held constant, heteromodal cortex represents individual narrative elements according not to their surface features, but to latent conceptual frameworks for understanding and interpreting narrative information.

Introduction

Identical sensory inputs can evoke different interpretations. Rather than being fully predictable from properties of the information itself (i.e., “stimulus-computable”), our interpretations–how we understand or make meaning of external information–are flexibly shaped by how that information interacts with our internal expectations, prior knowledge, and mental state. This is particularly evident for high-level, complex narrative stimuli.

Differences in brain activity across participants to the same external input (e.g., movies, auditory narratives, animations) have frequently been used to index different internal experiences of that input. In these studies, variability in neural activity across participants is often attributed to differences in interpretation, which arise from various sources including experimentally-imposed or endogenously-generated contexts and beliefs^1–6, life experiences^7,8, and stable personality traits^9,10. While across-subject analyses are informative, they can be confounded by idiosyncratic factors such as unmeasured traits and states, media preferences and past exposure, and even functional brain anatomy, making it challenging to fully attribute differences in brain activity to differences in interpretation. A within-subject approach, which compares the same individual to themselves before and after an update in interpretation, inherently controls for these factors and enables stronger inferences about neural representations of narrative interpretations.

Further, most neuroimaging studies treat narratives—and their corresponding interpretations—as monolithic entities. However, they consist of subcomponents such as characters and episodes^11,12, that dynamically interact with the broader narrative understanding. When these elements update (e.g., revelations about character motivations or unexpected events), they reshape the narrative model and overall interpretation. Simultaneously, the narrative model tracking the ‘gist’ of a story (sometimes classified as a situation model^13,14) provides the scaffold for interpreting these subcomponents. These elements likely operate on different timescales and engage distinct brain regions across the cortex^15,16. Therefore, while prior work has tended to focus on a priori chosen networks or brain regions, such as the default mode network (DMN)¹² and the medial prefrontal cortex (mPFC)¹⁷, understanding how different narrative elements are represented necessitates a comprehensive whole-brain approach.

Here, we aimed to identify where the brain represents distinct aspects of narrative interpretations. We used a unique narrative stimulus that contained a major twist halfway through that prompted participants to substantially shift their interpretations of the events preceding the twist. Participants then listened to the narrative a second time with this updated interpretation. Importantly, we held both the participant and the stimulus constant, enabling us to leverage within-subject shifts in neural activity between the first and second listen to understand how and where latent interpretive frameworks, independent of external sensory input, are reflected in the brain. Furthermore, by taking a whole-brain approach, we found evidence that different narrative elements—i.e., narrative models, episodes, and characters—are represented in somewhat distinct sets of brain regions. This work highlights how heteromodal cortex goes beyond sensory features to represent narrative elements according to latent conceptual frameworks for understanding and interpreting narrative information.

Results

Our overarching goal was to identify where and how the brain represents interpretations of narratives and their subcomponents using a within-subject approach. Thirty-six healthy adults listened to an auditory narrative twice in a row during functional magnetic resonance imaging (fMRI) scanning. The narrative featured a twist in the middle that recontextualized the earlier segments of the story. Initially perceived as a straightforward dialogue between a curmudgeonly dress-shopper (Steve) and a friendly, if pushy, shopkeeper (Lucy), the story later reveals a radically different reality: Steve is struggling to survive an apocalypse, and Lucy is a robot undermining his survival (See Methods section Stimulus description for further detail).

The dramatic shift induced by the twist required listeners to update their understanding of the broader narrative model, reevaluate specific episodes, and reassess the characters in light of the new context. We captured within-subject “shifts” – defined as between-listen changes in neural representations – that reflect updates to each narrative element to identify where each is represented in the brain.

Representations of the narrative model.

We first investigated where the overall narrative model—i.e., the broader gist of the narrative (in this case, “normal day at the mall” or “robot apocalypse”)—is represented. To this end, we compared within-subject neural and behavioral responses between the two listens. Given the twist in the middle, we split the narrative into three segments: pre-twist, twist, and post-twist. The twist changes the interpretation of everything that came before it, prompting participants to shift to a new narrative model that persists during the remainder of the first listen and throughout the second listen. As a result, the pre-twist segment, which is interpreted under a different model in the first (L1) versus second listen (L2), should be processed most differently between listens. In turn, the post-twist segment, which is processed with the same model across both listens, should exhibit more consistent neural and behavioral patterns (Fig. 1A). To identify where narrative models are represented, we therefore compared within-subject neural and behavioral shifts in each segment between listens (pre-twist_L1-L2 to post-twist_L1-L2), expecting greater shifts in the pre-twist segment than in the post-twist segment in both patterns of neural activity and behavioral ratings as a result of reinterpretation.

Figure 1. Neural and behavioral shifts reflect narrative model updating.

A. Hypothesized differences between segments. The same individuals listened to an auditory narrative two times. The narrative was divided into three segments: 1. pre-twist, 2. twist and 3. post-twist. Novel, recontextualizing information is learned during the ‘twist’ segment, inducing a shift in interpretation. This new interpretation (narrative model 2) is carried into the post-twist segment on the first listen and into the entirety of the second listen. Thus, greater neural shifts are expected in the pre-twist as compared to the post-twist segment. After each listen, participants were tasked with reporting the specific moments (episodes) that they reevaluated in light of the twist (see Fig. 2). B. Computing neural shifts. For each participant, neural shifts between listens were computed per region per timepoint as one minus the correlation between the multivoxel spatial patterns of activity in Listen 1 and Listen 2 (pattern intra-SC). C. Greater behavioral shifts in the pre-twist segment. Within-subject shifts between listens in behavioral ratings (continuous real-time reports of character impression) were greater in the pre-twist compared to the post-twist segment (paired t-test, *** indicating p < 0.001). D. Greater neural shifts in the pre-twist segment. The median pre-twist and post-twist neural shift value was taken for each participant and compared using a linear mixed effects model per region. Estimates plotted reflect the difference between the pre- and post-twist segments (set up as pre-twist > post-twist). Regions contoured in black show a significant effect at q < 0.05 (corrected for multiple comparisons using the false-discovery rate) for all matched-length sample comparisons between segments (see Methods). Regions contoured in blue show an effect at p < 0.05 (uncorrected) for all matched-length sample comparisons. E. Greater neural shifts accompany behavioral shifts. For each participant, we binarized timepoints into those with a behavioral shift (absolute difference in ratings between Listen 1 and Listen 2 > 0) and those without a behavioral shift (absolute difference in ratings equal to 0), then compared neural shifts between these two groups of timepoints. The observed median difference in neural shifts (behavioral shift “present” moments minus behavioral shift “absent” moments) across participants is plotted. Blue contours indicate regions showing significant relationships between neural and behavioral shifts (p < 0.05, uncorrected, determined via block permutations, n = 10,000).

Greater behavioral shifts in the pre-twist segment.

During both listens, participants reported their real-time impression (negative to positive) of the shopkeeper (“Lucy”) in a continuous rating task. In these behavioral ratings, impressions of Lucy generally moved from positive to negative across the story, reflecting the evolving understandings of the broader situation and indicating a transition from viewing Lucy as a store clerk dealing with a difficult customer to perceiving her as a robot with Steve struggling to survive.

As hypothesized, ratings shifted more between listens in the pre-twist compared to the post-twist segment, indicating a greater change in how participants perceived the situation (“behavioral shift”; paired t-test, t(34)=5.31, p < 0.001, 95% CI [0.3091, 0.7006]; Fig. 1C). Behavioral shifts were calculated as one minus the intra-subject correlation between each participant’s behavioral timeseries from the continuous rating task in Listen 1 and Listen 2.

Greater neural shifts in the pre-twist segment.

We operationalized neural representations as the multivoxel pattern of activity in each region at each timepoint. At each matched timepoint in Listen and Listen 2, we computed the within-subject correlation of these patterns (“pattern intra-subject correlation” (pattern intra-SC)^18,19) and calculated “neural shifts” as one minus this correlation (henceforth “intra-subject pattern distance”). As hypothesized (Fig. 1A), the intra-subject pattern distance was higher in the pre-twist segment than in the post-twist segment across the cortex, indicating greater neural shifts in response to the updated information (main effect of segment: estimate = 0.01, p < 0.001; whole-brain linear mixed effects model (LMEM) with region and participant as random effects). The regions that showed the strongest differences, suggesting a strong role in maintaining and updating the narrative model, included the left hippocampus, the angular gyrus, temporal parietal junction (TPJ), dorsomedial prefrontal cortex (dmPFC), and the bilateral posterior medial cortex (PMC)/precuneus (one LMEM per region with participant as a random effect; Fig. 1D). These regions align with findings from across-subject studies on contextual modulation of representations of situation models and schemas^20–23 and studies of interpretational shifts during auditory narrative processing^3,24.

Notably, we did not see significant neural shifts between listens in primary auditory cortex. This was expected given that the low-level sensory properties of the stimulus are identical across listens and also helps to mitigate concerns that participants may have simply been paying less attention during the second listen. Some effects, albeit weaker than those in multimodal association regions, were also seen in early and middle visual regions (e.g., V1, MT). These effects are likely due at least in part to differences in how participants were looking at the screen to report or consider reporting the changes in the continuous rating task; in both listens, there were more slider movements (rating adjustments) in the pre-twist segment than in the post-twist segment (t(34) > 5.63, p < 0.001 for both listens).

We ran a series of control analyses to ensure the robustness of our findings. We first aimed to rule out the possibility that differences in brain activity across listens were driven by participants’ movements on the continuous rating task. To address this, we regressed the movement of the slider from each participant’s neural timeseries in each listen, repeated the analyses on the residuals of this regression, and found that the results were largely unchanged (Supplementary Fig. 1A).

Next, we sought to dissociate the effects of our stimulus’ specialized model structure from the effects of simply re-listening to the same information. First, one may expect that given that participants have already heard this narrative once, they may become less interested on the second listen. However, both our hypothesis and our observed results work against expected attention or “boredom” effects: if participants were simply mind-wandering more as time went on during the second listen, we would expect to see greater shifts post-twist compared to pre-twist due to decreased engagement and more off-task (as opposed to stimulus-driven) activity. A second possibility is that regardless of any model updating, participants simply become more synchronized to themselves over time when relistening to the same narrative, which could also explain our pre- versus post-twist differences. We performed two analyses to help rule out this explanation. First, we turned to an independent dataset²⁵ where the same participants listened to an auditory story (from The Moth) multiple times, of which we used the first two listens (Supplementary Fig. 1B). Critically, this story did not contain a twist or any other feature that would induce a drastic model update akin to our stimulus. Encouragingly, we found (at a liberal, uncorrected threshold of p < 0.05) that only three regions showed a linear effect of time on pattern intra-SC: dorsolateral PFC and the bilateral auditory cortex. The latter region (auditory cortex) actually showed a decrease over time, potentially indicating reduced attention. Second, in our dataset, we tested for any linear effects of time within segments (pre/post-twist) that could inflate our findings. We split the pre- and post-twist segments into an early and late period and identified where, regardless of segment (pre- versus post-twist), there were greater pattern intra-SC values in the late as compared to the early periods (Supplementary Fig. 1C). The strongest early versus late effects emerged along the left temporal pole into the STS (q < 0.05), which notably did not show segment effects in our main analysis (see Fig. 1D). While canonical DMN regions including the left TPJ, PMC and mPFC also showed some differences in the direction of greater consistency for the late versus early period (uncorrected p < 0.05), effects were overall weaker than those in our main analysis. Results from this combination of analyses further strengthens the likelihood that our observed pre- versus post-twist differences were, in fact, driven largely by the narrative model updates induced by the twist in our stimulus, rather than simpler phenomena inherent to listening to the same stimulus a second time more generally.

Regions show greater neural shifts when individuals report behavioral shifts.

Our first two analyses showed that, as hypothesized, both neural and behavioral shifts are greater in the pre-twist than the post-twist segment, likely reflecting narrative model updates that changed how this segment was interpreted overall. In a follow-up analysis, we sought to detect a more fine-grained relationship between these two types of shifts: moment-to-moment, do greater neural shifts track with greater behavioral shifts? Towards this goal, we first binarized individual participants’ behavioral timeseries into timepoints where a shift was present (absolute difference in behavioral rating between the first and second listen > 0) or absent (difference in behavioral rating between listens = 0). We then compared neural shifts (intra-subject pattern distances) between these two sets of timepoints (see Methods section Linking moments of neural and behavioral shift), hypothesizing greater neural shifts at timepoints where a behavioral shift was also present.

Participants generally differed in how faithfully they complied with this behavioral task (see Methods for more information), limiting our power for this analysis. Although no regions withstood FDR correction, those that showed the strongest effects in the hypothesized direction were the bilateral precuneus/PMC, right TPJ, and the right medial PFC (p < 0.05), dovetailing with past work that implicates these regions in active contextual updating (Fig. 1D). Furthermore, regions that showed an effect in this analysis also showed stronger effects in the narrative model analysis (r = 0.27, p < 0.01; correlation of estimates across regions between analyses). This suggests that, as hypothesized, the regions showing greater model updating (pre- versus post-twist contrast) were also involved in tracking the changed perceptions of Lucy throughout the stimulus (compare Fig. 1D and 1E). Importantly, early visual regions did not show effects despite likely between-listen differences in task-induced eye movements towards the slider, suggesting that we are capturing higher-level cognitive mechanisms (e.g., model updating) that operate at a more abstract level than simple visual or motor behavior.

Representations of episodes.

Having detected evidence for representations of a coarse-level narrative model in certain brain regions, we next investigated if and where the brain represents interpretations of smaller units of a narrative, namely specific episodes²⁶. Here we defined episodes as punctate events with a clear beginning, middle, and end, that drove the plot forward.

After each listen, we prompted participants to identify specific moments that they reevaluated or reinterpreted in light of the twist (see Methods section Experimental procedures for more information). All episodes occurred within the pre-twist segment, aligning with the neural and behavioral results that suggested greater interpretation updating during this segment (Fig. 1).

We hypothesized that, over and above the generally greater neural shifts in the pre-twist relative to post-twist segment, neural shifts would be even more exaggerated specifically during the episodes that participants reported reevaluating between listens. To test this hypothesis, we selected five episodes that were reevaluated by the majority of participants and chose five control episodes of matched length that were also in the pre-twist segment that most participants did not report reevaluating (shown in Fig. 2A). Then, for each participant, we modeled patterns of brain activity during each individual episode in each listen using an event-related general linear model (GLM) and used the extracted episode-wise betas to compute a neural shift (intra-subject pattern distance between listens, Fig. 2B; see Methods section Computing shifts in the reevaluated episodes).

Figure 2. Episodes that are reevaluated show greater neural shifts between listens.

A. Identifying reevaluated episodes and matched controls. Using behavioral data from inside and outside the scanner, we selected the top five episodes that participants most commonly reported reevaluating in light of the information introduced by the twist and paired each one with a matched control episode that was nearby in the narrative and the same length, but not reported as reevaluated by most participants. All reevaluated and control episodes were within the pre-twist segment. We plot the temporal location and duration of each group-defined episode atop a raster-style depiction of individual participants’ behavioral reports (black lines correspond to moments that participants reported reevaluating; gray lines correspond to the first TRs which were removed to avoid initial transient effects). The five reevaluated and control episode pairs are highlighted and labeled. Subscripts: r, reevaluated; c, control. B. Computing neural shifts between the reevaluated and control episodes. For each participant, we used an event-related GLM to model each individual episode in each listen, then computed neural shifts as one minus the correlation between the spatial pattern of beta values in Listen 1 and Listen 2 (one value per region per episode). Greater neural shifts were hypothesized for the reevaluated, as opposed to the control, episodes. C. Reevaluated episodes show greater neural shifts. Plotted estimates show the strength of the difference between reevaluated and null episodes within participants. (Estimates reflect output from a linear mixed effects model in which within-subject neural shifts were predicted by episode type (set up as reevaluated > control), using participant and episode pair as a random effect.) Regions contoured in black show a significant effect at q < 0.05 (corrected for multiple comparisons using the false-discovery rate). Blue contours reflect a relationship thresholded at p < 0.05 (uncorrected). The distributions of neural shifts within the superior temporal sulcus are plotted in the inset. Dots represent participants’ median neural shifts across episodes of each type (reevaluated and control).

By comparing neural shifts between reevaluated and control episodes, we found evidence that interpretations of episodes are represented along the bilateral superior temporal lobes, in the left TPJ, and, at an uncorrected threshold, in the left superior frontal cortex (Fig. 2C; LMEM per region predicting the difference in the neural shift between reevaluated and control episodes, treating both participant and episode pair [reevaluated, matched control] as random effects). These findings align with related work identifying the left anterior middle temporal gyrus and the TPJ (among other DMN regions) as supporting ‘aha’ moments^27,28.

Compared to the broader narrative model, representations of specific episodes appeared to be represented in distinct, more left-lateralized temporal regions (compare Fig. 1D to Fig. 2C), which may be due to the left lateralization of language-mediated semantic representations^29–32. The one region that showed strong effects in both episode and narrative model representations was the left TPJ, a region suggested to be involved in binding of external information and managing competing beliefs^33,34.

Representations of Characters.

Having demonstrated that narrative models and models of specific episodes are represented in largely distinct brain regions, we next examined representations of characters. Characters generally link episodic details to the narrative model by embodying the motivations and goals that influence the narrative’s progression. We investigated how representations of Lucy (from shopkeeper to robot) and Steve (from annoying shopper to persistent survivor) are constructed and updated across the two listens. (See Methods section Stimulus description for further detail on the characters).

We expected that by the end of Listen 1, participants would have converged on a final interpretation of the two characters and that they would then “reload” this interpretation into memory at the start of Listen 2. We operationalized these assumptions into predictions about what neural activity patterns should look like in regions tracking latent representations of each character.

To this end, we split the narrative into “Lucy” or “Steve” conversational turns based on speaking onset and offset times and modeled each turn in each listen using an event-related GLM. We then designated a per-participant, per-region “template” neural representation of each character in Listen 1, when participants had all the information necessary to fully interpret (represent) their identity in light of the updated narrative model. Specifically, template representations were defined by taking the median activity in each voxel across the final three events (conversational turns) of Listen 1 for a given character.

We then correlated representations of each character at each turn in Listen 1 and Listen 2 with their corresponding template, yielding a series of correlations per region showing how character representations evolve toward the template in each listen (Fig. 3A). We considered a region as representing a character if it showed the following properties: (1) a steady increase over time in similarity between the character’s events and the template over Listen 1, (2) a reloaded template-like representation at the start of Listen 2, (3) a stabilization in representation (i.e., flatter slope toward the template) over the course of Listen 2, and (4) a dissociable representation of the other character (i.e., lower or negative correlations with the template that stay flat or decrease over time) in both listens (see Fig. 3B for a schematic of these criteria). For more information on these criteria and how they were tested, see Methods section “Computing updates in the representations of characters”.

Figure 3. Character representations are updated in light of the twist.

A. Computing representations of characters. The dialogue was split into ‘Lucy’ and ‘Steve’ events based on speaking onset and offset times. We designated a per-participant, per-region ‘template’ representation of each character based on the multivoxel activity pattern during their last three conversational turns in Listen 1 (L1). Multivoxel patterns during each event in each listen were then correlated with the template to test for a series of criteria (see panel B). B. Schematic of criteria. Red lines: Pattern similarity (correlations) between same-character events and the corresponding template (i.e., Lucy events to Lucy template; Steve events to Steve template) were hypothesized to be positive and to increase progressively over the course of the story (positive slope). In Listen 2, they were expected to start higher and exhibit a weaker slope compared to Listen 1, reflecting the “loading” of the character’s representation from the end of Listen 1. Gray lines: Pattern similarity (correlations) between opposite-character events and each template (i.e., Steve events to Lucy template and vice versa) were hypothesized to be non-existent or, if anything, to show a negative slope over time (as representations of the characters diverged). C. Regions track character representations. Estimates reflect the magnitude of the effect for our main criterion, which was that the similarity between a character’s events and their template should show a steeper slope over time in Listen 1 than Listen 2 (computed with a linear mixed effects model predicting event-template correlations from an interaction between listen and event number with a random effect of participant). Regions plotted meet all of our criteria at an uncorrected threshold of p < 0.05 (see B; Methods). Black contours indicate regions that show a significant effect at q < 0.05 (following correction for multiple comparisons using the false-discovery rate) for the main criterion. Regions that show an effect in the expected direction for all criteria are shown in Supplementary Fig. 2A.

We focused on regions that showed effects in the expected direction for all criteria at an uncorrected threshold and that exhibited a significantly steeper slope over time (i.e., a sharper evolution toward the template) in Listen 1 relative to Listen 2 at a corrected threshold (q_FDR < 0.05; Fig. 3C), which we considered the most important index for a region representing latent character interpretations. Regions meeting both of these standards emerged bilaterally along the STS into the temporal pole, the TPJ, angular gyrus and ventromedial PFC for both characters^22,35–37. Supplementary Fig. 2A shows the regions with effects in the expected direction for all criteria (but not necessarily that withstand correction for significance in our primary criterion, criterion 3), which includes most of the cortex, providing a reassuring sanity check (since effects in the opposite direction for any of the criteria would be difficult to interpret). Effects were robust to choice of exact window size for template definition from 2–4 events (see Supplementary Fig. 2B).

Taken together with the previous section, these results show that episode and character representations rely on somewhat distinct brain regions. Unlike episode representations which are relatively localized to the left superior temporal lobe and TPJ, characters are represented in more distributed and bilateral regions.

Dissociable neural substrates for representing distinct elements of narrative interpretation.

Results thus far reveal that neural representations of different narrative elements involve partially overlapping yet somewhat distinct sets of brain regions. This can be appreciated visually by comparing the maps for global narrative models, episodes, and characters (compare Fig. 1D to 2C to 3C). To quantify this dissociation between the three narrative elements, we first assessed the degree of overlap in the regions representing them by correlating effect estimates across each pair of analyses. (For character representations, we identified regions that showed effects in the expected direction across all criteria for both Steve and Lucy (Fig 3C.; Supplementary Fig. 2A) and then used the mean of the estimates from the two characters.) Regions representing the narrative model were distinct from those representing either episodes (r = − 0.04, n.s.) or characters (r = − 0.02; n.s.); in turn, regions representing episodes and characters were also distinct from one another (r = − 0.05, n.s.; Fig. 4A).

Figure 4. Representations of narrative elements are dissociable and linked to partially overlapping but distinct sets of brain regions.

A. Narrative elements are represented in distinct neural substrates. Correlations between region-wise normalized effect estimates across each pair of analyses show weak or nonexistent relationships, suggesting that representations of narrative models, episodes and characters rely on distinct neural substrates. Each dot indicates a region. Coloring of a dot (region) is based on the assigned cluster (see B). B. Groups of regions support different narrative elements. We clustered regions according to their relative involvement in representing the three narrative elements: narrative model (as indexed by the pre- versus post-twist analysis), specific episodes, and characters. A solution of k = 4 clusters was found. Cluster 1 represented all three elements with a stronger weighting towards the narrative model and characters. Cluster 2 represented episodes and characters relatively equally. Cluster 3 represented characters most strongly while Cluster 4 showed weak involvement in episodes and the narrative model.

To further explore these distinctions, we applied KMeans clustering to group regions based on their distribution (pattern) of estimates from each analysis; this yielded a stable solution of four clusters (Fig. 4B). Results reinforce that the three narrative elements are somewhat dissociable in where they are represented neurally. We also note that while the canonical default mode network (DMN) was involved in representing all three elements, its different sub-regions and sub-networks have distinct and variable contributions for representing each element.

To elaborate, the first three clusters had some involvement in representing all three narrative elements, but with varying relative weights. Of particular interest are the first two clusters. Cluster 1 represented all three elements but was relatively weighted toward narrative model and character representations. In turn, Cluster 2 showed stronger, relatively equal weighting for episode and character representations. Interestingly, these two clusters both comprise parts of the DMN; the “core” regions (bilateral AG, TPJ, and precuneus/PMC) in Cluster 1 and the STS into the temporal poles in Cluster 2. The former regions have been previously reported to facilitate broad contextual and “interpretation” updating across individuals, while the latter have been associated with maintaining representations of semantic information, identities, and mental simulations^38,39. In turn, Cluster 3 and 4 exhibited even greater specificity. Both clusters were involved in weakly representing narrative models, but each cluster was paired with a distinct narrative subcomponent: Cluster 3 with strong representations of characters and Cluster 4 with weak episode representations. Together, these analyses emphasize how sets of brain regions are differentially engaged to support the organization of narrative elements.

Discussion

In this study, we investigated where and how the brain supports latent representations of distinct narrative elements. To do so, we deliberately selected an auditory narrative that featured a mid-story ‘twist’, or shift in the ground truth, that fundamentally altered participants’ understanding of earlier events. Participants listened to the stimulus twice over, carrying forward the narrative model formed after the twist into the second listen. This within-subject design enabled us to directly compare each participant to themselves as they updated their interpretations and understand how this interpretational shift altered the representation of the same sensory input. We decomposed representations of the narrative into a general narrative model as well as two components, episodes and characters, and identified brain regions where multivariate activity patterns tracked not surface-level sensory features, but rather the conceptual understanding (interpretation) of these elements. We found that the broader narrative model exhibited the most widespread representations across the brain. In turn, episodes and characters relied on partially overlapping regions with each other, yet each also engaged distinct sets of cortical regions, suggesting a degree of specialization in neural roles for representing and integrating different narrative elements.

While narratives are increasingly used to study the neural integration of information over time^11,40, researchers have paid limited attention to how subcomponents of narratives are instantiated within underlying neural representations. Many studies inherently assume that narratives are represented as a unified whole. However, behavioral evidence from prior work⁴¹ suggests that different narrative elements can be updated independently, proffering the possibility that distinct neural systems may underlie the representation of specific narrative elements^22,37,42. We provide evidence for this idea: the default mode network (DMN) supports narrative representations broadly, but there are notable distinctions across transmodal cortex and the hippocampus in the degree to which specific regions are involved in representing these different narrative elements.

Much prior work has focused on across-subject differences in activity within the DMN during narrative processing to index underlying interpretations^3,4,43,44. For instance, Zadbood et al., (2022) used an across-subjects design and a movie with a plot twist to demonstrate that representations in specific core subregions of the DMN (e.g., TPJ, mPFC, temporal poles) varied based on participants’ prior knowledge of the twist and were updated with the new information gained via the twist. Similarly, Milivojevic, Vicente-Grabovetsky & Doeller (2015) showed that event representations undergo updates following moments of insight. Our findings align with and extend these prior across-subject studies. By localizing representations of distinct narrative elements within individuals, we provide greater specificity to how narratives are represented within the DMN. We show that different regions and subnetworks of this larger network preferentially represent some narrative elements over others. While the core cortical DMN regions track the broader narrative model, lateral temporal regions, such as the STS and temporal pole, appear to support more focal representations for episodes and characters. Taken together, our findings add to the longstanding evidence that the DMN comprises multiple, interacting subsystems with distinct functions^45–47.

These topographic distinctions may in part reflect differences in not only what types of information these transmodal regions are sensitive to, but also their temporal windows of information processing and integration⁴⁸. There is extensive evidence for a hierarchical processing architecture in the brain (see⁴⁹ for a review): regions earlier in the cortical hierarchy (namely, primary sensory and sensory association regions) are sensitive to fast fluctuations in the input stream, while higher order transmodal regions are sensitive to slower fluctuations, changing only in response to longer windows of prior stimulus context. While temporal receptive windows have been commonly characterized according to surface linguistic features (i.e., comparing words to sentences to paragraphs), these windows likely also abstract away from explicit units of language to support the processing of more implicit, conceptual-level narrative features conveyed by that language. The broader narrative model and character representations operate over longer timescales than episodes; narrative models sit at the apex, episodes serve as the fundamental building blocks, and characters function as dynamic agents driving transitions between episodes. We find that regions with shorter receptive windows track faster, more time-bound fundamental narrative units (e.g., left STS represents episodes) compared to regions with longer intrinsic timescales that process slower dynamics (e.g., the dorsolateral prefrontal cortex uniquely represents the narrative model; see^12,50 and⁵¹ for a review). Future work should directly manipulate the timescales at which these features operate and interact to investigate this more systematically.

Despite these dissociations in representation, the lateral posterior parietal and temporal cortex (regions including and around the TPJ) represented all three narrative elements regardless of their position in the narrative hierarchy (Fig. 4B, Cluster 1). Outside of its general association with the DMN, a recent proposal has termed this patch of cortex “gestalt cortex,” theorizing that it is specifically involved in supporting subjective experiences, or construals, by reconciling competing interpretations⁵². To our knowledge, we are providing the first within-subject evidence for “gestalt cortex,” highlighting that representational shifts within these regions reflect internal updating of construals.

There are several limitations to this work. First, our analyses rely on a single stimulus. Although this stimulus was carefully chosen for our study design, we acknowledge that some observed effects could be driven in part by idiosyncratic properties of this particular stimulus rather than more general features of narrative interpretation⁵³. We benefited from focusing solely on one, relatively long stimulus in the auditory domain, but future work may consider employing carefully crafted multisensory stimuli to broaden generalizability. Second, relatedly, our study specifically focuses on representations of two mid-level subcomponents of narratives—episodes and characters—that could be clearly defined in our stimulus. Future research could explore more fine grained features, such as distinctions between main and secondary characters or hierarchical (nested) episode and event structures. Third, participants were quite variable in their behaviors during the study, specifically how often they used the slider to report their character impressions as well as the number of and detail associated with episodes they reported reevaluating, which limited our ability to create individualized models of representations and how they were updated. Fourth, the inclusion of an active task during and after narrative listening was a deliberate design choice, intended to capture real-time updating and to strike a balance between structured engagement and fully naturalistic listening. However, this task may have altered participants’ natural engagement with the narrative, potentially encouraging more deliberate or frequent updating of mental models than would occur during fully passive listening. Lastly, issues of MRI data quality interfered with our ability to investigate subcortical regions. Future work should explore subcortical involvement in these processes, including regions such as the amygdala, which has been implicated in supporting episodic memories and narrative processing.

In sum, we took a rigorous within-subject approach to capture how and where latent interpretive frameworks are instantiated in the brain. By holding both the participant and the sensory input constant, we identified shifts in patterns of neural activity induced by new context for the same information, allowing us to pinpoint where interpretations of distinct narrative elements are neurally represented. Together, this work provides a foundation for understanding how exogenous input and endogenous belief frameworks interact to shape subjective experience.

Methods

Stimulus description.

We used the “Dark End of the Mall” episode (18:25 min:sec) from the podcast The Truth, which consists almost entirely of a dialogue between two characters, Lucy and Steve⁵⁴. (The non-speaking time is limited to moments of a dog barking and brief moments of a “song” playing in the background.) We chose this stimulus due to the feasibility of working with only two characters and, more importantly, its unique narrative structure. Specifically, the narrative contained a twist in the middle that required participants to globally update their narrative model of events that preceded the twist, creating three distinct and meaningful narrative segments (pre-twist, twist, and post-twist) for within-subject comparison. We provide a brief synopsis below.

The story starts off with a phone conversation between Lucy, a sweet but vapid bridal shop employee, and presumably her boyfriend (whom listeners do not hear) which gets interrupted by Steve running into the shop. Listeners initially perceive Steve as a cranky dress shopper who is abrasive toward Lucy as he tries multiple attempts to convince her that she should give him some of the food hidden in the shop. Lucy gets frustrated with Steve and calls mall security and tries to kick him out of the shop. Eventually, Steve asks Lucy if he can tell her a story. It is revealed via Steve’s story that Lucy is, in fact, a robot programmed to work in a 1950’s style bridal shop, that they are both living in an apocalypse in 2050, and that Steve is one of the last surviving humans and has figured out that bridal shops have hidden snacks that sustain his survival. He almost convinces Lucy to help him, but ultimately fails as she kicks him out of the shop where, presumably, he meets his death. Listeners last hear a distressed Steve confronting barking dogs and Lucy again on the phone with her boyfriend, but listeners now realize via the narrative that the dogs are likely zombies and the boyfriend is fictitious.

Participants.

All data was collected at the Dartmouth Brain Imaging Center. Participants (n=36; 11M; median age = 20, range = 18 to 33) were healthy individuals, with normal or corrected-to-normal vision and hearing and no recent psychiatric or neurological diagnoses or MRI contraindications. They were recruited from the local areas of New Hampshire and Vermont, including the Dartmouth College student body. The Committee for the Protection of Human Subjects of Dartmouth College approved the study, and all participants provided written consent.

Experimental procedures.

All participants listened to the same auditory stimulus twice. At the beginning of the study, they were told that they may hear the auditory narrative a second time, but that they also may have the opportunity to hear a second story. No participant actually heard another story. While in the scanner, we used Sensimetrics Model S14 insert earphones to present the sound, and participants were given a trackpad (Cedrus Lumina) to continuously indicate their impressions of Lucy from very negative to very positive throughout each listen. They were given minimal visual input: the screen displayed throughout both listens showed a static photograph of a bridal shop (to promote imagery and engagement with the story) and, underneath that image, the continuous scale used to rate Lucy impressions.

Continuous Rating Task.

Participants were tasked with rating the character of Lucy by answering the question: “Overall, how much do you like Lucy?” During the presentation of the stimulus, participants used the trackpad to update their rating along a scale from −3 (very negative) to 3 (very positive) while the stimulus played. We opted to do this task in real time in the scanner as opposed to in an independent dataset of non-fMRI participants because pilot participants showed considerable variability in their ratings. Furthermore, to emphasize the within-subject design of our study, we did not want to use other participants’ data as a proxy for fMRI participants’ ratings of the character.

Before the second listen, participants were instructed as follows: “For your 2nd story, you have been assigned to listen to the same story again and complete the same prompt. For this 2nd listen of the same story, consider how your impression has changed. Because you have already listened to the story, we expect that your impressions of Lucy are different than your 1st listen. Given what you know about this story, what is your impression of Lucy now? Please use the slider to continuously rate your impression.”

Tasks after each listen.

Each scanner run consisted of the entire narrative; after each listen, participants were asked to do a series of character rating questions and memory tests (maximum of 10 seconds each) and engage in a “reevaluation task” (Fig. 1A). For the character rating questions, participants were asked to report “Overall, how much do you like [Lucy/Steve]” as independent questions. For memory questions, participants were asked “1. What is Lucy?, 2. What does Lucy hear running throughout the story?, 3. What is the name of the shop where this story takes place?” after Listen 1 and “1. What caused the destruction of humankind?, 2. What dish does Lucy recommend Steve buy at the food court?, 3. Who does Lucy think she is talking to at the beginning of the story?” These questions were intended to be relatively challenging and therefore to serve as attention checks. Participants performed well on these questions (median score 100%; mean score = 93%). After completing these questions, participants then completed the reevaluation task after each listen. After Listen 1, they were instructed “Using the microphone, please describe the moments at the beginning of the story that you reconsidered after hearing the end.” After Listen 2, they were instructed, “please describe the moments of the story that changed for you after hearing the story once before.” They had 60 seconds to answer using free speech.

Post-scan tasks.

Outside the scanner, participants were presented with the transcript and asked to “highlight the 1-3 sentences that mark the moment in the story when the twist occurred.” They were also given all of the sentences in the pre-twist segment and tasked to indicate the ones that they reevaluated. Instructions stated “highlight the sentences that mark the moments in the story that you reinterpreted when listening to it a second time.”

fMRI data processing.

MRI acquisition.

All data were collected at Dartmouth College in a 3.0 Tesla Siemens MAGNETOM Prisma whole-body MRI system (Siemens Medical Solutions, Erlangen, Germany) equipped with a 64-channel head coil.

T1 image.

For registration purposes, a high-resolution T1-weighted magnetization-prepared rapid acquisition gradient echo (MPRAGE) imaging sequence was acquired (TR = 2,300 ms, echo time (TE) = 2.32 ms, inversion time = 933 ms, flip angle = 8°, field of view = 256 × 256 mm, slices = 255, voxel size = 3 × 3 × 3 mm isotropic). T1 images were segmented, and surfaces were generated using FreeSurfer⁵⁵.

fMRI acquisition.

fMRI data were acquired using a multi-echo T2*-weighted sequence. The sequence parameters were: TR = 1,000 ms, TEs = [14.2, 34.84, 55.48], GRAPPA factor = 4, flip angle = 60°, matrix size = 90 × 72, slices = 52, multiband factor = 4, voxel size = 3mm isotropic. To account for field stabilization and hemodynamic delay, an additional two TRs were added to the front of the stimulus and 10 TRs were added to the end.

Preprocessing.

Multi-echo data preprocessing was implemented in AFNI⁵⁶ using afni_proc.py for alignment, transformation, and optimization steps. Each participant’s data was processed to align the anatomical (T1) image and functional images, with motion correction based on the second echo and alignment parameters applied to all echoes. Functional data underwent despiking (3dDespike) for outlier attenuation, followed by the concatenation and extraction of functional time series for each echo. The three echoes were then optimally combined and denoised using multi-echo ICA via tedana^57–59. Signals were then normalized to percent signal change and spatially blurred (3dBlurInMask), with motion regressors applied to reduce artifacts in final volumes. Following preprocessing, to account for transitory changes at the start of the stimulus¹, we removed the first 18 TRs from the start of the stimulus for all of our subsequent analyses (also excluded in Fig. 1A, grayed out in Fig. 2A).

Defining regions of interest.

The Schaefer parcellation⁶⁰ was used to designate 100 cortical regions; five of these regions—around the ventral part of the brain—were removed because more than 50% of participants were missing more than 40% of the data in these regions. The Harvard-Oxford Atlas was used to identify the hippocampus in both the left and right hemispheres⁶¹. We were unable to include other subcortical regions, including the amygdala, due to data loss (almost 50% of participants (17/36) were missing signal in more than 40% of voxels). Parcel sizes ranged from 113 to 759 voxels. All results shown here were robust to parcellation granularity in that effects persisted when using a 400-region parcellation⁶⁰.

Computing the ‘twist’.

We defined the “twist” in the story as moment(s) when participants transition from one interpretation/narrative model to another—specifically, from believing the setting is a bridal shop to realizing it is a post-apocalyptic world. To capture this shift, participants were asked to identify the twist in a post-scan survey (see Experimental procedures for more information on instructions). Participant responses varied considerably, with some selecting multiple points in the story. To address this variability, we adopted a conservative approach to identifying the twist, defining its start as the point before the earliest event chosen by the majority of participants, and its end as the point after the latest event chosen by the majority. This approach allowed us to split the stimulus into pre-twist (length = 532 TRs) and post-twist segments (length = 355 TRs), plus a segment in the middle corresponding to the twist itself (length = 200 TRs). Pre- and post-twist segments were matched for length when appropriate (see Section Methods: Computing neural shifts to assess narrative model representations).

Computing behavioral shifts to assess narrative model representations.

We compared “behavioral shifts” between the pre-twist and post-twist segments using the timeseries from each participant’s continuous rating of Lucy acquired in both listens. See Experimental procedures for specific instructions on how this continuous rating task was conducted. We quantified the dissimilarity (“behavioral shift”) as one minus an intra-subject correlation (intra-SC) between the behavioral timeseries from Listen 1 and Listen 2 for each segment. We compared these within-subject intra-SC values between segments using a paired t-test.

Computing neural shifts to assess narrative model representations.

Our first goal was to quantify changes in the within-subject representation of the narrative (“neural shifts”) between listens and to compare the magnitude of these changes between the pre-twist and post-twist segments. For each participant and region, we correlated the multivoxel spatial pattern at each timepoint between listens, yielding a pattern intra-subject correlation (intraSC)^18,19 for each timepoint. This was converted into a “neural shift” at each timepoint by subtracting the pattern intraSC from one (distance).

To test for a difference between segments, we computed the median neural shift value within each segment for each participant and conducted a linear mixed effects model (LMEM; using lme4 in R;⁶²) where median neural shift was predicted by the segment (pre- or post-twist) it belonged to, using participant as a random effect. Note that taking the median neural shift from each segment, as opposed to using shifts from all timepoints, helps accounts for autocorrelation in the functional data. This model was run per region. The estimates from each of these LMEMs were plotted (Fig. 1D).

To ensure observed neural shifts were not driven by differences in length of the two segments (532 versus 355 TRs), we trimmed the pre-twist segment to match the length of the post-twist segment. Specifically, we generated all possible 355-TR subsets of the pre-twist segment by sequentially trimming the pre-twist data from the start, creating 178 distinct samples (532 – 355 + 1). For each sample, we ran an LMEM for each region to compare the pre-twist and post-twist segments. The p-values from these models were then corrected for multiple comparisons using false discovery rate (FDR) based on the number of regions in our analyses (97 total: 95 cortical and two hippocampi) using an alpha of 0.05. To be as conservative as possible, we only considered regions to be significant if they were q_FDR < 0.05 in all 178 matched-length samples (Fig. 1D).

Linking moments of neural and behavioral shift.

For each participant, we aimed to identify which neural regions track behavioral shifts in interpretation. To this end, we binarized timepoints into moments where behavioral shifts were present (i.e., absolute difference of behavioral rating between Listen 1 and Listen 2 >0) or absent (i.e., absolute difference = 0). We took this binary approach, rather than directly correlating the behavioral continuous response timeseries with the neural timecourse as has been done in other studies^63,64, for two reasons. First, this approach better isolates specific moments where shifts in character impressions occur, allowing us to directly link these discrete behavioral shifts to changes in neural activity. Second, participants exhibited variability across listens in their use of the sliders both in the range of values used and in the frequency of movements (Wilcoxon signed-rank test, p < 0.05). Participants also differed amongst themselves, though not statistically significantly (std. within Listen 1: 20.5 ± 24; Listen 2: 16 ± 32.4; Kruskal-Wallis test, p > 0.05). This variability introduced potential confounds, limiting the validity of direct correlations between the neural and behavioral timeseries. For example, two participants did not move the sliders at all in the second listen (yielding an n=34 for this analysis altogether) and several moved them very infrequently, violating the basic assumptions for such correlations. By adopting a binary approach, we circumvented these issues and instead focused on the presence or absence of meaningful differences.

For each participant, we then compared the median neural shift between the timepoints when a behavioral shift was present or absent, using this (present minus absent) as our observed difference. To account for the hemodynamic delay, we shifted the behavioral timeseries by 4 TRs (4 seconds) relative to the neural data. To evaluate the statistical significance of these observed differences, we generated a null distribution by randomly shuffling blocks of time (of length 10 TRs⁶⁵) of the behavioral data 10,000 times for each participant, effectively breaking any relationship between the neural data and behavioral labels. For each permutation, we recalculated the differences between the shuffled ‘shift present’ and ‘shift absent’ timepoints, to generate a distribution of differences that would be expected under the null hypothesis (i.e., H₀: no true relationship between the neural and behavioral data). The p-value was calculated as the proportion of null differences greater than or equal to the observed difference.

Control analyses: ruling out possible confounding effects of time on within-subject similarity.

To further ensure that the observed results, which were consistent with our hypothesized directionality (higher similarity in the post-twist relative to pre-twist segment), were due primarily to shifts in interpretation rather than other explanations, we investigated the alternative hypothesis that individuals simply become more similar to themselves over time when processing the same long-timescale narrative. Critically, we tested this hypothesis both in an independent dataset as well as in our own dataset, as described below.

Computing within-subject similarity over time in an independent dataset.

We used fMRI data from an existing dataset²⁵ in which participants (N = 8) listened to the same The Moth story (“Where There’s Smoke”) multiple times. Importantly, this story lacked a twist or any other feature that might induce interpretational differences, making it suitable as a control. We used data from the first two times participants listened to the story (run 1 from session 2 and run 2 from session 3), performed functional alignment using hyperalignment⁶⁶ with a leave-one-session-out cross-validation procedure, and again parcellated the data using the Schaefer parcellation. Taking the same approach as in our main analyses, we then computed the pattern intra-SC at each timepoint. To assess whether within-subject similarity changed with time, for each region, we fit a linear model for each participant predicting pattern intra-SC as a function of timepoint (TR). We then evaluated statistical significance using a one-sample t-test (two-sided) for each region on the resulting beta values across participants. We applied a liberal, uncorrected threshold of p < 0.05 to see which regions, if any, showed increased or decreased similarity over time (Supplementary Fig. 1B).

Comparing early versus late changes within segments of our narrative.

To further investigate possible confounding effects of time (rather than interpretation) on the similarity of within-subject neural representations, in our dataset, we further divided our pre-twist and post-twist segments into early and late halves. This resulted in four distinct periods: pre-twist early, pre-twist late, post-twist early, and post-twist late. Within each of these four periods, we took the same approach: for each participant, we computed pattern intra-SC values between listens at each timepoint within each region. We then tested the hypothesis that there would be higher intra-SC (similarity) in the late compared to early periods of a segment. To do so, similar to our main analyses, we fit a linear mixed effects model per parcel predicting median intra-SC from segment period (early versus late) with random effects of both segment (pre- versus post-twist) and participant. The goal was to see where, regardless of segment, there would be greater intra-SC values later rather than earlier. Effects are plotted in Supplementary Fig. 1C.

Computing shifts in the reevaluated episodes.

As discussed in further detail in Experimental procedures, after each listen, participants verbally reported episodes – distinct events with a clear beginning, middle, and end that advanced the storyline – that they reevaluated. Then, outside the scanner, they manually highlighted the text to indicate these episodes. Our goal was to identify where and how these episodes are represented in the brain.

Given that participants varied in the number of episodes they chose, we selected five episodes consistently noted by the majority (at least 25/36; ~70%) of participants in both their in-scanner verbal reports and post-scanner written highlighting task. These episodes varied in duration (8, 9, 11, 18, and 27 TRs) and were manually checked by the experimenter to ensure that they included the entirety of an episode, i.e., if a participant chose only one of the two sentences that comprised an episode, we considered the entirety of the episode if the majority of other participants reported reevaluating all of it. Importantly, all identified episodes occurred within the pre-twist segment (Fig. 2A).

We then selected a set of “control episodes” to serve as a comparison point for the reevaluated episodes. To this end, we identified episodes within the pre-twist segment that the vast majority of participants (no more than 11/36; less than 30%) did not report reevaluating. We intentionally chose these episodes to be matched in length and nearby in time to the reevaluated episodes to account for any neural drift in the signal and to ensure that both the “control” and “reevaluated” episodes were within the same pre-twist segment (which had more overall reinterpretation, see Fig. 1). A brief description of these episodes is provided below.

The reevaluated episodes include the following: 1r. a conversation that Lucy has with an imaginary boyfriend; 2r. when Steve calls her a robot (reinterpreted from ‘corporate drone’ to actual robot); 3r. when Lucy calls Steve skinny which participants begin to realize is because he has been in survival mode for years; 4r. the ‘emergency song’ which is not a ‘hit song’ of the summer, but rather an emergency signal in the apocalypse; 5r. when Lucy tells Steve that the reason she cannot give him food and water is policy (because she is programmed to prevent this). The corresponding ‘control’ moments are when 1c. Lucy welcomes Steve to the store; 2c. when Lucy is impressed with Steve’s knowledge of the store’s policies; 3c. when he asks if his trying on a dress is against their policy; 4c. when Lucy chastises Steve; 5c. and when he calls her kind.

To compute the timings of each episode, we first used WhisperX⁶⁷ to force-align the stimulus transcript with the auditory narrative. This process yielded an onset and offset timing for each word in seconds. We defined each episode as lasting from the onset of the first word to the offset of the last word. It is important to note that the start of the first reevaluated episode was excluded because it overlapped with the portion of the stimulus excluded to account for transitory delays (i.e., therefore, we only included the remaining portion of the episode).

Next, to directly assess whether the processing of reevaluated episodes showed greater differences between listens compared to control episodes, we applied a general linear model (GLM) analysis. Using a GLM, for each participant, we modeled all episodes (10 total; reevaluated and control) in each listen using individual regressors for each episode (implemented as an individual-modulated event-related analysis using AFNI’s 3dDeconvolve function). This allowed us to obtain voxel-wise beta values for each episode. We then calculated “neural shifts” for each episode as one minus the correlation (correlation distance) between the spatial pattern of voxelwise beta values in each region between Listen 1 and Listen 2 (Fig. 2B). Lastly, per region, we fit a LMEM to test the hypothesis that neural shifts would be greater for the reevaluated compared to the control episodes. This was set up using a main effect of episode type (set as a contrast of reevaluated > control), using a random effect of participant and episode pair. Here, episode pair refers to each pair of reevaluated and control episodes that were close in time and matched in length. P-values from the models were corrected for multiple comparisons using FDR with an alpha of 0.05, based on the number of regions analyzed (97; Fig. 2C).

Computing updates in the representations of characters.

In our final analysis, we aimed to track where and how characters are represented. To this end, we identified brain regions where, despite receiving the same sensory input on Listen 2, the representations of the character Lucy were updated given the interpretation gained throughout Listen 1. We translated this into a series of four criteria that a region had to meet to be considered as representing character interpretations. Three of these criteria were in regard to the character being tested and one was in regard to the other character (Steve for Lucy and vice versa); this final “other character” criterion served as a control to ensure that the representations of the characters were distinct from one another. All criteria are described in detail later in this section (see also Methods section “Criteria for character representation updating”).

To isolate the representations of each character, we segmented the stimulus into blocks in which either Lucy or Steve was speaking (conversational “turns”), defining these as “Lucy events” and “Steve events.” We identified the onset and offset of these events using the word-level alignment times provided by WhisperX⁶⁷ and manually verified each event. We excluded events shorter than 5 TRs (such as Steve saying his name), resulting in 41 events for Lucy and 33 events for Steve (median event length: 9 TRs; range 5–27 TRs).

For each participant, we ran a GLM for each listen with individual regressors for each speaking event (implemented as an individual-modulated event-related analysis using AFNI’s 3dDeconvolve function, similar to the episodes analysis described above). Then, within each region, we took the median across voxel-wise beta values from the three last character events of Listen 1 to define a “template” representation for each character (e.g., last three Lucy and last three Steve events). Specifically, in these events, participants should have a finalized understanding of who each character is under their revised interpretation following the twist.

Thus, for each participant, we correlated the multivoxel patterns of beta values for each non-template event in both Listen 1 and Listen 2 (either Lucy or Steve) to the participant’s own character-template event (Fig. 3A). Lastly, we leveraged these event-template pattern intra-SC values to identify character representations (tested using the following criteria). Our results remain consistent regardless of template size. Supplementary Fig. 2B shows similar outcomes for both characters when using either two or four events in the template.

Criteria for character representation updating.

We expected that representations of each character to naturally evolve for participants throughout Listen 1 (e.g. for Lucy, from shopkeeper to robot), and that the updated representation would be “loaded” back into memory at the start of Listen 2. We evaluated which brain regions exhibited this representational transition—and could therefore be considered to instantiate latent interpretations of characters—as defined by the following criteria.

Criterion 1— Representations of the character become more like the template throughout Listen 1.

To test if representations of a character become more like the template, we fit a LMEM for each region, predicting the event-template pattern intra-SC in Listen 1 from the event number (with higher numbers corresponding to later events) and treating participant as a random effect. We hypothesized a positive linear trend across character events, with later events showing stronger correlations with the template as representations converge toward the final template, reflecting participants’ learning about the character across the first listen. For this criterion to be met, the statistic had to be positive.

Criterion 2—Representations of the character during their first event are ‘updated’ in Listen 2.

To test if participants “load in” their updated representation of a character when starting Listen 2, for each region and individual, we compared the correlation to the template for the first character event between the two listens. By comparing the same character event (matched sensory input) across listens to the same template, we inferred that stronger correlations with the template in Listen 2 reflected a shift toward a more updated representation of that character. For each region, we performed a paired t-test comparing the distribution of correlation values across participants between Listens 1 and 2. For this criterion to be met, the statistic had to be positive (Listen 2 > Listen 1). We additionally performed a 1-sample t-test comparing the distribution of Listen 2 correlation values to 0. For this criterion to be met, the statistic had to be greater than 0.

Criterion 3—Representations of each character stabilize in Listen 2.

To test if representations of either character “stabilize,” we compared how event-template correlations evolved over time between the two listens. We fit a LMEM for each region, predicting the event-template correlation based on an interaction of listen (Listen 1 or Listen 2) and character event number, treating participant as a random effect. As noted in Criterion 1, we hypothesized that there would be a positive linear fit of event number—that is, later character events would be more correlated to the template as the representation built up over the course of the narrative. Here, additionally, we tested that the positive slope would be steeper in Listen 1 relative to Listen 2, given that in Listen 2 the character representation requires less updating and starts out closer to the template because it has been “preloaded” into memory. For this criterion to be met, the statistic of the interaction between Listen and event number had to be positive (slope in Listen 1 > slope in Listen 2) and we furthered delineated regions that were significant at q_FDR < 0.05. This was our most important criteria– see Combining these criteria.

We additionally expected that the slope of Listen 2 be positive or at least not significantly negative at an uncorrected threshold of p < 0.05. We tested this by fitting a LMEM for each region, predicting the event-template pattern intra-SC in Listen 1 from the event number (with higher numbers corresponding to later events) and treating participant as a random effect.

Criterion 4—Control: Representations of the two characters are distinct from one another in Listen 1 and Listen 2.

As a control, we compared representations of the other character to the final template of the character being tested (e.g., Steve events to Lucy template and vice versa). Specifically, we computed event-template correlations using the other character’s events from Listen 1 and Listen 2. We then fit two independent LMEMs, with participant as a random effect, predicting this correlation from the event number. Given that we expected representations of the other character to be distinct, we did not expect a correlation with the template. Therefore, for this criterion to be met, this relationship needed to be negative or at least not significantly positive at an uncorrected threshold of p < 0.05. This would indicate that the representations of the characters are becoming increasingly distinct.

A summary of the criteria and their corresponding representation in the schematic in Fig. 3B is below.

• Criterion 1 (solid red line in Fig. 3B schematic): the slope of character event-template correlations over time had to be positive.

• Criterion 2 (arrow pointing to ‘event 1’): the correlation between the first character event and the template had to be positive in Listen 2 and greater than the same correlation in Listen 1.

• Criterion 3 (dashed red-line relative to solid red line): the slope of Listen 1 had to be significantly greater than the slope during Listen 2 and the slope during Listen 2 had to be positive or not significantly negative.

• Criterion 4 (negative blue lines): correlations of character events with the other character’s template had to be negative or not significantly positive.

Combining criteria.

We had two steps to combining these criteria. First, to be conservative, we show only regions that fit all of the expected criteria (no effect in Criterion 4 and positive in Criteria 1–3) in our character representation map (Fig. 3C). Second, we considered Criterion 3 – Representations of the character stabilizes in Listen 2 – as the most critical, given its focus on within-subject, across-listen updating and reloading of character representations. Consequently, we used estimates from this analysis for plotting and for the following analyses (see Clustering results across analyses). We also used the corresponding p-values to correct for multiple comparisons using FDR with an alpha of 0.05 across the 97 regions analyzed (black contours in Fig. 3C; Supplementary Fig. 2A).

Clustering results across analyses.

In our final analysis, we aimed to assess the extent to which representations of the three narrative elements—narrative models (Fig. 1D), episodes (Fig. 2C), and characters (Fig. 3C)—rely on overlapping versus distinct brain regions. For the character analysis, we only included regions that fit the expected direction in all four of our criteria in both characters and then took the mean estimate across characters (Supplementary Fig. 2A).

We took two approaches to comparing region-wise involvement across the three narrative elements. In each of these, we normalized the data using min-max scaling; the scaled values allowed for consistent comparison across analyses. First, we correlated the scaled estimates across regions between all possible pairs of the three analyses. Second, we used KMeans clustering⁶⁸ to perform pattern vector-based clustering, grouping regions based on the similarity of their pattern of estimates derived from each analysis.

For this second approach, we excluded regions where effects were in the opposite of the expected direction in a given analysis, resulting in the exclusion of 3 regions (3/97; 3%) from the narrative model analysis, 24 regions (~24.7%) from the episode analysis, and 21 regions (21.6%) from the character analysis. None of the 97 overall regions were excluded from all three of our analyses. We set each of these excluded regions to 0 prior to scaling estimates across regions and performing clustering.

This approach captured the underlying structure of the relationships across regions and assigned a single cluster label to each region. To determine the optimal number of clusters (k), we first calculated the silhouette score for values of k ranging from 2 to 5 (see Supplementary Fig. 3A). We selected k = 4 based on the maximum score (s = 0.45); a k = 2 solution had a higher score, but simply clustered all meaningful regions together (see Supplementary Fig. 3B).

Visualization.

Motivated by a recent recommendation⁶⁹, we present largely unthresholded, whole-brain maps for all of our main figures and add contours to indicate regions meeting criteria for statistical significance (described in detail in the caption of each figure). Although our discussion mostly focuses on only those regions meeting statistical significance, we display results across the brain to provide insight into the directionality of effects and facilitate comparisons with past and future work.

To do so, we create a translucent map weighting the data by an opacity (alpha) mask using a threshold of 20% of the maximum range of the data. For each region, we calculate an alpha value (ranging from 0 to 1) to determine its transparency level. If a value exceeds a threshold of 20%, the corresponding region is assigned an alpha value of 1 (fully opaque) and is normally plotted. For values between 0 and our threshold, the alpha value is scaled proportionally from 0 to 1, with increasing transparency for weaker effects. In the narrative model and episode analyses, regions meeting an uncorrected threshold of p < 0.05 are outlined in a blue contour, while those meeting a corrected threshold of q_FDR < 0.05 are outlined in black. In the character analysis, all plotted regions meet the uncorrected threshold of p < 0.05, with contours indicating whether a region satisfies the main criterion at q_FDR < 0.05 (in black). The SurfPlot package^70,71 was used for visualization.

Funding Statement

Funding was provided by the National Institutes of Health (R00MH120257 to E.S.F.) and (1F31MH138084-01 to C.S.S.) and by the National Science Foundation Graduate Research Fellowship (Award #2236868) to C.S.S.

Data availability.

Data from this study, including raw MRI data, will be made available on OpenNeuro upon publication.