The causal structure and computational value of narratives

Naturalistic stimuli are used in neuroscience to approximate real-world experience and enable the study of complex cognitive processes. Most such stimuli are narratives, which are not merely sequences of events but networks of causally connected events across time. While event perception and segmentation have been widely studied, the central property of causal structure has received less attention. This review asks how causal connections shape narrative comprehension and memory, and how understanding these processes can inform computational models of learning, particularly credit assignment in reinforcement learning. The authors propose that integrating causal plausibility into models may increase ecological validity and clarify why narratives are so valuable for linking actions to outcomes in complex environments.

Behavioral literature: Early narrative research (1970s onward) showed that readers construct mental representations that go beyond surface text, relying on world knowledge; surface form and meaning constitute distinct memory traces. Coherence and schema-consistent organization improve memory. Network models of narratives represent events as nodes and causal relations as edges. Causal links can be defined by logical criteria (temporal priority, operativity, necessity, sufficiency) or by naive participant judgments. Findings include: forming a narrative from unrelated items markedly improves recollection; events with more causal connections are judged more important and are better remembered; causal chains organize event memory and influence retention; recall is influenced more by causal links than by mere temporal order; sentences that create causal breaks yield enhanced recall. Narrative comprehension involves complex causal reasoning with multiple potential causes and counterfactuals, often exceeding the scope of simple causal models, and not all causally linked events require explicit reasoning at the moment of processing. Preferential retrieval of coherent, causally linked trajectories supports efficient inference, judgment, planning, and credit assignment in high-dimensional state spaces. Neural literature: Evidence implicates default mode network (DMN) regions, medial temporal lobe (MTL) cortex, and hippocampus in constructing mental simulations that support causal reasoning in narratives. DMN and hippocampus are engaged during scene construction and imagination; hippocampal amnesics show deficits in imagining vivid scenes. During narrative communication, listeners and viewers share event-specific DMN patterns; DMN patterns update when new information changes interpretation. Long cortical timescales in DMN are disrupted by event-level temporal scrambling, possibly due to broken causal structure. Distinct neural correlates support different predictive structures: middle temporal gyrus tracks next-step predictive relationships (Kalman filter-like), entorhinal cortex tracks long-range associative structure (successor representation), with flexible switching depending on task demands. A medial prefrontal-hippocampal network scales with simulation demands in uncertain decisions; ventral temporal cortex reinstates possible outcomes, scaling with difficulty, and hippocampal activation corresponds to retrieval demands. Specific to causal structure, subjective comprehension moments in scrambled movies align with inferred causal connections and increased, integrated DMN network states. Event causal centrality in movies predicts recall and scales PMC/AG responses; hippocampal event-offset responses are larger for high-centrality events. Hippocampus shows greater similarity across narratively coherent distant events and reinstatement predicts later recollection detail. Schema-consistent narrative sequences (e.g., restaurants, airports) can be recovered in DMN during encoding and recall. Posterior medial DMN regions track temporal jumps and chronological proximity in nonlinear narratives, consistent with embedding causal structure in situation models. Reinforcement learning (RL) connections: Information that prunes possible futures carries value-like signals (information prediction errors) in dopaminergic neurons; humans and animals pay costs for information that increases certainty about outcomes. Narrative schemas confer causal plausibility, improving attention, fluency, and memory by enabling more efficient spread of value signals. Hippocampal activity increases with greater plausibility during episodic counterfactual thinking; people favor parsimonious causal models even at some accuracy cost. Dopamine signals may in some settings reflect predecessor representations (backward predictions of causally leading states), aligning with human reliance on backward planning in diverging state spaces and forward planning in converging ones. Narrative schemas function as structure priors that help overcome the curse of dimensionality by filtering irrelevant features and prioritizing relevant counterfactual simulations for credit assignment.

This article is a narrative review synthesizing behavioral, neuroimaging, and computational studies rather than reporting new experiments. Methods used across the reviewed literature include: (1) Narrative causal network construction: stimuli (stories, movies) are segmented into events based on participant-identified boundaries; participants judge causal relations between event pairs; event-by-event causal matrices (directed or undirected) are constructed with connection strengths (e.g., proportion endorsing a link); networks visualize causal structure and quantify metrics such as (weighted) degree centrality. (2) Behavioral memory testing: measuring recall, importance ratings, and retrieval order as functions of causal connectivity, causal breaks, and coherence. (3) Neuroimaging approaches: fMRI during narrative viewing, listening, recall, and imagination; inter-subject pattern similarity to compare event-specific DMN patterns; analysis of temporal receptive windows via event-level scrambling; Hidden Markov Models to recover schema-consistent event sequences; examination of hippocampal event-boundary responses; representational similarity analyses for narratively coherent event pairs; network-level analyses of DMN integration during “aha” moments. (4) Decision-making and RL paradigms: tasks assessing preference for information (information prediction errors) in primates and humans; tasks manipulating uncertainty and simulation demands to probe mPFC-hippocampal engagement; analyses distinguishing predictive rules (Kalman filter vs successor representation) across cortical regions; studies modeling forward vs backward planning in different state-space topologies.

- Causal structure is a core property of narratives that strongly shapes comprehension and memory. Events with higher causal connectivity are judged more important and are better remembered; causal chains organize recall and outweigh simple temporal adjacency effects. Constructing narratives from unrelated items enhances recollection. - DMN, hippocampus, and related MTL regions support mental simulation and situation models; their activity patterns track narrative events, update with reinterpretation, and exhibit long timescales sensitive to event-level structure. Moments of subjective comprehension in scrambled narratives coincide with inferred causal connections and more integrated DMN network states. - Event causal centrality predicts memory and scales responses in posterior DMN regions during recall; hippocampal event-boundary responses are greater for high-centrality events and for salient event boundaries. Narratively coherent distant events show greater hippocampal pattern similarity and reinstatement predicting detailed recollection. - Distinct neural representations support different predictive structures: middle temporal gyrus encodes next-step predictive relationships, entorhinal cortex encodes long-range associative structure; individuals flexibly shift horizons of prediction, consistent with narrative-supported multi-scale forecasting. - In value-based learning, information that reduces uncertainty elicits value-like neural signals (information prediction errors) and is preferred behaviorally, even without changing reward magnitude or delay. Narrative schemas provide causal plausibility that prunes state spaces, improving attention, fluency, and memory encoding. - Plausibility modulates hippocampal activation during episodic counterfactual thinking; people prefer compressed (parsimonious) causal models even at some accuracy cost, supporting efficient representation of causal chains. - Dopaminergic signals can reflect predecessor representations of causally leading states, aligning with human reliance on backward planning in diverging state spaces and forward planning in converging ones, suggesting that compact forward and backward predictive representations facilitate planning and credit assignment in complex environments. - Narrative schemas function as structure priors that mitigate the curse of dimensionality by filtering irrelevant features and selecting relevant counterfactual simulations, thereby improving credit assignment and plan selection.

The review demonstrates that causal structure is a crucial determinant of how narratives are processed and remembered, engaging episodic memory systems and DMN to construct and update mental simulations of event sequences. These findings address the central question by showing that causal connectivity enhances memory encoding and retrieval, and that neural activity in DMN and hippocampus reflects both the construction of situation models and sensitivity to causal relationships. Bridging to reinforcement learning, the authors argue that narrative schemas confer causal plausibility that narrows state and action spaces, enabling more efficient credit assignment and planning through both forward and backward predictive representations. Incorporating causal plausibility into learning models could better capture how organisms link actions to outcomes in naturalistic, high-dimensional environments. The significance lies in reframing naturalistic neuroscience to explicitly consider causal networks across events, offering mechanistic insights into why narratives are privileged in memory and how value-guided cognition leverages structured knowledge to navigate complex worlds.

Causal structure is foundational to narrative comprehension and memory and influences neural systems implicated in episodic simulation and recall. Recognizing and measuring causal connectivity can clarify prior findings with naturalistic stimuli and guide future experimental designs. Integrating narrative plausibility into reinforcement learning frameworks promises greater ecological validity and improved accounts of credit assignment and planning. Future research should test whether the effects seen with temporal scrambling arise from disruption of causal structure, delineate the role of the DMN in causal reasoning versus downstream consequences of enhanced memory processes, chart how narrative plausibility is acquired developmentally, and examine how different causal types (physical, biological, psychological, social, structural) recruit distinct cognitive and neural mechanisms. Considering causal structure may inform interventions leveraging narratives in mental health and illuminate how humans efficiently learn from complex, multi-event experiences.

Direct neural investigations of causal structure during narrative processing remain limited. It is challenging to disentangle whether DMN activity reflects causal reasoning per se or consequences such as enhanced memory encoding/retrieval or imagery. Retrospective ratings of causal links may not map precisely onto the timing of on-line causal reasoning, can change with later information, and vary across individuals. Current causal reasoning models often fail to capture the complexity of multi-cause, multi-step narratives. Many laboratory tasks lack cross-trial causal structure, potentially missing key mechanisms. The taxonomy of causal connection types and their distinct neural substrates is not well specified, and the developmental acquisition of narrative plausibility is incompletely understood.