The Representational Geometry of (Dis)Belief: A Multidimensional Space Organized by Belief Direction and Extremity

Technical Summary: The Representational Geometry of (Dis)Belief

Problem and Motivation
In an era characterized by competing narratives and misinformation, understanding how the brain represents belief states is critical. While prior research has extensively studied the processes of belief formation (e.g., how prior expectations shape acceptance), the representational geometry of belief—how information is encoded in neural patterns under different belief states—remains underexplored. Previous neuroimaging studies have largely focused on discrete judgments of true/false statements, which capture decision-making after information presentation rather than the evolving representation of meaning during naturalistic narrative comprehension. Furthermore, existing models often assume belief is represented along a single linear axis (belief vs. disbelief). This study challenges that assumption, proposing that belief representation may be multidimensional, organized by distinct axes such as directional position, extremity, or categorical proximity.

Methodology
The study employed a naturalistic fMRI approach, re-analyzing a dataset from Braun et al. (2025) comprising 68 participants listening to six extended juridical-style auditory narratives. Contextual information preceding each narrative manipulated speaker credibility to induce a broad distribution of belief ratings (0–100 scale).

The core analytical framework was Inter-Subject Representational Similarity Analysis (IS-RSA). The authors constructed five competing behavioral similarity models to test different hypotheses about how belief organizes neural activity:

1. $f_{dist}$ (Distance): Similarity based on absolute proximity along the belief-disbelief continuum.
2. $f_{mean\_belief}$ (Directionality): Similarity based on directional position (mean of ratings), distinguishing movement toward belief from movement toward disbelief.
3. $f_{side}$ (Shared Side): Similarity based on occupying the same pole (belief vs. disbelief) regardless of magnitude.
4. $f_{extremity}$ (Extremity): Similarity based on the strength of conviction, grouping strong believers and strong disbelievers together against ambivalent participants.
5. $f_{3cat}$ (Categorical): Similarity based on discrete bins (disbelief, uncertainty, belief).

Neural similarity matrices were computed for 1,026 cortical parcels. These were correlated with the behavioral model matrices using rank-transformed, z-scored vectors. Significance was assessed via phase-randomization and Mantel-style permutations with FDR correction.

To address whether neural patterns were dominated by a single axis or a mixture, the authors performed a Ridge RSA analysis, modeling neural similarity as a weighted combination of all five predictors. They calculated an "effective number of models" metric to quantify the dimensionality of the representation.

Finally, to interpret the neural finding of an "extremity" axis, a preregistered behavioral follow-up experiment ($N=105$) was conducted. Participants rated their belief and their certainty in their interpretation of the same narratives to test for a quadratic relationship between belief extremity and certainty.

Key Results

1. Dominant Axes: In a "winner-take-all" model competition, neural similarity was significantly explained only by two models: directional belief position ($f_{mean\_belief}$) and belief extremity ($f_{extremity}$). Models based on simple distance, shared side, or categorical bins did not yield significant results.
2. Anatomical Distribution: Significant effects were widespread, primarily located in the Default Mode Network (DMN), Control Network, and Temporal-Parietal regions. Both positive and negative RSA effects were observed, indicating that different brain regions track movement toward belief or disbelief, and different regions track extremity versus ambivalence.
3. Multidimensionality: Ridge RSA analysis revealed that significant parcels were rarely explained by a single model alone. The "effective number of models" was high (averaging 3.88–4.27), indicating that the neural geometry of belief is a multidimensional mixture of axes. However, the dominant predictors in these mixtures remained directionality and extremity.
4. Behavioral Validation: The behavioral follow-up confirmed a quadratic relationship between belief ratings and certainty in interpretation. Certainty was highest at both extremes (strong belief and strong disbelief) and lowest near the midpoint (uncertainty). This supports the hypothesis that the neural "extremity" axis reflects the degree of interpretive certainty or the resolution of a narrative into a stable mental model.

Significance and Claims
The paper claims to characterize the representational geometry of belief as a multidimensional, context-sensitive space organized primarily around two axes: belief directionality (distinguishing belief from disbelief) and belief extremity (distinguishing committed states from ambivalence).

Key contributions include:

• Reframing Belief Representation: Moving beyond the view of belief as a single linear dimension or a simple binary state, the study demonstrates that belief and disbelief are distinct representational directions that can both be associated with high certainty (extremity).
• Neural-Behavioral Link: The study provides a direct link between a specific neural representational geometry (the extremity axis) and a behavioral signature (certainty in interpretation), suggesting that strong belief and strong disbelief are psychologically similar in terms of the resolution of the narrative, even if they differ in content.
• Context-Dependence: The findings underscore that belief representation is not localized to a single "belief hub" but is distributed and varies depending on the specific narrative and the interpretive task. Different narratives elicited different neural maps, suggesting that abstract concepts like belief are dynamically constructed through the interaction of stimulus, task, and prior knowledge.

The authors conclude that belief representation is best understood as an interaction between the brain's latent structure and the specific context of the narrative, where directionality and extremity serve as the primary organizing principles for how listeners synchronize their neural responses to shared meaning.