Simulated Minds Split in Two: New Computational Model of Borderline Personality Disorder

A theoretical paper published in the journal Cognitive Computation on 23 October 2026 proposes one of the first unified computational accounts of borderline personality disorder (BPD), the prototypic condition within the Cluster B group of personality disorders. The framework, called Hierarchical Interoceptive–Social Inference Dysregulation (HISID), was developed by Marco Fronzi and is grounded in the Free Energy Principle and active inference, mathematical approaches that describe brains as prediction machines constantly minimising the surprise of their sensory inputs. According to the proposal, BPD arises when the brain’s confidence in two streams of evidence falls out of balance: signals from the body are chronically down-weighted, while signals of social threat are pathologically amplified, leaving the sense of self anchored almost entirely to volatile interpersonal cues.

The motivation for the framework lies in a persistent gap. Decades of neurobiological and psychotherapeutic research have documented the individual ingredients of BPD — instability in affect, self-image and relationships, rapid oscillation between idealisation and devaluation of others, abandonment fears and interpersonal hypersensitivity — but no single mechanistic theory has explained how these features arise together. BPD affects an estimated 0.7 to 2.7 percent of the general population, rising to roughly 12 percent in outpatient and about 22 percent in inpatient psychiatric settings, and suicide rates in affected individuals approach 10 percent. Comorbid Cluster B pathology more than doubles the odds of non-response to depression treatments, with an odds ratio of 2.27 in a pooled sample of 5,419 patients, underscoring the clinical stakes of understanding the underlying mechanisms.

HISID rests on four formal postulates. The first states a chronic precision dissociation: the precision, or inverse uncertainty, assigned to interoceptive signals is far below that of healthy controls, while precision assigned to social-affective signals is elevated. A second, distinct postulate adds an asymmetric, threat-conditional gain, so that when an observation falls below a threat threshold, social precision spikes further — a computational expression of the interpersonal hypersensitivity long described in the clinical literature. The third postulate recasts splitting, the binary oscillation between idealising and devaluing views of others, as a bimodal prior on partner valence: rather than holding a continuous belief about another person’s intentions, the modelled agent entertains only two discrete poles. The fourth postulate supplies a developmental origin, proposing that early adverse interpersonal experiences install high-precision threat priors while chronic stress-related dysregulation of the stress-response system suppresses interoceptive precision during a sensitive window.

The self-model consequences follow directly from the mathematics. In the framework, the agent’s belief about itself is a precision-weighted blend of an interoceptive belief and a social belief. Because interoceptive precision is suppressed while social precision is elevated, the social term dominates almost completely, so the self-model is determined nearly entirely by volatile social-affective inputs. From a dynamical systems perspective, this produces a free-energy landscape with a double well in the social dimension — two attractors corresponding to the idealisation and devaluation poles — and a flattened interoceptive axis that can no longer stabilise self-model updates through bodily signals. A coupling term drags the interoceptive belief toward whichever social pole is currently occupied, explaining how interpersonal events can trigger global reorganisation of self-concept and somatic experience.

To test whether this configuration is sufficient to generate BPD-like patterns, the paper presents a Level-1 simulation embedding two active inference agents — a healthy control agent and a BPD agent — in a 300-trial social-threat paradigm modelled on tasks such as Cyberball or probabilistic social feedback. The environment follows a Markov chain over neutral, accepting and threatening states, with a partially sticky threat state, while an interoceptive signal drifts according to an Ornstein–Uhlenbeck process weakly coupled to the social state. Forty independent agents were run per group. The BPD agent carried suppressed interoceptive precision, an asymmetric threat-conditional precision gain, the bimodal social prior and a hysteresis commitment rule requiring seven consecutive trials of evidence before switching poles.

The simulation produced three headline results. Across the forty agents per group, the BPD agent showed a 4.22-fold increase in self-model variance, operationalising identity instability; a 10.33-fold increase in per-agent median threat-recovery time, measured as the trials needed for the social belief to rise above a recovery threshold after threat onset; and a 4.45-fold increase in unweighted cumulative prediction-error accumulation, alongside a 2.61-fold increase in precision-weighted free energy. The authors report the unweighted measure deliberately, because the elevated precision weighting itself mechanically inflates the free-energy figure; the unweighted residual sum demonstrates a genuine increase in mismatch rather than an artefact of scaling. All group differences were assessed with one-sided Mann–Whitney U tests using one summary value per agent, avoiding pseudoreplication.

A key architectural finding emerged from what the author describes as a negative result. Gradient descent on a bimodal potential — the standard continuous updating scheme — failed to produce behaviourally plausible splitting, instead yielding high-frequency oscillation inconsistent with the extended pole-occupancy seen clinically. Behaviourally consistent splitting required two components: exact Bayesian mixture belief updating on a binary partner-state model, and a hysteresis commitment rule that locks the agent into a pole until sustained counter-evidence accumulates. Together these produce event-driven switching with long dwell times in each state, a computational signature the paper offers as a candidate formalisation of splitting as discrete pole-occupancy inference. The author is careful to note this is an existence proof of sufficiency, not a demonstration that the mechanism is unique; alternatives include continuous non-linear potentials, noise-driven switching in bistable systems, and higher-level attractor dynamics.

Because the healthy and BPD agents differed simultaneously in both parameters and architecture, the paper includes a factorial control study crossing precision profile with generative architecture in a two-by-two design, forty agents per cell. The decomposition was revealing: self-model variance was driven primarily by the bimodal-prior architecture, not by precision dissociation alone — an agent with BPD precisions but a Gaussian architecture actually showed lower variance than the healthy baseline — whereas threat-recovery slowing was driven primarily by the asymmetric threat-conditional precision, with the mixture agent at healthy precisions recovering in 3.5 trials, statistically indistinguishable from controls, versus 29.0 trials for the full BPD agent. The full phenotype reflected a super-additive interaction, with the combined effect exceeding the sum of the marginal contributions. The author maps this onto a broader distinction between state-like mechanisms, which alter the operating regime of an intact system, and capacity-like mechanisms, which alter its representational structure. Sensitivity analyses across a three-by-three precision grid and sweeps of the structural parameters confirmed that the qualitative signatures hold across the plausible parameter range.

The framework generates five operationalised empirical predictions testable with existing methods. Fitting the model to heartbeat-detection and probabilistic social-feedback tasks should reveal the inverse precision signature in patients. Trust-game or reversal-learning paradigms with computational phenotyping should expose bimodal partner-valence posteriors, assessable with Hartigan’s dip statistic. A third, Level-2 prediction — requiring an action-selection extension absent from the current simulation — proposes that self-harm may function to temporarily restore interoceptive precision, with post-harm affective relief correlating with normalisation of heartbeat-evoked potential amplitude and a transient shift toward the idealisation pole. Resting-state coupling between interoceptive and social networks is predicted to relate non-monotonically to symptom severity, and successful mentalisation-based therapy is predicted to reduce the fitted threat-precision gain and flatten the bimodal posterior, potentially providing a computational biomarker for treatment stratification.

The author is explicit about the limits of the work. The simulation is a minimal, hand-tuned generative testbed designed for interpretability, not an ecological model of human social life, and the reported statistics demonstrate that the implementation functions as designed rather than validating the theory empirically. The framework is formally specific to BPD; extension to narcissistic, antisocial and histrionic personality disorders remains theoretical, with each predicted to occupy a distinct parametric region of the same architecture. The model does not resolve whether interoceptive suppression precedes or follows adversity-driven prior formation, BPD encompasses at least three empirically distinguishable subtypes that may require modified parameterisations, and elevated social-threat precision may be shared with conditions such as PTSD and social anxiety, requiring direct transdiagnostic comparison. Elevating HISID from sufficiency demonstration to candidate mechanistic explanation will require construct validity, trajectory correspondence, parameter recovery, model comparison against plausible alternatives, and out-of-sample tests. If the five predictions hold in human data, the framework would offer a unified computational account of BPD pathophysiology and a roadmap for computationally informed psychotherapy design.

Subject of Research: A computational psychiatry framework and active-inference simulation of borderline personality disorder

Article Title: Hierarchical Interoceptive–Social Inference Dysregulation: A Computational Psychiatry Framework for Borderline Personality Disorder with Level-1 Active Inference Simulation and Implications for Cluster B Pathology

Article References: Fronzi, M. (2026). Hierarchical Interoceptive–Social Inference Dysregulation: A Computational Psychiatry Framework for Borderline Personality Disorder with Level-1 Active Inference Simulation and Implications for Cluster B Pathology. Cognitive Computation, 18(1), Article 120. https://doi.org/10.1007/s12559-026-10668-6

Image Credits: AI Generated

DOI: 10.1007/s12559-026-10668-6

Keywords: borderline personality disorder, computational psychiatry, active inference, free energy principle, interoception, splitting, Cluster B personality disorders, predictive processing, mentalisation, simulation study, precision weighting, identity instability