Computational Psychiatry

Deciding when to stop gathering information and commit to a choice is a fundamental challenge in decision-making under uncertainty. Normative characterizations such as Partially Observable Markov Decision Processes (POMDPs) prescribe mathematically optimal stopping rules; however, human evidence gathering systematically departs from optimality. Pathological departures -- such as the excessive indecisiveness characteristic of obsessive-compulsive disorder (OCD) -- offer an important opportunity to investigate the cognitive mechanisms involved in stopping. We extend a POMDP framework to incorporate key candidate suboptimalities: a biased prior belief, transient evidence exaggeration, progressive forgetting, boosted costs of error, temporal regulation (patience and urgency), and misperception of a deadline. We evaluate this model in a pre-existing dataset comprising 105 participants spanning healthy controls, generalised anxiety disorder, and the OCD spectrum performing an information gathering task with controlled, stochastic, deadlines. Model comparison reveals that human sequential choices are broadly governed by subjective risk penalties and time-dependent urgency, with a smaller and less certain contribution from an over-weighting of recent evidence, which a random-effects comparison does not support at the population level. Individuals differ in how that over-weighting is implemented: in one deadline condition, subjects divide almost evenly between models carrying a transient exaggeration of the newest sample, models carrying progressive forgetting of older evidence, and models carrying no recency mechanism at all. Crucially, while risk sensitivity and choice stochasticity act as shared mechanisms across conditions, mechanisms such as belief bias and patience are more variable. Finally, using OCD as a clinical case study, we demonstrate that simulating choices from the fitted exaggeration model reproduces model-agnostic regression signatures of clinical indecision, which the forgetting and no-recency accounts do not. These findings offer a generative foundation for dissecting clinical departures in information gathering across the obsessive-compulsive spectrum.

Metacognition is the ability to reflect on and evaluate our own cognitive processes. It is often altered in psychopathology. Yet, the computational mechanisms underlying these alterations remain unclear. In this work, we extend Hierarchical Gaussian Filter (HGF) models to jointly fit trial-by-trial predictions and confidence ratings in a predictive inference task, providing an individualised characterisation on metacognitive processing. Applying our cognitive computational model to a large subclinical open dataset (N=430), we are able to achieve, on average, excellent fit of prediction responses [Formula] and a moderate to good fit of confidence ratings [Formula]. Analysis of experimental change-points revealed that our model accurately captures confidence self-reports dynamics around these change-points. Posterior parameter estimates reveal a negative effect of sensory input prediction errors and a positive effect of sensory input prediction precision on confidence ratings, respectively. In addition, we replicate state-of-the-art findings related to compulsivity as measured by a transdiagnostic factor score, such as inflated confidence and a decoupling of action updates (here, prediction errors) and confidence in compulsivity. These results demonstrate the robustness of our methodology and the potential of joint prediction-confidence modelling to uncover latent metacognitive alterations in psychopathology.

The present study examined the relationship between the DSM-5 diagnostic criteria for gambling disorder (GD) and gambling severity via an item response theory (IRT) analysis in two large German population survey data sets (Buth et al. (2022, 2024)). IRT-based person fit analyses may reveal atypical response patterns (e.g. endorsing criteria linked to higher levels of disorder severity, but not criteria linked to lower levels). We examined the link of such atypical response patterns and mental health as measured by the MHI-5, employing a 2-parameter-logistic (2PL) IRT model and a linear mixed model with random intercepts. Results largely replicated previously reported item severity rankings across both samples: GD criteria such as loss chasing and a preoccupation with gambling were generally linked to lower severity levels, whereas criteria such as withdrawal symptoms or job/family problems where generally linked to higher severity levels. Modelling revealed a reduced assessment sensitivity in lower gambling severity ranges. Furthermore, person fit analyses suggest that atypical symptom patterns may be linked to poorer mental health (MHI-5). Implications for the interpretability of total scores of endorsed criteria and the validity of diagnostic practices determining eligibility for treatment and financial compensation are discussed.

Effective decision-making requires evaluating options based on relevant features while reducing attention to those that are not. However, how attention prioritizes one source of information over another during value-based decisions remains unknown. Across 7 experiments, we show that irrelevant features both constructively and destructively interfered with how values are assigned to choice options. When relevant and irrelevant features agreed, choice accuracy and speed were facilitated; when they conflicted, accuracy decreased and choices slowed in proportion to the value of the irrelevant feature. We modeled this interference as a weighted sum of relevant and irrelevant feature values. This model explained individual differences in choice behavior and, critically, tracked psychiatric symptom severity related to substance abuse and obsessive-compulsive disorder. This suggests that these symptoms may, in part, reflect an impaired ability to weigh relevant against irrelevant information when assigning value to choice options - a process we term feature-based valuation.

Catatonia offers no principled basis for sequencing interventions when first-line benzodiazepines fail. Electroconvulsive therapy (ECT) is the established next step, but specifies what to escalate to, not what to stabilise first. Here we show that recovery is a constrained progression across five precision domains of hierarchical inference: sensory ({pi}s), policy ({beta}), motivational ({pi}m), and fast and slow volatility precision ({pi}v_fast, {pi}v_slow). In twenty-five consecutive inpatients managed without ECT, an order {pi}s [-&gt;] {beta} [-&gt;] {pi}m [-&gt;] {pi}v_fast [-&gt;] {pi}v_slow) held without inversion in every patient. Bush-Francis Catatonia Rating Scale (BFCRS) scores fell from 26.3 {+/-} 5.6 to 2.0 {+/-} 2.4 (p < 0.001), and functional recovery tracked restoration of organized action rather than symptom suppression. The framework predicts, untested in this uniformly remitting cohort, that interventions effective at one stage may destabilise another. Within stated conditions, a single inversion falsifies the ordering.

Importance. Suicide is a leading cause of mortality, yet its prevention remains a challenge. Suicide attempters exhibit disrupted decision-making, but the specific cognitive processes driving this impairment and their utility for risk stratification are unclear. Objective. To determine which computational decision-making mechanisms are disrupted in high-risk suicide attempters, and whether they distinguish attempters from depressed and healthy controls. Design. This cross-sectional study fit reinforcement-learning models to Iowa Gambling Task performance across two independent studies: SPAD (August 2017 to May 2020); SUICIDE-DECIDE (March 2021 to September 2023). Data was analyzed between January 2025 and July 2026. Setting. SPAD was conducted in Nimes (France). SUICIDE-DECIDE was a muticenter study across Jena (Germany) and Nimes. Clinical groups were recruited from academic inpatient psychiatry units and outpatient clinics, and healthy controls through local advertisement. Participants. SPAD included adults aged 60 years or older; SUICIDE-DECIDE middle-aged adults stratified by attempt means (violent/nonviolent). Consecutive participants were recruited into three groups: suicide attempters, patient controls, healthy controls. Main outcomes and measures. The primary outcomes were group differences on parameters reflecting learning from rewards and losses, sensitivity to win frequency, choice perseverance, and forgetting of past choices. We trained a multinomial logistic regression model in the SUICIDE-DECIDE study using self-reported depression and parameters that differentiated attempters from controls across studies, and evaluated its performance in the SPAD sample. Results. SPAD included 57 participants (40 females [70.2%]; mean [SD] age, 70.1 [8.2]). SUICIDE-DECIDE included 124 participants (77 females [62.1%]; 32.1 [11.0]). Suicide attempters forgot past choices faster than controls (SPAD: F(2,36) = 13.5; P < .001; SUICIDE-DECIDE: F(3,57) = 46.6; P < .001). No other parameter specifically differed in attempters across studies. In SUICIDE-DECIDE, greater forgetting of past choices was associated with attempter status only (RRR = 4.2, 95% CI [1.5, 11.5]). This parameter paired with depression severity correctly classified 87.2% of participants in SPAD (95% CI [74.3%, 95.2%], 41/47). Conclusions and relevance. Short-sighted decision-making may represent a cognitive phenotype associated with suicide in high-risk populations and a potential target for prevention and treatment.

Despite its strong link to neuropsychiatric conditions, frustration remains critically understudied in humans and animals alike. Therefore, there is an urgent need to develop tools to understand and therapeutically target frustration-related functions. Interestingly, humans and rats respond similarly during frustrative nonreward by increasing barpress durations. We previously validated barpress duration in rat operant tasks as a reliable measure of frustration-related behavior; however, it is wellknown that in addition to duration of responding, emotional states such as frustration alter other aspects of responding such as force of pressing. One-dimensional, static measures such as maximum force could miss rich information contained within operant data. Thus, the objective of this study is to apply machine learning (ML) to force/time profiles to discriminate frustration-related barpresses from non-frustration-related barpresses. Results showed an AUROC for FR1 (i.e., non-frustrated) vs. extinction (frustrated condition) for individual barpresses of 0.65 that improved to 0.84 with a chunk size of 10. The model generalized well to progressive ratio responding, a different kind of frustration procedure. We conclude that force/time profiling does provide utility beyond one dimensional measures of duration or force separately, meaning that we can indeed infer the internal state of frustration from behavior using ML techniques. Importantly, this project will also serve as proof-of-concept for applying ML to predict other internal states from barpress data.

Background: Substance use disorders (SUDs) are associated with impaired goal-directed decision-making, but the specific cognitive mechanisms underlying these deficits remain unclear. Forward planning, a key component of goal-directed control, requires strategies to contain computational demand and ensure feasibility through heuristics such as reducing planning depth, pruning unlikely options, and discounting probabilistic outcomes. This study investigated whether individuals with alcohol use disorder (AUD) and tobacco use disorder (TUD) differ from healthy controls in their use of these heuristics during sequential decision-making. Methods: We examined participants from the general population meeting criteria for at least moderate AUD (n = 45), TUD (n = 49), or both (n = 45), and compared them with matched healthy control groups. To assess forward planning, we used a previously published probabilistic three-step Markov decision task. Computational modeling was used to infer individual planning depths and compare model variants incorporating low-probability pruning, probability discounting, or both. Results: A combined model including both low-probability pruning and probability discounting best explained participants' behavior and was subsequently used for analysis. Compared to matched healthy controls, all groups with SUD exhibited significantly reduced planning depth. Moreover, individuals with SUD tended to show stronger probability discounting, with this effect reaching statistical significance in the comorbid group. Conclusion: Individuals with AUD and/or TUD engaged in forward planning using similar strategies as healthy controls but exhibited reduced planning depth and a tendency toward stronger probability discounting. These findings suggest the presence of altered planning processes in SUD, which could contribute to maladaptive decision-making. Future research is needed to clarify whether such differences reflect pre-existing cognitive traits, consequences of substance use, or both, and to explore their potential relevance for intervention.

BACKGROUND: An individual's adaptation to threatening environments under uncertainty is reflected in stress responses. Predictability (the ability to anticipate events) and controllability (the ability to control outcomes) are central to how one adapts, yet their joint influence on aversive learning remains unclear. METHODS: Thirty healthy adults completed a probabilistic aversive learning task in which cue-outcome contingencies varied across levels of predictability and controllability, i.e. whether shock intensity depended on prediction accuracy. Prediction accuracy, reaction time, subjective stress ratings, and skin conductance responses were recorded throughout. Trial-wise learning dynamics were estimated using the Volatile Kalman Filter. RESULTS: Prediction accuracy reduced as environments became less predictable and negatively associated with higher learning rates across predictability levels, with the strongest relationship observed in highly predictable blocks. Skin conductance responses showed that moderately predictable environments elicited responses like those in highly predictable environments when accurate predictions reduced shock intensity, but resembled responses in unpredictable environments when shock intensity was uncontrollable. Model comparison revealed a double dissociation between subjective stress ratings and skin conductance responses. Subjective ratings were best explained by model-derived volatility when prediction accuracy determined shock intensity and by belief uncertainty when it was independent of prediction accuracy, whereas skin conductance responses showed the reverse pattern. Reaction times were best explained by belief uncertainty when predictions influenced shock intensity. Higher anxiety was associated with elevated learning rates in highly and moderately predictable blocks when predictions did not control shock intensity.

Suicide risk assessment in clinical conversations relies on the sequential structure of therapist-patient dialogue, yet formal models of how suicide-related content emerges, transitions, and re-ciprocates across conversational turns are lacking. We applied network analysis to C-SSRS-coded utterances from psychotherapy sessions to characterize the topology and dynamics of suicide-related clinical communication. Transcripts of 110 psychotherapy sessions from the SPEAK-SAFE study were classified for suicidality relevance using the Qwen3-32B large language model with an adapted Columbia-Suicide Severity Rating Scale (C-SSRS) schema. From the suicidali-ty-relevant utterances, we constructed a weighted directed category transition network. Centrali-ty, reciprocity, clustering, assortativity, and path length metrics were computed. Independent human expert ratings for a session-level subsample of 30 sessions from this cohort showed sub-stantial agreement with the LLM classifications (Spearman {rho} = 0.885, ICC[2,1] = 0.708). The network was dominated by therapist inquiry. Reciprocity was high, and closeness centrality de-creased monotonically with C-SSRS severity. Active suicidal ideation functioned as a bridge node from which escalation pathways to method, plan, and attempt categories originated. The network exhibited a hub-and-spoke topology with small-world characteristics. Suicide-related clinical dialogue is structured as a therapist-centered hub-and-spoke network with high reciproci-ty and a consistent severity-peripherality gradient. The extension of network theory from intra-individual symptom structures to inter-individual communication dynamics represents a method-ological innovation demonstrated here as a proof of concept, with potential applications to clini-cal training, automated decision support, and multimodal risk assessment. clinical interviews clin-ical interviews

Recent computational approaches to speech in psychosis generate large multidimensional feature spaces capturing semantic and acoustic aspects of language production. However, the clinical relevance of individual measures often becomes difficult to interpret due to redundancy, interaction effects, and high intercorrelation among features. Building upon previous work constructing a single composite index derived from semantic features based on language model embeddings, we here build a second acoustic index derived from speech acoustic features. Our aim was to evaluate the differential performance of both indices in conjunction in PANSS symptom prediction in psychosis in a cross-linguistic setting, including positive symptoms (P1, P2, P3), negative symptoms (N1, N4, N6), and general measures (G5, G9). The dataset comprised five languages and 221 patients with schizophrenia spectrum disorder (SSD). Both indices showed predictive power for individual PANSS scores, while also demonstrating clinically important complementarity: semantic indices were more strongly associated with positive symptom dimensions (P2, P3, Total Positive), whereas acoustic indices showed stronger relationships with negative and general symptoms (N1, N4, G5, G9). Both domains shared predictive overlap for global measures, such as PANSS Total scores. These findings suggest that both indices capture complementary and partially overlapping dimensions of psychopathology. The proposed composite index framework contributes to the advancement of low-dimensional speech-derived markers of symptom severity variation, potentially informing vulnerability to relapse and remission in psychosis.

The rat gambling task (rGT) has been widely used to investigate the neural mechanisms underlying risky choice and motor impulsivity. Here, rats sample between four options (P1-P4) that vary in the size and probability of reward and time-out penalties. The optimal strategy is to avoid risky options that may yield higher per-trial gains, but deliver longer and more frequent time-outs. Previous reports suggest pairing wins with salient audiovisual cues increases risky decision making, but behavioural variation is high, and it is unclear whether motor impulsivity is also affected. Here we leveraged rGT data from over 750 rats to characterize behavioural performance across sex and cue condition. We compared different methods of classifying rats as optimal or risk-preferring, using either a unitary decision score variable or specific P-choice preference, and applied drift diffusion modeling (DDM) to explore whether divergent cognitive mechanisms underlie risky decision making across subgroups. We confirmed that risky choice is higher on the cued rGT, partly due to a greater proportion of risk-preferring rats, but also because net optimal decision-makers chose the risky options more often. Risk-preferring rats made more impulsive, premature responses regardless of cue condition, as did males. Optimal decision-makers made more premature responses when cues were present, such that premature response rates were higher overall on the cued rGT. DDM and response latency data suggest divergent cognitive processes underpinning risky decisions across sex. Wider decision boundaries were associated with both highly optimal and highly risky choice patterns, indicating risky choices are made deliberatively by highly risk-preferring individuals. Similar results were obtained regardless of classification method.

Adults who stutter can speak fluently when speech is not addressed to another person, but stuttering emerges when they aim to convey information to a listener. The value of the information being conveyed to the listener also affects the likelihood of stuttering. Why should the mere absence of a listener neutralise a profound motor deficit, and why does a word's predictability affect whether it is spoken fluently? To resolve this socio-motor paradox, we develop a computational model of stuttering within an active inference architecture. The model represents the communicative context, including whether a listener is present and whether the agent is speaking or listening. It was designed around two candidate mechanisms for stuttering, a prior for silence and rigid phoneme sequencing precision. Using both, the model produced fluent private speech and more stuttering-like events during social speech. In the same parameter regime, the model also showed more stuttering-like events on words with higher information value, and produced a word-length effect, in which disfluency increased with longer words. To our knowledge, this is the first model of stuttering to generate both the private speech and the information-value effect from inferred communicative context. By representing the listener as a hidden state that makes the sensory consequences of resuming speech ambiguous, the model offers a computational link between social cognition and speech-motor instability, and suggests that speech fluency depends on whether the speaker believes anyone is present. Clinically, it may offer testable hypotheses and a route to personalising treatment, since the same overt severity can arise from different combinations of parameters.

Background. The SRQ-20 screens common mental distress in low- and middle-income countries, yet its equivalence across groups, especially armed-conflict exposure, is rarely tested with methods that separate true invariance from an underpowered null. We evaluated its psychometric properties and measurement equivalence in Colombian adults. Methods. In 10,865 adults from the 2015 Colombian National Mental Health Survey, we assessed dimensionality, fitted a two-parameter logistic (2PL) model, and tested equivalence across armed-conflict exposure, sex, and region using multiple-group models with purified anchoring and freely estimated group means, separating true distress differences from item bias. Item functioning was equivalence-tested against a {+/-} 0.10 band on signed expected-score differences (SIDS), with test-level differential test functioning (DTF) plus severity-graded and design-weighted sensitivity analyses. Criterion validity used design-weighted ROC against 12-month CIDI diagnoses. Results. The scale was essentially unidimensional (one-factor CFI = .945, rising to .971 with four content-redundant item pairs modelled; explained common variance = .71) and fit the 2PL well, with high conditional reliability at the cut-points (.93-.94). Once true distress differences were separated from item bias, the SRQ-20 was equivalent across armed conflict exposure (all |SIDS| < .10, maximum .03; net DTF {approx} 0.1 points) and region, holding even among directly victimised adults; sex was partially invariant (three items; DTF {approx} 0.85 points). Design-weighted AUC was .88 (major depression) and .84 (any disorder). Conclusions. The SRQ-20 measures distress equivalently across armed-conflict exposure (including direct victimisation) and region in Colombian adults, supporting exposed-non-exposed comparisons within this population; raw-total sex comparisons carry a small, quantifiable bias. The 20-item form is recommended.

Background: Interview-based suicide assessments are insufficient for near-term risk detection, but there are no evidence-based, high-performing alternatives. The Death Implicit Association Test (DIAT) has been proposed as a suicide risk assessment tool. However, traditional DIAT scoring methods exhibit limited association with suicidality. Methods: We tested whether temporal structure in reaction times (RTs) from the Brief Death Implicit Association Test (BDIAT) tracks active suicidal ideation (SI), which was ascertained by ecological momentary assessment (EMA). Seventy-seven participants who either did or did not endorse active SI during EMA (SI+ n=22; SI- n=55) completed 18 alternating blocks of the BDIAT (9 Life + Me, 9 Death + Me). We computed the conventional D-score, an aggregate index of self-life versus self-death associations, and separately treated RTs as time-varying signals, examining entrainment to block alternation and habituation during conflict-related trials. We then developed a bilinear logistic regression classifier to extract these temporal RT features and trained it to predict SI status, with performance validated using leave-one-out cross-validation. Results: The conventional D-score did not reliably classify SI+/SI- individuals. In contrast, temporal RT dynamics reveal group differences in block-level entrainment and trial-level habituation. A bilinear logistic regression classifier leveraging these dynamics predicts SI status with 77% balanced accuracy and AUC 0.799. Conclusions: Results demonstrate that temporal dynamics of RT during the BDIAT track active ideation and can be used to distinguish SI status at the individual level, implicating RT temporal dynamics evoked during BDIAT as a candidate behavioral marker of near-term suicide risk.

Autism Spectrum Disorder (ASD) is a heterogeneous neurodevelopmental condition defined by differences in social communication and restricted, repetitive behaviours. As diagnostic criteria have broadened, ASD is now recognised across a wider range of individuals, raising key questions about its structure: does ASD have discrete sub-types, or is it better conceptualised as a continuous, possibly multidimensional, condition? We aim to explore whether a multidimensional continuum model more accurately captures the variability within ASD. We analysed a large SPARK phenotypic dataset of medical history and diagnostic surveys (background history, SCQ, RBS-R; n=36,710 individuals). We apply and compare two traditional statistical approaches, Factor Analysis and Gaussian Mixture Models, with a modern machine learning technique, the Variational Autoencoder (VAE). VAEs reconstructed unseen test data with ~4-fold better accuracy than Factor Analysis, and ~8-fold better accuracy than Gaussian Mixture Models. We identified four stable latent factors across 100 independently trained VAEs. These four dimensions provide an individual behavioural profile that can be visualized using radar-plots, offering a compact way to compare profiles at the person level. Through further analysis, we found evidence for 3 overlapping clusters or subtypes of ASD identified within the 4D latent space. This work aims to inform new ways of modelling ASD using a VAE that will be able to discern between a continuum or a clustered output and that go beyond binary diagnosis, instead reflecting the complex range of trait profiles, with implications for personalised diagnosis and intervention.

Current dimensional approaches to psychiatric disorders have largely focused on explaining differences between individuals, whereas it remains unknown whether symptom dynamics within individuals are organized by the same underlying dimensions. In PTSD, temporal symptom variability may represent a clinically meaningful source of heterogeneity relevant to spontaneous recovery, chronicity, and treatment response. Our reciprocal inhibition model of PTSD proposed that both between-individual heterogeneity and within-individual dynamics may be organized along a dimension reflecting the relative balance between re-experiencing and avoidance symptoms (symptom imbalance), potentially corresponding to shifts between states of emotional under- and overmodulation. Here, using seven longitudinal and two cross-sectional PTSD cohorts spanning disorder development, chronicity, and recovery, we examined whether symptom heterogeneity between individuals and within individuals over time is organized along shared latent symptom dimensions. Principal component analysis (PCA) performed separately on between-individual variability (individual differences) and within-individual variation (temporal variability) consistently recovered the same two axes: the first indexing overall symptom severity and the second reflecting the proposed symptom imbalance. To enable direct comparison across cohorts and between-individual and temporal scales, we integrated cohort-specific covariance structures using hierarchical multi-group PCA yielding universal axes (uPC1/uPC2). Mapping treatment trajectories onto this shared symptom space revealed that two first-line psychotherapies: cognitive processing therapy (CPT) and eye movement desensitization and reprocessing (EMDR): produced comparable reductions in overall symptom severity (uPC1), but opposite shifts along symptom imbalance (uPC2). These findings suggest treatment-related symptom trajectories that are not captured by severity alone and provide a quantitative basis for treatment stratification grounded in symptom imbalance dynamics, motivating prospective tests of state-dependent intervention in PTSD and related psychiatric disorders.

Pavlovian bias, a motor bias driven by emotionally-valenced context, has long been considered a proxy for emotion-action processing, which may become aberrant in affective disorders and suicidality. The orthogonal valenced Go/NoGo paradigm was developed to isolate this bias and has been widely adopted in translational psychiatry research under this interpretation. However, several unresolved theoretical inconsistencies highlight the need to dissociate potential confounding factors within this paradigm before assigning mechanistic and clinical interpretations to its findings. Using a novel paradigm that fully orthogonalized valence, action, and aimed transition in a normative sample of 25 individuals, we found that the "active escape bias" previously observed in aversive contexts--and interpreted as reflecting suicidality in the clinical population--was reproduced in appetitive contexts using trials with the same transition structure. Furthermore, computational modeling indicated that the bias patterns isolated by the orthogonal paradigm are primarily explained by the combined effects of aimed transition and cue salience, without requiring emotional-valence modulation. These findings call for a fundamental reconsideration of how orthogonal Go/NoGo paradigms are interpreted in translational psychiatry, particularly how the behavior they isolate is mechanistically linked to affective symptoms.

The development of large language models (LLMs) offers promising opportunities for predicting human behavior across diverse contexts. However, most prior work has emphasized behavioral imitation, with limited attention to transparent or interpretable models of the cognitive mechanisms underlying human decisions. In this study, we introduce MindEvolve, an autonomous workflow designed to predict behavior in social interactions by generating interpretable symbolic models of cognition. We systematically evaluate the modeling capabilities of multiple LLMs across a battery of socioeconomic games covering four core domains of social cognition: economic preferences, social preferences, social reasoning (theory of mind), and recursive planning. The symbolic models proposed by the LLMs are assessed by expert human evaluators for interpretability and theoretical coherence. Our results show that while most LLMs robustly capture economic and social preferences in relatively simple strategic settings, and in some cases generate novel models that integrate broader knowledge than those proposed by human experts, their capacity to model more complex psychological processes remains limited. However, a subset of state-of-the-art models demonstrates promising performance in capturing higher-order reasoning processes such as theory of mind and recursive planning. Together, these findings highlight the emerging potential of LLMs not only to imitate behavior, but to generate interpretable, mechanistic accounts of human social decision-making. Our results provide a roadmap for advancing LLM-based cognitive modeling toward human-expert-level theory construction.

This document describes the Statistical Analysis Plan (SAP) for PETRA, a randomised controlled trial in people with generalised anxiety disorder comparing pregabalin plus an antidepressant and standard care, with placebo plus an antidepressant and standard care, with respect to the primary outcome of the GAD-7 score at week 12.