Poster Session

As we navigate uncertain and changing everyday environments, we constantly need to make decisions and thus base these decisions on judgments about uncertain, changing information. One of the risks in this process is to "jump to conclusions" (JTC), i.e. to make judgments not just about how to act, but also about what is the case, based on little evidence. A classic task for studying JTC is the beads task. In it, one of two or more hidden urns with varying proportions of e.g. red or blue chips is selected at random, and chips are successively drawn from it with replacement. The task is to decide how many draws to see before making a judgment about which urn was selected ("draws to decision", DTD). Past research found that JTC - operationalised as few DTD - correlates with conspiratorial and racist beliefs, authoritarianism, political conservatism and need for closure. However, previous work often lacked a normative benchmark to define well-callibrated decisions; and seldom distinguished between discrimination ability and decision thresholds. We report an extension of two previous studies in which the DTD score was complemented with trial by trial judgments of discrimination ability, and compared to the normative benchmark of Bayesian belief updating. Although the above associations remained unchanged overall, people differed in the relation between their discrimination ability and decision thresholds, and they tended to be sensitive to the fact that fewer DTD are more rational when discriminability between urn proportions was higher.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

The circular diffusion model (CDM; Smith, 2016) models evidence accumulation as a two-dimensional Wiener process on a disk and jointly predicts response angles and response times in continuous-report tasks. A central question in decision-making research is whether behavioral biases reflect differences in how evidence is processed (drift-rate shifts) or differences in the initial state of the decision maker (starting-point shifts). Although the mathematical solution for the CDM with a non-origin starting point has been available for over a decade (Smith, 2016; Yin & Wang, 2009), no existing software provides computationally tractable estimation of the starting-point parameters. We derive the likelihood for the starting-point-extended CDM in a form suitable for gradient-based Bayesian estimation and develop a computationally efficient evaluation scheme based on a hybrid polynomial and asymptotic Bessel function approximation, achieving an eightfold speedup over a naive implementation. We validate the method by demonstrating accurate parameter recovery and apply the model to two empirical datasets: perceptual orientation judgments under predecision cues (Kvam, 2019) and consumer valuations with context effects (Izakson et al., 2024). We also discuss further extensions of the model, including trial-level regression on starting-point parameters, mixtures of drift rates, and across-trial variability. The open-source Python implementation is available via PyMC.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Research on individual differences aims to recover stable, person‑specific parameters from noisy experimental data. Hierarchical (multilevel) models improve reliability through partial pooling, but they are typically evaluated mainly at the population level. Consequently, the quality of individual‑level estimates often remains unchecked, leading to unstable individual estimates that produce attenuated correlations and conclusions that are driven more by modelling choices rather than true psychological variation. Therefore, evaluating models with inferences about individual differences must assess whether the model supports inferences about individuals. We propose a structured workflow for developing and evaluating Bayesian hierarchical models that treats validity of individual-level inferences as a main objective and is organised into three stages: (1.) before fitting the model, (2.) fitting the model, and (3.) evaluating the (fitted) model. (1.) Before fitting the model – We address defining the experimental design, selecting an initial model, and specifying informative priors for both population‑ and individual‑level parameters. We also discuss prior‑predictive checks to verify that the model can generate plausible heterogeneity and correlation structures. (2.) Fitting the model – We focus on parameter recovery, computational diagnostics, and model validation. We discuss simulation-based recovery analyses, different computational diagnostics and model validation (such as cross validation), with a focus on recovery and validation of individual-level estimates. (3.) Evaluating the model – From the wide range of options to evaluate the quality of a Bayesian model, we focus on posterior prediction, Bayesian fit indices, Bayesian model comparison using the Bayes factor, and sensitivity analyses. While many of the tools are well established, their integration into a workflow focused on recoverability, robustness, and interpretability at the individual level has, to the best of our knowledge, not yet been done.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Tasks in which humans make numerical judgments given numerical information have historically been analyzed using a linear regression framework (Karelaia & Hogarth, 2008). The regression approach involves the identification of information sources related to a specified outcome. The information sources are assigned weights, and the dot product of these weights with their respective values are then combined with a static intercept term to produce an estimate of the outcome value. However, it is also plausible that in some situations people may assign weights based on the rank-order of the information provided (Broomell & Wagner, 2024). To test if humans can use such an approach, participants were tasked with emulating the behavior of either 1) an order-weighted average (OWA) aggregator, 2) a linear-weighted average (LWA), or 3) a control condition with no clear averaging pattern. Results showed that participants were able to emulate the order-weighted average pattern even more effectively than both the control and the linear-weighted average. Nonetheless, participants were able to emulate the linear-weighted average more effectively than the control, suggesting that humans can use either type of averaging procedure. As a result, a model which allows for simultaneous consideration of order- and linear-weighted averaging procedures may be appropriate for modeling the dynamic nature of human judgment in numerical judgment tasks. Such an approach may have a stronger theoretical basis for predicting human judgment in dynamic environments.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Understanding how cognitive processes evolve over time is essential for modeling dynamic decision-making; however, dynamic cognitive models are often complex and computationally demanding. In this study, we propose a more accessible approach to capturing temporal dynamics using the EZ-diffusion model within a Bayesian framework. Rather than attempting to mimic the underlying data-generating mechanism, we incorporate random-walk priors into the EZ-diffusion model, allowing parameter estimates to evolve flexibly over time while remaining agnostic to the specific form of the generating process. Temporal dynamics were captured using sliding windows across trials, with EZ-diffusion parameters estimated within each window and subsequently aggregated to derive parameter trajectories over the course of the experiment. To evaluate the model, we conducted a simulation study in which the true data-generating parameters followed sinusoidal and exponential functions across trials. Parameter recovery was assessed across different window sizes and prior settings to determine whether the proposed approach could accurately track the underlying dynamic patterns. Finally, we applied the model to empirical data from Gulith et al. (2009), who investigated practice effects. By combining the simplicity of the EZ-diffusion model with random-walk priors in a Bayesian framework, this approach provides an accessible method for studying parameter dynamics over time in experimental settings.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

In interacting with the environment, we are forced to make decisions based on information processed through various channels—visual, tactile, auditory, etc. While evidence accumulation is more readily studied in visual and tactile modalities, due to the constraints of the visual field and sensory receptors in the skin—i.e., the only information available is that which the eyes have direct access to in front of you or that which makes contact with skin—auditory information is decidedly more difficult to study because one can perceive sounds from the entire surrounding environment. As a result, the ratio of auditory signal to environmental noise has relatively high variability. To understand how auditory features are perceived and evidence accumulated when making classification decisions, subjects were presented with audio recordings under various naturalistic background noise conditions and asked to decide which of four target signals were present. The time course and accuracy of subjects’ decisions, as well as the features of the auditory signal until a decision is recorded, give some indication about what features are important when classifying a signal as one target or another. It also provides information about what background noise conditions make classification decisions easier or more difficult. Human data are currently being entered into a geometric drift diffusion model to formalize the auditory evidence accumulation process under noisy conditions.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Stereotype Threat describes the negative impact on cognitive performance caused by the activation of negative social stereotypes. This study investigates the underlying cognitive mechanisms of gender-specific Stereotype Threat effects. Four online experiments were conducted with 1,164 participants, randomly assigned to either a Stereotype Threat or a control condition. To model the underlying cognitive processes, parameters of the drift diffusion model were estimated and effects on accuracy and response times were reported. Results showed no differences in accuracy as an effect of Stereotype Threat. However, women in the Stereotype Threat condition consistently responded more slowly and, in one experiment, exhibited a higher threshold separation when completing the mathematical task, indicating more conservative response tendencies, compared to men and the control group. The study addresses the need for potential interventions such as stereotype awareness to mitigate these effects and calls for further research into cultural and social influencing factors.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Learning from feedback is a fundamental mechanism by which individuals make decisions under uncertainty. An important open question is how people integrate prior information with repeated feedback to improve their decisions. To address this, we conducted an experiment in which we manipulated whether participants had structural knowledge about the decision environment: half of the participants were told the number of distinct outcomes each option could produce, whereas the other half received no such information and had to infer the outcome structure from feedback alone. To assess whether structural knowledge changes subjective uncertainty, participants also provided trial-by-trial confidence judgments. Structural knowledge reduced extreme preferences under partial feedback, shifting the proportion of risky choices toward indifference between options, but had no effect on choice under full feedback or on confidence. We then developed a Bayesian model that accounts for participants' initial beliefs about outcome probabilities before receiving any feedback. The model updates beliefs as outcomes are experienced and generates both choices and confidence judgments from each option’s outcome distribution and subjective uncertainty. The model captured choice and confidence patterns across different decision environments as well as individual differences. Allowing initial beliefs to vary improved model fit and partially accounted for behavioral differences between participants with and without structural knowledge. Together, these results suggest that structural knowledge shapes learning by altering prior beliefs, thereby linking prior information to experience-based risky choice.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

A model’s failure to recover its own parameters might be taken as evidence of its inherent limitations. However, parameter recovery can also depend on factors beyond the model itself, such as the structure of the stimulus space. Although the Generalized Context Model (GCM) is one of the most prominent models in category learning, little is known about how category structure affects its parameter recovery. To this end, we report a series of simulations with a naturalistic rock-category dataset, wherein each rock exemplar comprised eight separable dimensions. Specifically, we conducted a large-scale parameter recovery analysis of the GCM across all 435 possible pairs of rock categories. Simulation conditions were designed to reflect typical category-learning experiments with few exemplar presentations (1 vs. 2 vs. 5 vs. 10); to evaluate recovery under near-ideal sampling conditions, we also simulated conditions with a larger number of exemplar presentations (100 vs. 1000 vs. 10000 vs. 100000). Preliminary results indicate that parameter recovery varies systematically as a function of the distance between categories in a multidimensional scaling space, with greater category separation producing poorer recovery. Critically, these findings show that parameter recovery in the GCM depends on category structure and highlight the importance of evaluating parameter recovery when designing experiments that are intended to support inference from models of category learning. Moreover, these findings suggest that extant category-learning studies do not include enough repeated exemplars to enable successful parameter recovery, which can limit the insights that follow from formal models of category learning.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Quitting involves being forced to stop despite wanting to continue. In laboratory tasks, quitting is often conflated with stopping decisions. They are modeled as optimal threshold-setting that balances expected reward and cost on the task. In real-world contexts, these types of decisions are known to be dependent on psychological factors besides utilitarian ones. In this work, an observational study was conducted where endurance runners self-reported incidences of thoughts of stopping a run as and when they experienced them. The runners were instructed to report the time for each stopping thought while engaged in a running session and could decide to stop running whenever they wanted to. The endurance experience for each runner was controlled by a constant running pace measured from their subjective endurance levels. Psychological variables were measured using scales for personality traits (NEO-FFI), procrastination (Pure Procrastination Scale) and impulsiveness (Barrett Scale). Participants who quit their runs ahead of when they had wanted to stop displayed a distinctive cascading pattern of stopping thoughts, represented by two dimensions on a t-SNE space. An exponential rate model estimated each runner’s probability of quitting based on two parameters, rate of stopping thoughts and the growth rate of the decision variable. The rate parameter showed negative associations with procrastination and anxiety factors and positive association with the conscientiousness factor. A single t-SNE dimension loaded positively onto the rate parameter suggesting the influence of stopping thoughts on a subject’s quitting propensity. These results suggest that stopping thoughts bias the decision process towards quitting and need to be incorporated into models of stopping and quitting, for greater phenomenological discriminability.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Although expertise supports efficient performance in tasks like visual search, the benefits of expertise are often perceptually specific. To better understand the mechanisms by which expertise supports visual cognition, we studied the performance of both experts and novices in a music-based change detection task. Participants had to search for and fixate on a single note that changed between alternating images of musical scores, half of which were in an unfamiliar, 90-degree rotated orientation. We estimated parameters of a drift diffusion model by fitting it to accuracy and response time (the total time until the “flickering” target note was found amongst the surrounding, static foils). Experts showed increased response caution for rotated relative to upright stimuli and higher response caution than novices in the rotated condition, suggesting that, even when experts are confronted with unfamiliar displays, they are nonetheless better able to adapt their search strategies that novices. This apparent strategy difference was also evidenced by measures of eye movement behavior, where experts made significantly more fixations across rotated stimuli than novices.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Implementing complex cognitive experiments–particularly those requiring continuous task presentation or multi-window environments–typically demands substantial programming expertise. We present CogFlow, an open-source platform comprising a graphical Builder and portable Interpreter that enables researchers to configure and deploy experiments via human-readable JSON schemas, no coding required. The Builder provides drag-and-drop construction of experimental timelines from modular components (supporting paradigms like Random Dot Motion, Flanker, SART, N-back, Stroop, Simon, and others). Key innovation: unlike existing tools (PsychoPy, jsPsych, Lab.js), CogFlow supports continuous stimulus presentation and precise subtask scheduling via temporal constraints, enabling realistic multi-tasking environments. For example, our SOC Dashboard task simulates concurrent Security Operations Centre workloads across vigilance monitoring, log triage, and traffic analysis subtasks implementing paradigms like the N-back or Wisconsin Card Sorting paradigm. The JSON schema functions simultaneously as experimental controller and machine-readable specification, directly linking task configuration to computational modelling pipelines. Features include block-level parameter sampling, configurable response modalities, eye-tracking integration, and flexible deployment architecture: Cloudflare Workers for token-authenticated config and asset storage, JATOS integration for participant management, option of SharePoint export for enterprise environments, or standalone local browser deployment for in-house studies. Schema-validated specifications enhance reproducibility, enable systematic paradigm parameter exploration in simulation studies, and lower barriers for researchers in cognitive modelling, cybersecurity human factors, and applied attention research. CogFlow addresses a critical gap between GUI-based tools (inflexible) and code-based platforms (inaccessible to non-programmers).

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

In a dream social network made from a series of dreams of an individual, each person dreamt of is a vertex. An edge between two people indicates they occurred in a dream together. Such networks share properties with waking life social networks, such as short path lengths. In some social networks, highly connected people are highly connected to each other. In other social networks, moderately connected people are highly connected to each other, but highly connected people are not well connected to each other. In dream social networks, where are the highly connected people? The degree of a vertex is the number of edges incident with it. A measure of connection is the normalized rich club coefficient. For degree k, it is the number of edges between vertices of degree higher than k divided by the mean of that number over comparable random networks. A rich club effect occurs at degree k if its rich club coefficient is greater than 1. For dream social networks of four individuals a rich club effect was found for vertices of moderate degree, but not for vertices of high degree. A plausible explanation is that when two people occur together in a dream, their strength of connection in memory is refreshed to maintain it. There is little need to refresh strength of connection between high degree people. (One isn’t likely to forget that their brother and sister are associated.) So the limited time in dreaming sleep is allocated to refreshing moderate connections.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

How do we selectively attend to goal-relevant information and ignore irrelevant information? Some theoretical work has posited a “shrinking spotlight” of attentional focus to answer this question. However, the psychological process underlying such an attentional shrink is not specified by this model. Additionally, it is unclear why such a spotlight would reset its width between trials of a selective attention task, though this is an implicit assumption of the model. A key to resolving these issues may come from model comparison work, which generally finds that the shrinking spotlight model fits best when goal-relevant and irrelevant stimuli are spatially distinct, suggesting that spatial perceptual processes may be important to these dynamics. Drawing on this, we developed a novel model combining eye movements with simple drift diffusion processes that—across a number of trials—produces patterns of data consistent with a shrinking spotlight. Additionally, this new model was a good fit to empirical eyetracking and behavioral data from a flanker task. Together, these results suggest that the effects of fixations aggregated across trials may produce what appears to be a shrinking attentional spotlight without any such latent psychological process actually occurring.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Subjective time perception is integral to navigating everyday life, yet understanding remains limited. Current theoretical and methodological difficulties invite a computational approach that can manage paradigm complexity and clarify how neural dynamics create individual time estimates and influence resulting behavior. With this computational goal in mind, we implemented a dynamical systems model that utilizes bistable units to simulate time accumulation, producing a biologically plausible method for tracking elapsed duration. The units form a hierarchical system that relies on a pacemaker to propagate activity, resembling the theoretical mechanism of an internal clock, where elapsed time is calculated from unit cycles during an interval. This system is also sensitive to overestimation and underestimation. The model was assessed using data from a time interval study showing the presence of anchoring effects on time perception. The model fit well for individual participant estimates and very well for anchoring effects on duration estimates. These fits indicate the presence of individual differences that are not fully explained by experimental groups and show our model’s ability to represent and predict time while using biologically plausible mechanisms. Preliminary analyses indicate that the model bias parameter accounts for anchoring influence, and the model gain parameter (internal clock speed) reflects individual differences in human participants. This suggests that while the anchor influences internal time estimates, individual differences beyond experimental groups arise from an internal clock. To further explore individual differences, a future model iteration will include a neural pathway to create a complete, biologically plausible account of subjective time perception.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Similarity assessment is central to human conceptual behavior, including concept learning, categorization, problem solving, and decision making. While many approaches to similarity focus on comparisons between individual objects, others emphasize the similarity between relations among objects. Building on this latter relational perspective, previous research has investigated relational similarity judgments on dimensionally consistent stimuli as a function of Boolean concept structures while implementing a formal mathematical model (Vigo, 2015; Yadav, 2024; Vigo et al.,2026). The present study tests Vigo’s (2015) model using relational similarity judgments on the 3-2-3 family of Boolean concept structures under both dimensionally consistent and dimensionally inconsistent assignments. Dimensionally consistent assignments use identical sets of objects across comparisons (e.g., bugs vs. bugs), whereas dimensionally inconsistent assignments use different object sets (e.g., bugs vs. flasks). Each concept structure in this family is defined over three binary dimensions and three objects. Participants rated the relational similarity of paired concept structures. Preliminary analyses show that the model accounts for approximately 70% of the variance in similarity judgments for dimensionally consistent stimuli, with a substantial reduction in explained variance for dimensionally inconsistent stimuli. This decrease in explanatory power suggests that relational similarity judgments are sensitive to dimensional consistency. Because the current model does not incorporate the degree of consistency in dimensional assignments, these findings point to an important limitation and motivate the need for model extension. More broadly, the results highlight the importance of dimensional consistency in shaping similarity judgments and in refining formal accounts of relational similarity.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Attentional drift diffusion models (aDDMs) provide a mechanistic account of how visual attention influences value-based decision-making. However, full Bayesian inference has been computationally intractable due to trial-wise unique within-trial drift rate dynamics driven by gaze patterns. This intractability forced researchers to rely on time-averaged drift approximations (TADA), whose parameters were treated as if they reliably mapped onto aDDM parameters. In recent work, we derived an algorithm that speeds up proper aDDM likelihood computation by three orders of magnitude (Liu et al., 2026). We further demonstrated that TADA methods are statistically inconsistent and showed with targeted counterexamples how they can bias conclusions about attentional effects when the true data-generating process is the aDDM (Liu et al., 2025). Here we present an application and translation of this research. First, we provide a software integration bridging the gap toward practical application: our methods are now available to be used naturally via the probabilistic programming library PyMC. We extended our algorithm to an autodifferentiable JAX implementation, enabling gradient-based MCMC methods like NUTS. Second, we present a comprehensive, systematic Bayesian parameter recovery study comparing TADA against proper aDDM inference. This provides clear guidance for experimentalists on when TADA might remain appropriate and pinpoints specific risks to scientific conclusions when it is applied for computational convenience. Third, we conduct a re-analysis of two empirical datasets across species to showcase the effects on parameter inference in realistic settings.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

How can cognitive models account for the complex interplay of communicative and non-communicative processes in collaboration? This paper introduces a blueprint for an ACT-R dialogue manager capable of navigating the interplay between communication and action in the dynamic task Co-op Space Fortress. Our approach integrates language processing, valuation, and discovery into a unified cognitive framework. Central to the model is an information state, which includes task knowledge, shared plans and transactive memory. Agents actively maintains this information state through the selection of specific dialogue acts. By defining a typology for these acts, we provide a path toward developing agents that collaborate with humans through natural, context-aware dialogue.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Autonomous mobile robots operating in real-world environments must handle incomplete knowledge and dynamic obstacles. This work presents an empirical study of an integrated cognitive model that combines ROS 2 actions and sensory data with the cognitive architecture ACT-R (pyactr) for a Turtlebot4 robot. The robot operates in a discrete matrix-based environment containing both solid and fake obstacles, while initially only assuming the existence of unknown obstacles. Path planning is performed using the A* algorithm, and encountered obstacles are evaluated through physical interaction. Based on repeated collision experiences, the robot adapts its navigation strategy by switching from shortest-path strategy to a risk-aware one that avoids all obstacles, even unknown ones. Experimental results demonstrate that the cognitive extension enables adaptive strategy selection and memory-based learning of obstacle properties. These findings highlight the potential of cognitive approaches for interactive task learning in mobile robots.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Computational models are powerful tools in cognitive science. They provide formal, testable accounts of human behavior and allow for nuanced interpretations of behavioral data that go beyond raw behavioral metrics. Yet this approach faces significant reliability and validity issues. Beyond well-known issues of reliability, generalizability, and statistical significance, we highlight a more concerning problem: many modeling results, even when robust, may simply be invalid. We present results from a simulation-based study showing that even basic, widely used models can “go rogue”: break out of intended constraints, generate unintended behaviors, and produce outright false conclusions. We present three simple cases in which this occurs, illustrate how and why these failures arise, and discuss the serious implications for modeling practice and interpretation. We conclude by emphasizing that, unless researchers adjust current modeling interpretation conventions, computational modeling-based studies risk systematically producing false results.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

This paper presents a computational, biologically plausible model of dual-process metacognition. Although Type 1 and Type 2 distinctions in metacognition have been widely discussed, and recent work has begun to model some components computationally, the mechanisms linking these processes remain under-specified. We argue that dual-process metacognition is best understood as an emergent phenomenon produced by the coordination of modules and information types, rather than as two architecturally separate metacognitive systems. Using a computational cognitive architecture, we illustrate these dynamics in metamemory and metareasoning by specifying how implicit signals can trigger the retrieval and application of explicit metacognitive strategies. The model offers a parsimonious account of how dual-process metacognitive control may be realized in human cognition.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

As automated systems in aviation are expected to shift toward agentic teammates, pilots’ cognitive work may increasingly involve monitoring, verification, and reliance decisions rather than direct task execution. To investigate this transformation, we present a cognitive model of a pilot interacting with an autonomous agent during takeoff, grounded in human–automation interaction research. The model is implemented in QN-ACT-R, a queuing-network extension of ACT-R, and integrates a SEEV-based visual attention mechanism to govern gaze allocation across cockpit Areas of Interest. Verification is modeled through competition between cross-check and non-cross-check productions whose utilities update via ACT-R utility learning. Cross-check behavior is initialized under neutral trust and updated as a function of interaction outcomes: reliable autonomy gradually reduces verification probability (complacency), whereas detected failures increase crosscheck rate (distrust), capturing dynamic trust calibration. The model also implements reactive reliance: the pilot does not wait for automation but proceeds with task execution and integrates support opportunistically if it becomes available before decision completion. We evaluated the model in simulations manipulating autonomy reliability and support delays. High reliability conditions led to progressive reductions in verification frequency, while detected autonomy’s failures triggered sharp increases in monitoring and longer execution times. Delayed support reduced mean decision time only when the aid arrived before task completion, consistent with reactive reliance. We are currently conducting Human-In-The-Loop Simulations using the same experimental conditions. At ICCM 2026, we will present initial empirical comparisons of predicted and observed verification frequency, support integration, and time on tasks.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

How can we better represent the impact of sociocultural structures on decision making in computational cognitive models? Modeling this impact requires traversing multiple levels of semantic representation, however it is not immediately clear to a modeler which levels of representation are most salient to a given situation. Though large language models and cognitively grounded corpus models can represent broad semantic associations through co-occurences, the role of self representations in memory should be accounted for to determine how cultural associations shape decision making. We propose a declarative memory system to be used in the ACT-R cognitive architecture that represents semantic associations at multiple levels via a vector-symbolic autoencoder. We use a simple HRR operation to encode episodic memories differently from semantic memory vectors extracted from text to produce a final chunk activation for a memory request. We use ACT-R cognitive models of a racially contextualized implicit association test (IAT) to test this new declarative memory system.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Computational models of brain-related phenomena often depend on latent representations whose geometry determines what structure can be inferred from complex data. In glioblastoma, tumours with similar molecular profiles frequently exhibit divergent imaging characteristics and clinical trajectories, suggesting intrinsic phenotypic structure not captured by standard labels. This work formalizes unsupervised radiomic clustering as a latent-state inference problem, examining how representation geometry influences the detectability of meaningful subgroups. Using multi-parametric MRI data from the UPENN-GBM cohort (n = 599), 432 radiomic features were extracted across enhancing tumour, necrotic core, and peritumoral edema regions. Linear Principal Component Analysis (PCA) and non-linear Uniform Manifold Approximation and Projection (UMAP) were systematically compared using topology-preservation metrics and clustering stability analyses. While PCA preserved global variance and pairwise distances, it retained the high-dimensional concentration of measure that limited density-based clustering. In contrast, UMAP introduced controlled topological distortion that increased structural contrast and facilitated manifold unfolding, producing a task-optimized latent space for HDBSCAN clustering. The resulting phenotypes were structurally stable and exhibited significant radiomic differentiation, with approximately 90% of features differing between clusters (p < 0.01). Kaplan–Meier analysis demonstrated these imaging-defined states were independently prognostic for overall survival (p < 0.05), remaining statistically independent of age, sex, IDH1 mutation status, and MGMT promoter methylation. This framework provides a principled foundation for representation-aware modeling and explainable AI in neuro-oncology.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Typing layouts are often evaluated using full prototypes, but many research layouts rely on specialized hardware, sensors, or non-public systems, which makes controlled comparison difficult. We test whether cognitive models can serve as a practical evaluation tool for comparing typing layouts under the same architecture and parameter settings. We developed four ACT-R models with layout-specific encodings: QWERTY Bimanual typing, Hunt-and-peck (one-finger) typing, FastType mid-air gestures, and Abacus gesture typing. Motor and visual modules were disabled to focus on representational differences in declarative and procedural knowledge. Each model typed 10,000 words generated from random sentences. We computed average words-per-minute (WPM) in 10 blocks of 1,000 words to produce learning curves and long-term performance estimates from later blocks. When available, model WPM ranges were consistent with reported human data. We also report on the number of declarative chunks and production rules to summarize declarative and procedural knowledge across layouts. These results support the feasibility of cognitive modeling for early-stage layout evaluation.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

An extensive literature demonstrates that human probability judgment deviates systematically from the rules of classical probability theory (CPT). Violations of CPT include the con- junction fallacy, disjunction fallacy, sub-additivity, violations of binary complementarity and classical probability identities. Despite their focus on breadth and integration, cognitive ar- chitectures have not been tested in the domain of probability judgment. To fill this gap, we developed an ACT-R model in which probability judgments are generated through repeated sampling from memory. In this model, violations of CPT arise from a combination of partial matching and retrieval failures. We tested the ACT-R model against a rich dataset from Huang et al. (2024) consisting of two sets of 78 judgments based on three binary events each. ACT-R was able to capture the gen- eral pattern of violations of classical probability identities, the law of total probability, and binary complementarity, but also showed consistent evidence of under/over estimation in some cases.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Using mathematical models to examine human behavior is part of the growing discipline of cognitive modeling. The Drift Dif- fusion Model (DDM) is a mathematical model that displays human behavior in binary reaction-time tasks. The maximum likelihood approach is one of many methods that have been developed to estimate and evaluate parameters of the DDM. A direct optimization algorithm that is suitable to find minima in a large search space is the Particle Swarm Optimization (PSO). The algorithm is able to locate the negative log-likelihood min- imum for DDM parameter estimates. However, it has rarely been used for this purpose. The realm of this paper is to show how different applications of PSO algorithm can fit DDM pa- rameters, using replicable and accessible pseudo-code.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Spreading activation and sequential sampling are fundamental and ubiquitous elementary mechanisms in computational models of cognition. This paper develops the core of an axiomatic theory of evidential structures with the goal of identifying a minimal and unifying set of foundational assumptions for cognitive models making use of these mechanisms. It is argued that bodies of evidence, as ordinarily understood, satisfy the axioms of closed extensive structures and that they may therefore be represented on an additive ratio-scale. Furthermore, a notable decomposition of the relevant ratio-scale representation is presented, under which spreading activation and sequential sampling emerge as elementary computations on bodies of evidence.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

This paper presents a computational cognitive model of repeated strategic interaction in Rock–Paper–Scissors (RPS) with bounded learning constraints. Although classical game theory provides a framework for strategic choice in simple zero-sum games, human behavior departs from these predictions and instead reflects history-dependent, instance- based learning. Prior work demonstrates that different cognitive architectures converge on similar qualitative dynamics, yet the mechanisms that constrain learning are left unexamined. We explore this gap by adding bounded utility and find that this produces novel emergent behavior not observed in earlier models. Using epiNgen, an ACT-R–inspired architecture, we replicate classic Lag1/Lag2 effects and show that imposing utility floors and ceilings can qualitatively reshape interaction dynamics. These findings highlight bounded utility as a simple, neurally plausible constraint that can substantially change learning and outcomes in simple zero-sum games.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Most computational models of social behavior begin from the assumption that agents are fundamentally self-interested, and that cooperation or compassion must be engineered through additional mechanisms such as incentives, norms, or explicit social preferences. This assumption is often treated as an evolutionary fact. The present paper argues that it is instead a modeling choice—one that significantly constrains the space of possible social dynamics. Drawing on the Free Energy Principle, the paper proposes an alternative motivational primitive: sensitivity to instability within a system’s representational field. On this account, undifferentiated regulatory responsiveness precedes explicit self–other differentiation, while self-interest emerges through representational structuring and precision-weighted narrowing. This reverses the usual explanatory order. Cooperation does not require special-purpose altruistic machinery layered onto a selfish core; rather, selective self-interest requires additional mechanisms that restrict an initially broader regulatory field.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

When participants are asked to estimate the time interval between two events, they tend to give smaller estimates when the first event consists of a voluntary action. This temporal binding effect is one of the most widely used implicit measures of the sense of agency. However, despite its widespread use, the cognitive mechanisms that generate this effect remain unclear. The rational analysis of memory instructs us that the strength of a memory trace should track the recency and frequency of its usage. Furthermore, emotional arousal is known to produce stronger memory traces. In this study, we therefore hypothesize that participants perform interval estimation by comparing the availability of experimental events in memory. In this light, temporal binding effects may be attributed to differences in emotional arousal at the time of encoding due to the type of event. For instance, voluntary actions may be accompanied by slightly higher arousal, leading to stronger traces of the first event relative to the involuntary case. We present an online interval-estimation paradigm aiming to test this hypothesis as well as a computational model of this process, which predicts that the size of the binding effect is commensurate with the magnitude of the difference in arousal between the two conditions.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Many environments are racialized, even if such racialization occurs beyond awareness, and this racialization extends to environments in which human-AI interaction occurs. We investigate how the racialized framing of AI systems impact performance on a human-AI collaboration task with the help of an ACT-R model. Participants (N=1008) played a cooperative decision-making game (Pigchase) with AI agents that were explicitly racialized, followed by a post-game survey designed to understand their strategies. Instead of just the typical phenotypical (e.g., visual) representations of race, we used both knowledge-based and visual methods for racialization, such that we could understand the impacts of latent racialized representations on behavior. We used a 3 x 7 experiment design, where we see how participants' self-identified race (White, Black, Other) interacted with both visual and verbal cues for racialized framing of AI. To develop a cognitive process level understanding of behaviors exhibited during the experiment, we built instance-based learning ACT-R model that incorporates social notions of self. We use this model to understand pro-self or pro-social attributes of the participants, trust accumulation over trials, and differences seen from survey responses. We show that the model is worth taking seriously.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Decades of work investigating terror management theory (TMT) have produced strikingly mixed experimental and meta-analytic results supporting the existence of mortality salience (MS) effects—the finding that reminders of death produce compensatory behaviors. Little is known, however, about whether potential theoretical mechanisms might explain this heterogeneity. We explored some of these mechanisms by developing an agent-based cognitive model designed to operationalize three potential ways of responding to MS: bolstering one’s beliefs, derogating the beliefs of others, and avoiding engagement with MS entirely. We ran 10,000 independent simulations investigating the consequences of these MS response styles. Results showed that these response styles could lead to distinct outcomes in defensiveness: individual differences in responding produced variability in the extent to which worldview and self-esteem defenses were activated. Such individual differences may help explain some contradictory empirical findings in the TMT literature, highlighting a need for greater precision in how the core components of TMT and other theories of psychological defense are specified.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Humans can learn and develop stable categories in noisy environments, reflecting the stability-plasticity phenomenon. One must remain stable under small perturbations (noise) but must adapt if important information is given. This poses a challenge for cognitive models like recurrent associative memory neural networks. Such memories are built to store information as an attractor. This allows it to be noise-tolerant and to complete missing information. Currently, it has no mechanism to distinguish if a pattern is relevant or not. As such, each difference is seen as a different pattern. Over time, the incorporation of variations will saturate its memory, and the network will lose its attractor associative property. To address this limitation, we propose an adaptable distributed filter that produces invariant representations of quantity. The network will only store a pattern as a new memory when the difference is more than the predetermined threshold. Otherwise, the network will partially incorporate the patterns with its recalled memory. The network was tested on 26 binary correlated patterns, where each pattern was noised before being presented to the network. Performance was assessed by computing the number of attractors and their radii of attraction. Results show that the network can discard noise and develop the correct number of memories if the variability with noise (within) is not greater than the variability between the patterns and if the memory load is medium (~20%). Therefore, using an invariant quantity representation given by the filter, the network is closer to solving the stability-plasticity dilemma in a distributed fashion.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

To be complete, a model of communicative behav- ior requires a framework describing overall rational action, of which com- munication is a part. We use a model from Bayesian deci- sion theory: A partially observable Markov decision process (POMDP). A POMDP describes how an agent, human or ar- tificial, should revise its beliefs about the world given its ob- servations, and defines rational action as one that maximizes the agent’s expected utility given its beliefs. We build on the POMDP framework by adding to it the theory of mind (ToM) - a collection of variables describing another agent and al- lowing predictions about its behavior. Further, we formalize the agents’ capability to communicate, forming communica- tive interactive POMDP (CIPOMDP). In our framework the ToM models are also (CI)POMDPs. This allows us to use the known Bayesian decision theory techniques to describe how agents model others while making their own rational decisions, and it leads to natural nesting of ToM models. In particular, the processing of a communicative act by a listener consists of Bayes update of listener’s beliefs given new information and their ToM of the speaker. The update is general: The listener does not need to assume that the speaker is well informed and sincere; the Bayesian math quantifies these elements precisely. The framework also applies to the speakers: They use the ToM they have to simulate the listener’s belief update to choose the communicative act that maximizes their, i.e., the listener’s, ex- pected utility.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

To better understand how environmental adversity shapes decision-making, we propose a Contextualized Rational Inattention (CRI) model of cognitive control. Traditional approaches often model structural disadvantage phenomenologically, treating it as random noise or fixed linear biases. In contrast, drawing on Ecological Rationality and Scarcity theory, we formalize environmental adversity as a structural constraint on an agent's information-processing channel capacity. Using a dynamic optimization framework, we model the marginal cost of mutual information as a direct function of the agent's context. A computational simulation study demonstrates that while standard descriptive models fail to generalize across shifting environmental base rates, the CRI model successfully predicts optimal adaptation. We apply this framework empirically to $N=32,276$ trials of the Dimensional Change Card Sort (DCCS) task, using the Area Deprivation Index (ADI) as a measure of objective scarcity. Results confirm that neighborhood deprivation significantly increases the information-processing cost of cognitive control. We discuss these findings in the context of the predictive-versus-explanatory trade-off in mathematical modeling, arguing that heuristic reliance in high-adversity contexts reflects mathematically optimal bandwidth allocation rather than cognitive deficit.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Quantum states are a powerful way for complex information representation, particularly with the involvement of rotations and entanglement. We prepared a challenging minimal benchmark for assessing the scalability of a transformer-like model. For this purpose, a small dataset of state vectors based on 2-qubit ansatz circuit was prepared to handle the quantum depth with three levels from shallow to deep. The generated dataset was used to train and evaluate two types of models, feed-forward networks with or without self-attention module, also varied in their parameter growth, by reducing the number of hidden dimensions from 12 to 8 or the number of attention heads from 4 to 1. The performance was measured in terms of speed for learning in reaching a designated threshold for mapping inputs to targets. The simulation results showed a gradual performance drop as the depth level increased. The hidden dimension and attention heads could compensate each other to some extent, i.e., attention facilitated deeper mappings without increasing hidden dimensions.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Explaining expert performance in complex cognitive–perceptual–motor domains remains a central challenge for cognitive science. Experts appear to respond both rapidly and accurately in environments that are noisy, high-dimensional, and time-constrained, yet they do so with the same biological hardware as novices. This paper presents a computational account of expertise grounded in information theory and rate-distortion principles. We model skilled performance as bounded-optimal perception–action control, in which an agent must efficiently compress world states into internal representations and map those representations to utility-maximizing actions under explicit information constraints. The model formalizes two coupled sources of cognitive cost – representational complexity and policy complexity – using mutual information terms, and optimizes an objective that trades off expected utility against those complexity costs. Using a two-stage Blahut–Arimoto optimization procedure, we examine how performance changes across paired representational and action-capacity limits in a simplified competitive task domain. Results show that high performance depends on coordinated increases in representational differentiation and policy selectivity rather than on unconstrained computation. The paper also introduces an information-to-time bridge linking modeled information costs to human reaction-time limits. Together, the framework offers a tractable and cognitively interpretable account of expertise as improved value per unit information rather than faster raw processing.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).

Are brains like computers? The answer depends on what you mean when you ask. Brains are probably not structurally organized like a smartphone, but they compute, since computing is, by definition, something humans can do (Turing, 1937). Are computers like brains? Neuromorphic computing systems (Zhang et al., 2020) claim to be, but it seems like you can have computers that aren’t. In fact, you can even have one made out of neurons. We present an interpreter for the Forth programming language made out of model neurons, called North. Cognitive modellers are often interested in neuromorphic computing, because it promises a brain-like substrate in which one might validate cognitive models. But neuromorphic computing is Turing-complete (Date et al., 2022), so potentially any computer model or software, like a Forth interpreter, could be neurally intstantiated. When one makes an effort to make a model brain-like, one must pay careful attention to the analogy one is drawing between the brain and cognition, and one’s model. Our interpreter is interesting; it uses a vector-symbolic architecture (Gayler, 2004) and the Neural engineering framework (Eliasmith & Anderson, 2003), and an encoding loosely based on Tomkins-Flanagan and Kelly (2025) to allow for neurons to execute code that is human-readable and conventionally programmed. But, it does not compute in any way we would dare to relate to how the brain does things. That said, it opens the door to building more complex or unconventional computing systems that may have a better claim to doing things “like” the brain.

This is an in-person presentation on July 18, 2026 (14:40 ~ 16:20 EDT).