Poster Session

Configural information provides relational information that facilitates individuals' visual processes. Symmetry is one type of configural information, but it is unclear when, if ever, perceiving multiple lines relies on symmetry. If people perceive pairs of lines using their relative slope (e.g., if they are symmetric), then we would not expect independent nor separable perception of the individual line’s slope. This study uses General Recognition Theory (GRT) to examine slope perception when the varied slopes lead to symmetrical or asymmetrical pairs of lines. GRT enabled us to examine two types of perceptual relationships between pairs of lines: failures of perceptual separability and failures of perceptual independence. The former indicates perception of one dimension does not vary across levels of another dimension. Perceptual independence focuses on the perceptual effect of single stimuli: it is violated when the perceptual effects on each dimension are stochastically dependent . From the perspective of configuration processing theory, we expected violations of both perceptual independence and perceptual separability across all conditions because we assumed people would rely relational information. Indeed, tests of perceptual separability and independence all failed for all participants. Furthermore, people tended to make more errors within responses with the same symmetry status than would be predicted by a local model. Together, these results indicate separate, independent processing of the lines is likely not occurring; instead, people do seem to rely on symmetry in our task.

Evidence accumulation models (EAMs) are valuable tools to understand the cognitive processes underlying human decisions and to predict choice-RT distributions. However, EAMs typically do not account for how humans learn and adapt to feedback. Recent studies suggest that learning can be incorporated into EAMs, providing an adaptive account of speeded decision making. This has been established in tasks which focus narrowly on reward learning, leaving it unclear how broadly such adaptive EAMs could apply. In this study, we test an adaptive EAM of a more complex decision-making task. Specifically, participants performed an air traffic control conflict detection task in which they decided whether aircraft would remain safely separated during their flight. Participants were encouraged to learn the probability of aircraft conflict, which was varied across nine blocks. There were strong effects of conflict probability on choices and RT distributions, as well as a “recency effect”, implying strong influence of recent experience. We attempted to account for such effects with a racing linear ballistic accumulator model with classic ‘delta rule’ learning affecting trial-by-trial evidence-accumulation parameters, such as thresholds to make decisions. The model provided a promising account of key adaptive effects. Incorporating adaptive processes has potential to increase the predictive scope and utility of EAMs.

Grounded in resource theory and the notion of limited multitasking capacity, this study investigates how visual search tasks impact auditory working memory performance, highlighting the competition for shared cognitive resources between visual attention and auditory processing. Further, the primary task load of one modality influences the performance of the second task in another modality. In this work, we investigate how the difficulty of the visual search (VS) task influences ones’ ability to complete an auditory Sternberg working memory (WM) task. Specifically, we assess how behavior and cognitive efficiency changes in a single bound (go/no-go) aural WM task at a constant difficulty (set size = 6) when in context of a single bound (go/no-go) VS task of varying difficulty (i.e., size of search area and number of items) compared to the isolated context. Using evidence accumulation models, we capture parameters (such as drift, threshold, and non-decision times) for each task, at each level of difficulty (VS-low, -medium, -high) and in both single-/dual-task contexts. We find that performance decreased, and parameters of the cognitive models changed in both the WM and VS tasks when completed in multi-tasking context versus isolation. These decrements were heightened as the VS task difficulty increased. By understanding the cognitive resource limitations involved in multitasking, designers can optimize visual interfaces to minimize interference with auditory tasks. Additionally, workload protocols can be refined to reduce cognitive overload, enhancing efficiency and accuracy in tasks that require simultaneous visual and auditory processing.

Stereotype threat refers to the negative impact on cognitive performance caused by the activation of social stereotypes. This study investigates the cognitive mechanisms underlying gender-specific stereotype threat effects using a drift diffusion model analysis. An online experiment with 612 participants (307 men, 305 women) examined performance in a mathematical task and an emotion recognition task under either a stereotype threat or control condition. The drift diffusion model was used to estimate parameters reflecting decision-making processes, such as speed of evidence accumulation (drift rate) and response tendencies (threshold separation). Women performed worse than men in the mathematical task, while exhibiting a lower threshold separation. However, results showed no significant differences in accuracy due to stereotype threat. Nevertheless, women under stereotype threat exhibited increased threshold separation in the mathematical task, indicating more cautious decision-making compared to men and the control group. This aligns with regulatory focus theory, suggesting that stereotype threat induces a prevention-focused mindset, leading to more conservative response strategies. No substantial changes in drift rate were observed, contradicting theories linking stereotype threat to working memory depletion. These findings highlight that stereotype threat influences response strategies rather than overall performance, suggesting that traditional accuracy-based measures may not fully capture its cognitive effects.

To assess subjective value, people dynamically allocate attention to different attributes of their available options. As a measurable proxy for attention, eye fixation offers an external window into what information is attended and incorporated into preferences. Some preferential choice models, such as attentional drift diffusion models, incorporate gaze information to better understand both choice data and fixation patterns. However, the connection between gaze and other tasks, such as pricing, is not well understood. To shed light on this question, we created and tested a model for multi-attribute pricing, applying it to a risky-intertemporal pricing task. In this task, 31 participants were to respond how much they would pay to acquire delayed and/or risky payoffs (e.g., 30% chance of $20 in 100 days). Participants showed a strong preference for looking at the payoff attribute first and switched gaze mostly between payoffs and either secondary attribute (risk or delay), suggesting anchoring based on payoff during information sampling. This behavior was captured well by a model integrating process-based interaction of attention, information accumulation, and anchoring. We show that the diverging patterns of attention allocation between pricing and choice models can help explain preference reversals across preference elicitation procedures.

Understanding the cognitive mechanisms underlying urgency is critical to developing evidence accumulation models. To date, most investigations of urgency have focused on collapsing decision boundaries in the two-boundary diffusion decision model. In this study, we demonstrate that when data generated by an accumulator model is fit using a diffusion model (or vice versa), it can lead to incorrect conclusions about within-trial decision boundary dynamics parameter mis-estimation. First, we show that even when choice and reaction time data are generated by an accumulator model without a collapsing decision boundary, comparing diffusion models with and without collapsing boundaries can erroneously suggest the presence of a collapsing boundary. Second, we applied both diffusion and generalized accumulator models – with and without collapsing boundaries – to real data. In these real datasets, we observed (a) that diffusion models with collapsing boundaries were favored over ones without, but (b) that generalized accumulator models without collapsing boundaries provided the best overall fits. Additionally, we found that the two types of models sometimes provided conflicting conclusions about the effects of experimental manipulations on drift rate and decision boundary, which are believed to have distinct psychological meanings. Our findings highlight the importance of considering representations of evidence when inferring decision boundary dynamics, and in some cases suggest that collapsing boundaries may not be possible to identify.

Environmental pollution and climate change are sustainability issues that modern society must address. Therefore, the importance of decision-making "for future generations" has been emphasized (Saijo, 2017). The tendency of individuals to consider the benefits of others is called Social Value Orientation (SVO) (Van Lange, 1999). According to Inoue et al. (2021) and Miki & Shimizu (2023), there is no significant difference in the degree of prosociality when comparing SVO toward present others and SVO toward others in the future (e.g., x years later). However, it has been shown that individuals tend to prefer more egalitarian distributions when distributing rewards to present others. When imagining "future generations", people may envision abstract future individuals rather than specific people living x years later. Thus, this study aims to examine whether differences in the imagined "future generations" affect SVO. In this study, SVO was estimated using the Equal Equivalent Measure (Mizuno & Shimizu, 2024), and a model comparison using Bayesian factors was conducted. The experiment involved three conditions: the general future condition, where participants imagined "future generations"; the 40 years future condition, where participants imagined "others 40 years later"; and the present condition, where participants imagined "present others". The results showed that a model assuming SVO to be different across all three conditions had a better fit to the data than a model assuming similar SVO for both future generations and others 40 years later. Details of the model comparison and SVO will be presented on the day of the presentation.

In recent years, growing social divisions have raised concerns. Many people hesitate to discuss controversial topics for fear of criticism, which hinders constructive debate (Tsuda, 2024). On social media, extreme viewpoints tend to dominate, while moderate voices remain unheard, creating a misleading impression of public opinion. This pattern has been documented in discussions on political issues in the United States, such as abortion, climate change, and gun control (Robertson, 2024), and similar trends have been observed in Japan (Yamaguchi, 2022). However, this reluctance to engage in political discussions is not confined to online spaces. Research on political discourse has examined both private and public settings (Yokoyama, 2023), and studies suggest that political discussions are relatively rare in Japan (Ikeda, 2016). Yet, whether the same discrepancy between actual and expressed opinions exists offline remains uncertain. Previous studies (Robertson, 2024; Yamaguchi, 2022) primarily relied on Likert scales, which are susceptible to response biases (Van Herk, Poortinga, & Verhallen, 2004). In response, recent research has explored the use of the Thurstone method (e.g., Kashiwabara & Shimizu). This study constructs a scale using a multidimensional unfolding approach with the best-worst scaling method to assess attitudes toward selective marital surname systems and patterns of opinion expression in both online and offline contexts. The goal is to provide a clearer understanding of biases in expressed opinions.

The hierarchical diffusion model (HDM) is widely used in cognitive modelling. However, sample size planning for HDM (based on power analysis) remains challenging, as it typically relies on simulation-based approaches that require substantial computational resources. This study addresses this limitation by integrating composite likelihood with amortized Bayesian inference to improve computational efficiency. Composite likelihood approximates the full likelihood by focusing on subsets of data and events, allowing the likelihood of HDM to be partitioned into smaller units. Such an approach can speed up training phase and take advantage of the fast inference phase in BayesFlow. To evaluate its effectiveness, we conducted parameter recovery studies using simulated data from both simple and hierarchical DDMs, comparing composite and full likelihood estimation in BayesFlow and Stan (as a benchmark). Preliminary results suggested that composite likelihood achieves parameter recovery comparable to full likelihood in both BayesFlow and Stan, while significantly reducing inference time in BayesFlow compared to Stan. These results highlight the potential of composite likelihood for efficient power analysis in HDM, providing a promising direction for future research.

Perceived threats from immigrants tend to increase anti-immigration attitudes among citizens of host countries (Esses et al., 1998, 2001). However, most previous research has focused on attitudes measured by scales, with few studies employing incentivized experimental games to measure behavioral responses. In Kashihara and Shimizu (2024), a new experimental game was developed based on the preemptive attack game used in outgroup aggression studies (Simunovic et al., 2013) to capture reactions to potential attacks by immigrants. This incentivized game, played in pairs, offers three choices: attack, defend, or keep. In the present study, participants who were citizens of the host country (N = 301) were divided into two conditions: one group played the game against another host-country citizen (host condition), and the other group played against an immigrant (immigrant condition). Additionally, this study employed computational modeling to analyze the behavioral data. Specifically, we derived a decision-making model in which participants calculate expected payoffs for themselves and their counterpart—weighted by altruism based on Ackerman et al.’s (2016) social preference model—while factoring in the perceived probability of an attack (threat) from the opponent. We then estimated threat and altruism parameters by performing Bayesian modeling on the behavioral data. At the behavioral level, participants were more defensive toward immigrants under certain experimental conditions. Parameter estimation also revealed a heightened threat motivation in the immigrant condition. This study offers a novel method for measuring perceived threat from immigrants through an experimental game paradigm and discusses implications for future research.

Multiple tasks have shown that listeners rely on lexical memory - their knowledge of words - to aid the perception of unclear speech (Ganong, 1980; Ishida et al., 2016; Warren, 1970). Giovannone and Theodore (2023) investigated the test-retest (TRT) reliability of three tasks – the Locally Time-Reversed Speech task (LTRS), the Ganong task, and the Phoneme Restoration (PR) task – using lexical reliance levels estimated from linear mixed models (e.g. random slopes). They found moderate reliability in LTRS and Ganong and low reliability in PR. Improvements in the TRT reliability of several cognitive tasks have been found using generative modeling, which models participants’ response behavior (Guest & Martin, 2020; Haines et al., in press). The current study evaluated whether generative modeling could also improve the TRT reliability of the three lexical reliance tasks. We built task-based generative models of the three tasks. Higher TRT reliability was found compared to Giovannoene and Theodore (2023: LTRS: 0.96 vs. 0.72, Ganong: 0.76 vs. 0.74, PR: 0.70 vs. 0.37). To evaluate model quality, we tested the three models on their ability to recover task parameters from synthetic participant data. Results showed that the generative models more accurately recovered parameters compared to the traditional analysis method of averaging across trials. Generative modeling was successfully extended to lexical reliance tasks and was shown to be advantageous in assessing their reliabilities.

In the attraction effect, a decoy option increases the choice share of a similar, dominating target option at the expense of a dissimilar, competitor option. The attraction effect has been demonstrated across numerous choice domains, including simple perceptual choice, i.e., selecting the rectangle with the largest area. Recently, researchers have also demonstrated a repulsion effect in perceptual choice, where the decoy increases the competitor’s choice share. Crucially, in these experiments, the target and decoy are more similar and thus easier to compare, which may generate correlations in their perceived areas. Such correlations can lead to reduced target choices, even if, on average, the perceived target and competitor areas remain equal. To examine the possibility that the repulsion effect is generated by correlated target-decoy perceptions, as opposed to a choice bias for the competitor, we employed a psychophysics experiment, coupled with Bayesian hierarchical modeling, to estimate the parameters of a multivariate Gaussian choice model. This work shows that judgements of the decoy and target areas are more strongly correlated than target-competitor or decoy-competitor areas. Furthermore, the model naturally produces a repulsion effect that is qualitatively similar to the results of a choice experiment using the same stimuli. These findings suggest that (this form of) the repulsion effect may stem from fundamentally different processes than the attraction effect.

Episodic memory and schema knowledge are known to interact when we recall everyday events – such as the location of an object in a scene. Recent Bayesian models of memory have assumed that this interaction is a function of the trade-off between the strength of episodic memory and schema knowledge. Ramey et at. (2022) empirically quantified this relationship using the recollection-familiarity paradigm as a proxy for memory strength. They also manipulated the congruency/incongruence of object locations in natural scenes as a proxy for schema strength. Here we replicate their findings and then model the effects using a Hierarchical Bayesian model of memory. We model familiarity/recollection as memory strength, and the congruence versus incongruence as having different priors - with the congruent prior linked to accuracy for congruent new scenes in the experimental data, and the incongruent prior linked to accuracy for incongruent new scenes. The model successfully captures 1) the greater accuracy for schema-congruent versus incongruent object locations 2) the decreasing difference in accuracy between congruent and incongruent scenes across familiarity ratings, and 3) the elimination of the accuracy difference for recollected scenes. We also evaluated the dual process claim that recollection responses reflect a distinct recollection process. We found that the best fitting parameters for memory precision had a curvelinear relationship where memory precision gradually improves over the familiarity rating and then substantially changed for recollected responses.

We used novel musical material to investigate how perceptual information is remembered, enables predictions, and how acquisition of such representations relates to motivational factors like preference. Memory may be critical to forming expectations and the aesthetic experience of music, which may in turn reflect a general tendency to experience reward for making correct predictions (Huron, 2006). In two experiments, we addressed three questions concerning relationships between memory, prediction, and preference: How do we learn implicit contingencies within continuous streams of information? Do contingencies shape preference differently for information with different predictive value? How does memory develop when predictive information is differentially informative about what one should expect to hear next? In the first experiment, participants listened to a continuous stream containing single melodies and contingent melody pairs (e.g., melody A was always followed by melody B). After exposure, participants preferred contingent melodies over novel melodies and correct memory for pairs was correlated with preference for melodies that confirmed (melody B) rather than enabled (melody A) predictions, consistent with the hypothesized relationship between memory and preference for prediction confirmations. In a second experiment, currently underway, the exposure stream contained pairs with varying levels of contingency ranging between 100% (B always follows A) and 50% (B follows A half the time). This experiment sheds light on the time-course contingencies are learned; we predict faster learning of stronger contingencies. Experiment 2 also tests if an inverted-U relationship exists between contingency and preference, where moderate levels of contingency are preferred over strong or weak contingencies (Berlyne, 1971).

In the fields of cognitive psychology and cognitive science, it has been pointed out that reasoning with conditionals (if A, then B) can be approximated by a probabilistic approach rather than through material implication (i.e., “not A, or B”) (e.g., Oaksford & Chater, 2001). Moreover, it has been suggested that in everyday communication, inferences are expressed not through formal logic but through social reasoning (Oaksford & Chater, 2020). Accordingly, this study constructed an axiom system for probabilistic logic, which posits that people evaluate logical inferences not as binary true/false judgments but as probabilistic beliefs. In this axiom system, in order to represent conditional sentences as conditional probabilities, propositions representing events are distinguished from propositions representing relationships between events—because conditional probability expresses the relational aspects of event probabilities rather than the probability of an event itself. Next, to capture social reasoning, rather than relying on the semantic or syntactic approaches of traditional formal logic, a pragmatic axiom was constructed by applying Luce’s (1959) choice axiom to the selection of propositions. Furthermore, from these axioms concerning proposition selection, it was demonstrated as a theorem that: (1) beliefs regarding propositions can be represented as probability measures, and (2) conditionals can be represented as conditional probabilities.

In a classic traveling salesman problem (TSP), the goal is to find the shortest tour through all cities, or nodes, in a given graph. We incorporate a two-color paradigm into classic TSP: each node is assigned one of two colors, and transitioning between nodes of different colors incurs a cost equal to double the Euclidean distance between the nodes. This modification makes the color TSP non-metric, and challenges humans’ reliance on spatial proximity cues, thus providing new insights into human visual perception and problem-solving processes. We conducted experiments in which several human subjects were tasked with constructing tours for various color TSPs. Their performance was then evaluated against both the optimal tours and the tours generated by a multiresolution graph pyramid model. Originally developed for TSP with obstacles, which is metric but not Euclidean, the current pyramid model offers a familiar hierarchical clustering approach to approximating human-like solutions. Our comparative analysis at this point finds that some heuristic strategies adopted by human participants are effectively captured by the pyramid model. Other behaviors however, particularly those relating to clustering changes caused by the increased cost of color transitions, are not as well represented – resulting in the pyramid model underperforming compared to some human results. These findings raise questions about whether the pyramid model can be further modified to emulate certain aspects of human visual perception such as path planning, and how human decision-making actively adapts to the non-metric nature of color TSP.

We present a novel study aimed to adjudicate between model predictions of why memory improves with repetition. When repeated items are studied with same versus different source context such as background color, recall probability is greater for items repeated in mass if they are studied in different contexts. However, spaced items do not always benefit from being studied with different contexts (e.g., Verkoeijen et al., 2004). Here we reevaluate these results from the perspective of event segmentation theory (Zacks et al., 2007). For list-learning experiments, a sequence of items studied with the same background forms an event, and an item studied in a new background forms an event boundary. Such event boundary items boast greater memorability (Heusser et al., 2018). Whereas the second occurrence of a massed item with different backgrounds is necessarily an event boundary, a spaced item repeated with different backgrounds may have neither item at an event boundary. In contrast to prior studies which randomized source context, here we provide event structure to repeated items. Participants performed free recall on lists of items with each item having one of two backgrounds. Matched to massed repetitions with different backgrounds, lists included massed repetitions with the same backgrounds but the first item was an event boundary item. Further, some massed items with the same background occurred in the middle of an event. Whether event boundary information contributes to memorability of repeated items has implications for models of episodic memory as well as models of event segmentation.

Understanding how people covary in performance across experimental tasks is a critical component of psychometrics and individual-difference psychology. A key goal, therefore, is the accurate and precise measurement of correlation coefficients in real-world settings. The difficulty is that real-world settings in experiments contain multiple sources of variation such as those from trials, conditions, individuals, and tasks; moreover, not appropriately modeling these sources leads to asymptotically attenuated estimates with nonsense confidence intervals. In these contexts, Bayesian hierarchical models are essential. The problem addressed here is how the choice of prior affects posterior correlation distributions for correlations in real-world settings. We compare through simulation the performance on Inverse Wishart and LKJ priors across a range of settings, and find both priors do well with reasonable settings. The advantage of the Inverse Wishart is computational speed (especially in large designs); the advantage of the LKJ is greater robustness to variation in prior settings. Our recommendation to use LKJ rather than the Inverse Wishart as a default unless speed is prioritized and some scaling information about the data is known a priori.

The increasing misuse of Deepfake technology has made it essential for humans to accurately distinguish between real and fake images. However, the high realism of Deepfake images, especially when image quality varies, poses a significant challenge for human recognition. While machine learning models have been developed for Deepfake detection, humans often struggle to make immediate and accurate judgments. This study investigates how humans learn to detect fake images and how they collaborate with AI in Deepfake detection tasks. Specifically, we examine the processes underlying human-AI collaboration by manipulating AI model accuracy to determine whether humans always benefit from AI assistance,. First, an implicit learning experiment was designed to assess whether individuals could improve their ability to recognize Deepfake images through trial-by-trial feedback. Additionally, we implemented a human-AI collaboration experiment by developing two anti-Deepfake AI models with different accuracy levels. These models provided participants with probability-based assessments of an image’s authenticity to assist in decision-making. Using Systems Factorial Technology (SFT), we applied the single-target self-terminating (STST) stopping rule to quantify the efficiency of human-AI collaboration by comparing decision-making performance with and without AI assistance. The results revealed that participants significantly improved in both accuracy and response time as they gained experience through implicit learning. However, AI assistance had mixed effects on decision efficiency: while a low-accuracy AI model hindered performance by inducing limited-capacity processing, a high-accuracy AI model had no noticeable impact, suggesting an unlimited capacity process. Interestingly, only when images present without background, decisions made with AI assistance outperformed those made through implicit learning alone, highlighting the potential benefits of AI in Deepfake detection. Taken together, this study explores the feasibility of human-AI collaboration in Deepfake recognition and highlights the limitations of undertrained AI models in aiding human decision-making. Our findings contribute to the development and application of anti-Deepfake AI models and provide insights into the potential for future human-AI collaboration.

Aging is notably associated with a slow-down of sensory processing, but there can be an increased benefit from multisensory integration. This benefit is not always observed, however, and can depend on the stimulus and task type. The purpose of this study is to further evaluate multisensory integration in older vs. younger adults using an exhaustive (AND) processing paradigm and the tools of systems factorial technology. Reaction times were measured to 440-Hz auditory targets and red-circle visual targets presented at three levels (loud/soft/absent and bright/dim/absent). In an AND condition, targets were combined factorially with three levels, and participants responded with one button when visual and auditory targets were both present and a different response otherwise. In each of two single target conditions, either auditory or visual stimuli were presented, and participants responded with one button when a target was present and another button otherwise. The workload capacity coefficient, a measure of the efficiency of responding to two stimuli versus one stimulus compared to a standard parallel model, demonstrated wide heterogeneity in both groups in preliminary data. A subset of participants revealed a multisensory advantage over the standard parallel model in the form of super capacity. Results will be discussed in the context of applications of systems factorial technology to exhaustive processing in the visual system and models of auditory visual processing in young and older adults.

Psychophysical research on color filters tends to focus on information lost directly due to a filter. However, many interactions, particularly those relying on color, depend on relationships among and configurations of information sources. In the current research, we focus on narrow-band color filters and how those filters may disrupt the perception of target shape cues. By adapting multiple facets of stimuli, we examine how these disruptions influence attention and information accumulation through the application of a Linear Ballistic Accumulator Omissions Model. We will report on the impact of disruption on visual stimuli with filtered overlays within a search detection task and their influence on varying perceptual cues that may affect the rate of information processed by an individual. We find performance with unfiltered search arrays to be more accurate and faster. LBAO model fits indicate that both information accumulation rate and response caution increase when color filters are applied.

Sense of Agency (SoA) is a core concept related to our experience as intentional agents in our environment. Explicit and implicit measures have been used to study SoA. Recent findings suggest that the most common implicit measure, namely Temporal Binding (TB), may reflect memory processes rather than SoA. Here, we implemented two TB measures and an explicit measure in a novel goal-directed extended action task to better understand SoA measures. Participants either observed or produced dot movements to a target of choice. They then estimated one of two possible durations; for Temporal Binding version 1 (TB1), they estimated the duration between the end of the dot movement and a tone that played either 300, 500, or 700 ms later (akin to traditional TB studies). For TB2, they estimated the duration between the start and end of the dot movements. After every 10 trials, they also provided explicit SoA ratings. The results indicated that participants reported stronger explicit SoA during active than passive movements. Results from neither TB version aligned with prediction based on TB accounts as a reflection of SoA. We discuss memory-based and scaling accounts as alternative interpretations for our data.

We explored strategies used to divide time and effort in a cooperative puzzle completion task involving dyads. In the task, partners must allocate puzzle pieces (28 rotated tetromino shapes) to one of two puzzles within 60 seconds. One puzzle rewards the individual $0.10 for each piece correctly allocated, and the other puzzle rewards the individual and their partner $0.05 each. Pieces unplaced in the latter puzzle also penalize one’s partner $0.03. Many pieces are similar or rotated duplicates, and participants must chose which puzzle to allocate pieces. Puzzle pieces return to the staging area without penalty if a piece is misallocated. Participants were able to review their partner’s performance after each trial. We observed participants (n = 22, 11 dyads) switching back and forth between the two puzzles during each trial. We used a Markov process to model transition and successful allocation probabilities. We modeled the participant success rates as a simple exponential learning function over trials. We modeled the transition probabilities between puzzles as a function of partner performance on the group puzzle. We used STAN to sample from the posterior of the implied model conditioned on the data. We found evidence that individuals are less likely to switch puzzles after misallocating a piece compared to a successful allocation. Although aggregated scores suggest that increased partner contributions to the group puzzle does predict individual contributions in the next trial, the current model of within-trial behavior does not reveal this relationship.

How participants use endogenous cues to guide episodic retrieval remains an open question. Emerging answers to this question vary with recall instructions. In free recall, in which participants recall a list of presented items in any order, retrieved context models provide a leading explanation (Lohnas & Healey, 2021). In contrast, in serial recall, in which participants recall the items in their presented order, positional coding models are the prevalent account (Hurlstone et al., 2014). Despite theoretical differences, both recall tasks share similarities in serial position and transition effects (Bhatarah et al., 2008). Further, recently developed retrieved context models can account for some effects previously interpreted as evidence in favor of positional codes (Osth & Hurlstone, 2023; Lohnas, 2024).  Here we present a novel experiment to further bridge computational models and empirical effects across recall tasks. With serial recall, participants studied lists with phonological similarity manipulations which have classically been taken as evidence against retrieved context models (Henson et al., 1996). Importantly, participants also studied lists of the same design but then performed free recall, the paradigm in which retrieved context models should play a stronger role.  Analyses of recall by serial position and recall transitions reveal established and novel results across paradigms. These analyses inform future model developments.

Modern Bayesian inference involves a mixture of computational techniques for estimating, validating, and drawing conclusions from probabilistic models as part of principled workflows for computational modeling. This poster introduces the new and completely revamped BayesFlow software for amortized Bayesian inference (ABI). Along with direct posterior and likelihood estimation, the framework now includes multi-backend support for PyTorch, TensorFlow, and JAX, the fastest generative networks for sampling available, complete customization and high-level interfaces, as well as new capabilities for hyperparameter optimization, design optimization, and hierarchical modeling. A selection of cognitive benchmarks is presented.

In previous work, we developed a probabilistic EZ drift diffusion model (EZ-DDM) that allows for Bayesian hierarchical extensions with latent variables and metaregression structures. The EZ-DDM provides closed-form estimators for the drift rate, boundary separation, and nondecision time parameters using three summary statistics: the accuracy rate and the mean and variance of the response times. The EZ hierarchical Bayesian drift diffusion model (EZ-HBDDM) is a computationally efficient proxy model for the hierarchical three-parameter DDM built from the sampling distributions of the EZ summary statistics. In this work, we evaluate the efficacy of the EZ-HBDDM for hypothesis testing. We report on numerical experiments in which we generated data from a hierarchical DDM, with person-specific boundary and nondecision time parameters and a within-subject design for drift rate across two or more conditions and with various effect sizes. Hypotheses were evaluated using Bayes factors capturing the ratio between the prior to posterior mass near zero. We additionally evaluated the efficiency of the procedure for various data sizes, from very small to very large. We also evaluate the performance of a robust implementation of the EZ-DDM.

The attraction effect occurs when the preference between two options is altered by the inclusion of a third, decoy option that is similar to, but worse than one of the original options. It challenges rational choice theories by showing that preferences are not stable. The attraction effect is typically studied in situations where individuals have full access to option information during choice. However, in many real-life situations, people must make choices based on their memory of the options. We investigated the attraction effect when people had to rely on their memory when making choices. Bayesian modeling with Stan was used to estimate the attraction effect in different conditions. In two experiments, we show that relying on memory eliminated the attraction effect. In addition, more accurate memory did not lead to the recovery of the attraction effect. We hypothesize that attribute comparisons become more cognitively demanding in memory-based choice because visual information about options is missing.

We introduce an ex-ante model of discretization where the decision maker partitions the state space into discrete categories given their prior subjective belief, the stakes, and the cost of accruing information. Our theory endogenizes how finely decision makers partition the state space, including the frequency of each category and associated decision rules. In a motivating example, we characterize the optimal partition under a uniform prior with cost of accruing information as Shannon entropy. This characterization allows us to test several key hypotheses of the theory with a set of experiments. In the first study, subjects explicitly choose the number of categories given varying levels of stakes and costs. In the second study, subjects reveal how many categories they implement through a perceptual matching task. Across both studies, we find evidence that we can classify subjects as discretizers, questioning whether the classical approach is a good approximation of behavior.

Much research in data science has been devoted to data visualization techniques but not to their effectiveness from a perceptual, and more generally, cognitive standpoint. In this study, we discuss and test a mathematical procedure for replacing human judges in the assessment of stacked bar data graphs. Specifically, we used a formal representational method introduced by Vigo (2019) that makes it possible to apply a similarity measure to stacked bar graphs. To test the effectiveness of the measures used in the procedure, we ran an experiment where participants were asked to judge the similarity between pairs of stacked bar graphs to determine how consistently and accurately these graphs are interpreted perceptually. Preliminary results indicate that the proposed mathematical procedure reasonably predicts people’s similarity judgments with respect to pairs of stacked bar graphs and that these judgments are consistent with the quantitative relations conveyed by such graphs regardless of the ergonomically chosen scale. This procedure may be used to replace judges in situations where researchers wish to assess the congruency of such graphs. A computer program implementing the procedure is proposed as a future direction.

This study explores how a mathematical model of category learning difficulty can be useful in providing insight into the nature of the performance observed when the process of concept formation is constrained by exposure time windows of “conceptual apprehension”. More specifically, we applied the law of invariance from Generalized Invariance Structure Theory (GIST; Vigo, 2013, 2015, 2022, 2024) to results from an experiment in which participants were tasked with learning six types of categories defined over three binary dimensions whose content was shown all at once (i.e., parainformatively) within seven prescribed exposure time conditions ranging from 200 milliseconds to 20 seconds. As expected, results from the experiment indicated that as the exposure time increased, the learning difficulty of the six category structures gradually culminated in the canonical learning difficulty ordering proposed by Shepard et al. (1961). Furthermore, the specific learning difficulty ordering associated with each temporal window suggests a gradient of structural discrimination in concept formation as a result of the time that may be needed by observers to fully detect the specific invariance structure underlying each of the six tested category types. This hypothesis is corroborated by estimates of the structure discrimination parameter from the parametric variant of the law of invariance in GIST.

Dyslexia is a neurodevelopmental disorder that affects reading, spelling, and phonological processing, with growing evidence suggesting deficits in both visual and auditory domains. Traditional diagnostic methods primarily focus on phonological deficits, which may overlook the role of multimodal integration in dyslexia. This study examines whether integrating visual and auditory stimuli improves task performance in dyslexic individuals, providing insights into cognitive processing mechanisms. Participants completed a pseudoword identification task under three conditions: visual-only, auditory-only, and combined visual-auditory. In non-dyslexic participants, response times were faster in the auditory condition, and accuracy was highest in the combined condition. Dyslexic participants demonstrated slower response times and lower accuracy in unimodal conditions, but their performance improved in the combined condition. An ANOVA revealed a significant effect of stimulus type on response times (F (2, 2499) = 303.41, p < .001), and a logistic mixed-effects model confirmed stimulus type significantly influenced accuracy. These findings support the role of multimodal integration in mitigating dyslexia-related deficits and highlight the need for computational and mathematical models that account for multisensory processing in dyslexia. This work contributes to the broader understanding of cognitive mechanisms underlying reading disorders and informs potential interventions.

A long-standing debate in recognition memory research concerns whether item or context noise plays a direct role in retrieval (e.g., Dennis & Humphreys, 2001; Shiffrin & Steyvers, 1997). While both likely contribute to interference (e.g., Criss & Shiffrin, 2004), the underlying cause of the test position effect— also known as output interference, a gradual decline in memory accuracy over the course of testing—remains unresolved (Criss et al., 2011; Tulving & Arbuckle, 1972). An item-noise model, Retrieving Effectively from Memory (REM), suggests that early recognition impairs later accuracy due to the encoding of new traces and degradation of untested ones (Criss et al., 2011; Shiffrin & Steyvers, 1997). In contrast, Osth and Dennis’ (2015) model, which incorporates three sources of interference—item noise, context noise, and background noise—attributes the test position effect to increasing context noise from shifts in context representation rather than item noise alone. In a reanalysis of the Murdock and Anderson (1975) data, Osth and Dennis (2015) argued that a model with only item-noise cannot account for the test position effects across multiple alternative forced-choice tasks. We show that REM can successfully account for these results. Moreover, we report novel contrasting tests of these two models in which we manipulated task type (two- and four-alternative forced-choice tasks), test composition (one vs. three old items in a four-alternative forced-choice task), and target base rate at test.

Context effects--the attraction, similarity, and compromise effects--show that the relative attractiveness of a target option over a competitor option changes when a third option is introduced. Specifically, an inferior yet non-dominated distractor increases the target’s attractiveness (attraction effect); a distractor similar to the target but dissimilar to the competitor decreases the target’s attractiveness (similarity effect); and positioning the target between a distractor and a competitor increases its attractiveness (compromise effect). Because most prior studies relied on narrow, hand-picked stimuli, it is unclear how robust these effects are across the full attribute space and whether they are domain general or domain specific. I addressed these gaps by systematically sampled options from the attribute space to empirically construct context effect maps in three domains: belief, preference, and perception. These maps allowed for a detailed quantification of context effects and tested their domain specificity. At the aggregate level, a similarity effect emerged only in the perception domain. Nonetheless, all three context effects appeared conditionally, moderated by the target–competitor relationship in both the perception and belief domains. Specifically, attraction and compromise effects were stronger when the target was larger/better than the competitor, whereas the similarity effect was stronger when the target was relatively smaller/worse. The qualitative pattern of moderation was shared across domains, though effect magnitudes varied. Beyond the classic effects, a general context effect was observed across all three domains: the relative attractiveness between two options shifted with the presence of a third, with domain-dependent strength and form. Taken together, these findings suggest a broader, more nuanced context effect that is nevertheless domain-specific.

Large Language Models (LLMs) and Vision-Language Models (VLMs) have the potential to emulate human cognitive abilities for robots and thus significantly advance the evolution and use of cognitive architectures. This opens up opportunities for using human-like judgment and decision-making capabilities such as instance-based learning and intuitive decision making inherent in cognitive architectures with social robots. Using a combined system of an Adaptive Control of Thought-Rational (ACT-R) model and a humanoid social robot, we show how content from the declarative memory of the ACT-R model can be retrieved per association of real-world data obtained by the robot via the image recognition capabilities of an LLM. Such recollections can be fetched and processed according to the procedural memory of cognitive model productions and then returned to the robot as instructions, for example to add the prompt to LLM-driven utterances to keep them more contextual. In addition, visual impressions captured by the robot can be stored in the cognitive model.

In cognitive neuroscience, there has been growing interest in adopting sequential sampling models (SSM) as the choice function for reinforcement learning (RLSSM), opening up new avenues for exploring generative processes that can jointly account for decision dynamics within and across trials. To date, such approaches have been limited by computational tractability, due to lack of closed-form likelihoods for the decision process and expensive trial-by-trial evaluation of complex reinforcement learning (RL) processes. By combining differentiable RL likelihoods with Likelihood Approximation Networks (LANs), and leveraging gradient-based inference methods including Hamiltonian Monte Carlo or Variational Inference (VI), we enable fast and efficient hierarchical Bayesian estimation for a broad class of RLSSM models. By exploiting the differentiability of RL likelihoods, this method improves scalability and enables faster convergence with gradient-based optimizers or MCMC samplers for complex RL processes. To showcase the combination of these approaches, we consider the Reinforcement Learning - Working Memory (RLWM) task and model with multiple interacting generative learning processes. This RLWM model is then combined with decision-process modules via LANs. We show that this approach can be combined with hierarchical variational inference to accurately recover the posterior parameter distributions in arbitrarily complex RLSSM paradigms. In comparison, fitting a choice-only model yields a biased estimator of the true generative process. Our method allows us to uncover a hitherto undescribed cognitive process within the RLWM task, whereby participants proactively adjust the boundary threshold of the choice process as a function of working memory load.

When interpreting others' actions, humans often rely on beliefs about personality traits or “personality types” to predict mental states and behavior (mentalization). However, individual differences in these intuitive inferences are often overlooked and unmodeled. To address this, we developed a controlled paradigm using Minecraft to investigate how participants' beliefs about player types and personality dimensions influence their judgments of others’ player types. A multinomial regression revealed that the interaction between participants’ ratings of targets on two personality dimensions significantly predicted player type classifications. This was supported by using frequentist and Bayesian hierarchical models taking into account participants as random effects. We tested three competing Finite Mixture Models, each incorporating participants’ elicited beliefs as probability distributions. The models were evaluated using standard metrics, including Leave-One-Out Cross-Validation, BIC, Cross-Entropy, Adjusted Rand Index, RMSE, and correlation, based on their fit and predictive accuracy on unseen data. This novel approach provides a structured way to assess the extent to which participants' reported beliefs explain variability in mentalization performance.

Human languages exhibit ample differences yet this variation appears to be constrained by robust organizing principles. Research in semantic domains such as color or kinship suggested that the world’s languages tend to exhibit a near-optimal trade-off between simplicity and informativeness, suggesting that language evolution is influence by such principles. The origin of these pressures, however, is not fully understood, and may stem either from the human cognitive system or from external factors in the communicative environment, coupled with evolutionary dynamics. To characterize this process, we present a computational model that simulates the emergence of a communication protocol (or ‘language’) between two neural network agents. We focus on the domain of kinship, hence agents are trained to develop a communication system to refer to members of their kin. Our agents do not have any prior linguistic knowledge; therefore, in order to communicate, they send signals by choosing symbols (‘words’) which are initially meaningless. Over the course of training, we reward interactions that lead to successful communication, hence over time agents develop a system of signal-meaning associations that allow them to refer to different members of their kin. Notably, we do not guide the agents to specific signal choices, thus their behavior arises from the domain-general statistical learning mechanism in our neural network agents. We observe that, provided agents are allowed to use a large enough vocabulary, the emergent languages may be initially complex but become simpler over time, reaching the optimal trade-off after sufficient training.

This paper investigates the cognitive processes for abacus gestures (a set of mid-air gestures representing numbers 0-99). We developed multiple cognitive models using ACT-R architecture with vision and motor modules and PyIBL with reward-based learning to simulate human learning and performance. By enhancing previous research models with VisiTor (Vision + Motor) framework, we added visual attention and motor movements involved in abacus gestures learning. Our results demonstrate that ACT-R models can show the learning curve with the contribution of the retrieval processes. The PyIBL model successfully simulated the learning curve with an increasing reward trend. The study discovers the retrieval processes, rather than vision or motor modules, dominate the learning curve in abacus gestures learning processes. The findings through cognitive modeling provide insights into gesture-based interaction design and human cognition during mid-air gesture learning.

Lindstedt and Byrne (2018) proposed a simple, minimally disruptive, generally applicable framework within ACT-R to assign visual object groupings based on proximity, while accounting for temporal relations across new scenes. However, this approach forms groups entirely based on proximity; i.e., objects that are close together tend to group together. This is a useful starting point but many other visual properties can cause objects to group. In this work, we further expand on the original system by incorporating alignment as a fundamental grouping principle, enhancing its ability to model human visual cognition. Our revised model introduces new mechanisms that enable the system to detect and utilize alignment, improving its flexibility and predictive power. We evaluate the impact of this addition through comparisons with empirical findings, demonstrating how alignment influences visual grouping. Furthermore, the updated design prioritizes extensibility, allowing for future integration of additional Gestalt principles such as shape and enclosure. By refining the underlying mechanisms and broadening the system’s scope, we contribute to a more comprehensive framework for modeling visual grouping. This work advances theoretical understanding of perceptual grouping in cognitive architectures and has implications for applications in human-computer interaction and cognitive modeling.

Active Inference is an ambitious research program in cognitive science and mathematical psychology aiming to provide a unified account of perception, goal-oriented action, and cognition. In this analytical and critical study, we argue for the inadequacy and incompleteness of active inference's theory of perception. Specifically, we show that `perception as inference,' considered as an explanatory psychological/philosophical theory of perception, faces the `problem of prior knowledge' - the problem of accounting for the origin and nature of knowledge assumed in the Bayesian agent's generative model. We trace the conceptual inadequacies to the psychology/philosophy of Helmholtz and adumbrate an alternate theory of direct perception based on the psychology/philosophy of Whitehead.

Bayesian inference is a powerful mathematical framework for modeling cognitive processes, and it has been widely used in computational models of sensorimotor integration. However, higher-level cognitive functions may also rely on processes that resemble Bayesian integration. In this study, we explore this framework in the context of belief updating about one's own sense of control. We conducted an experiment in which a spaceship has to be navigated through sections characterized by loss of control. Participants have to use their knowledge about the environment to project the path of the spaceship several steps ahead. Before each trial, participants were asked how well they expected to perform, and afterward, how much control they felt to have. Through model selection, we identified that measures for previous and current performance affect given responses. Additionally, we found that scores on Neuroticism were associated with expectancies but none of the big five personality traits had an effect on control responses. We discuss how the Bayesian framework can be used to integrate expectancy and actual performance in order to infer the final control response and how the control response of the previous trial translates into the current expectancy. Our goal is to integrate this framework into a cognitive model using the ACT-R architecture. This is ongoing work, and we share initial findings.

Computational cognitive architectures are useful tools for capturing the structures and processes of the mind computationally as well as for simulating behavior. One such cognitive architecture, Clarion, incorporates a two-level structure consisting of both symbolic and sub-symbolic representations (at the two different levels respectively) and bottom-up learning that goes from sub-symbolic to symbolic representations (using the Rule-Extraction-Refinement algorithm). This work explores the integration of Large Language Models (LLMs) into the Clarion framework to enhance its capabilities. The present paper specifically explores a new rule extraction method within Clarion, with SBERT (Sentence-BERT) incorporated into the bottom level of Clarion and a sliding window approach to extract n-gram rules for the top level of Clarion. This modified version of the Rule-Extraction-Refinement algorithm is used to carry out bottom-up learning within the new Clarion. Ongoing experiments on the Pennebaker and King essays dataset (for personality prediction) demonstrate the potential for improved performance and increased explainability when incorporating LLMs into Clarion.

Memory tasks such as the Deese-Roediger-McDermott (DRM) paradigm demonstrate how semantic associations can induce false recognition(Roediger & McDermott, 1995). Although cognitive architectures such as ACT-R have modeled various memory processes (Stewart & West, 2007; Kelly, Arora, West, & Reitter, 2020), they often struggle with large-scale semantic representations and thus lack the capacity to reliably predict false recognition. In contrast, hyperdimensional computing approaches effectively capture semantic similarity but lack mechanisms for control processes (Reid & Jamieson, 2023; Dodhia & Metcalfe, 1999). We propose a computational model that integrates distributional semantic models with ACT-R through holographic declarative memory (HDM, Kelly et al., 2020) to simulate the DRM paradigm. Our model uses embeddings generated by the BEAGLE algorithm (Jones & Mewhort, 2007) to represent pre-experimental semantic knowledge, and incorporates control processes, including rehearsal and decision strategies, to capture recall and recognition effects (Atkinson & Shiffrin, 1968; Lehman & Malmberg, 2013). We demonstrate that the model reproduces primacy and recency effects, and we expect high false recognition rates for semantically related lures and intra-list, interlist, and extra-list phenomena during recall. This work bridges cognitive architectures like ACT-R, which model human decision-making and skilled behaviour, and distributional semantics, enhancing computational models of false memory.

Neural specificity is the ability to differentiate multiple representations in a neural population such that individual neurons provide some constituent feature of those representations (Park, Carp, Hebrank, Park, & Polk, 2010; Kleemeyer et al., 2017). Two different theories of neural specificity have been developed to explain interregional brain differences that trend with performance loss: neural dedifferentiation (Koen, Hauck, & Rugg, 2019) and system segregation (Chan, Park, Savalia, Petersen, & Wig, 2014; Wig, 2017). Although both estimates demonstrate a significant trend with performance, neither adequately explains the mechanisms by which neural specificity causes performance differences. To identify these mechanisms, we have developed a neurocognitive model in the NengoSPA framework that performs a go/no-go task. To simulate neural dedifferentiation and system segregation, two interventions were performed on the model. This resulted in three simulations: a control simulation, and two intervention simulations. The first intervention was neural population reduction (NPR) in all network regions, while the second intervention was representation localization reduction (RLR) between neural populations in different network regions. Performance in both intervention simulations decreased compared to the control simulation, although in different ways. The NPR simulation demonstrated consistent omission errors, while the RLR simulation demonstrated consistent commission errors. These findings suggest that a reduction in neural population causes the omission errors typical of neural attenuation, while a reduction in neural specificity causes the commission errors typical of neural broadening, both of which are implicated in neural dedifferentiation.

The Attentional Training Technique (ATT) has been shown to enhance attentional control and reduce maladaptive cognitive patterns but lacks a well-defined computational explanation. This paper applies a metacognitive skill model within the ACT-R cognitive architecture to clarify the procedural mechanisms underlying ATT. Grounded in proceduralization theory, we propose that ATT transforms declarative attentional strategies into automatic procedural skills, enhancing metacognitive control and emotional regulation. This framework advances our understanding of the computational and cognitive mechanisms supporting ATT, its applications in psychotherapy, and the process of metacognitive skill learning.

In dynamic multi-object task environments, human memory retrieval is shaped by both cognitive processes and environmental influences. Search time, as the interval between the moment participants first checked the subtask and when they began interacting with the target object of the current subtask, serves as a key indicator of retrieval efficiency. Analysis of data from 10 participants using an AI2-THOR simulated household setting revealed substantial variability in search time, reflecting differences in memory processes and action implementation. Memory dynamics (i.e., all changes of memories due to changes in environmental) were modeled with both a conventional base-level learning equation and a modified version that integrates entropy-based uncertainty and observation duration. While the conventional model did not capture observed fluctuations in search time, the modified equation demonstrated a closer alignment with behavioral data, as evidenced by corresponding changes in activation values and fluctuation rates. These findings underscore the importance of incorporating uncertainty and temporal factors into memory models to more accurately simulate human memory retrieval in dynamic multi-object environments.

We created Scaper, a static code analysis tool for production system architectures. Using this tool, we successfully analyzed 26 ACT-R models. This process took ~5 seconds to complete and identified 3,619 components including chunk-types, declarative memories, and productions. In general, the analysis shows appropriate levels of compliance to the Scaper’s tests. Models’ compliance scores are above 94% for component usage, and above 78% for production design. However, further analysis revealed that a 3.9% of 152 chunks are not used, and 4.1% of the references from productions have discrepancies with the chunk-type definitions. Although we did not analyze whether such issues affect models’ performance, their detections at least suggest code improvements opportunities. Moreover, Scaper can support the development of production systems by (a) providing a better visibility of the components, (b) identifying discrepancies in the components’ relationships, and (c) fostering models’ code readability, maintainability, and scalability.

What communication methods have been used between cognitive architectures and environments and which methods might be most suitable for disaster recovery scenarios? We discuss some existing approaches for communication between Cognitive architectures such as ACT-R and modern simulation environments. As a part of this analysis and discussion, we also evaluate communication methods’ potential efficacy for disaster recovery simulations and then related real-world applications. By systematically analyzing both technological capabilities and practical constraints, we provide useful insights to researchers and practitioners in selecting optimal communication systems for some applications, especially disaster recovery applications. This practical analysis provides a useful connection between theoretical cognitive models and their practical implementation in high-stakes environments like disaster recovery and should generally be useful for cognitive modelers seeking to understand the functional landscape of connecting their cognitive model to a digital environment.

Despite the continued anthropomorphizing of AI systems, the potential impact of racialization during Human-AI interaction is understudied. In this poster we explore how human-AI cooperation may be impacted by the belief that data used to train an AI system is racialized, that is, was trained on data from a specific group of people. During this study, participants completed a human-AI cooperation task using the Pig Chase game, a variant of the popular Stag Hunt game. Participants from different self-identified demographics interacted with AI agents whose perceived racial identities were manipulated, allowing us to assess how socio-cultural perspectives influence participants’ decision-making in the game. After the game, participants completed a survey questionnaire to explain their strategies which were used while playing the game, and to understand their perceived intelligence of their AI teammate. Statistical analysis of the task behavior data revealed a statistically significant effect of the demographic of the participant, as well as the interaction between this self-identified demographic and the treatment condition (i.e., the perceived demographic of the agent). We built a cognitive model of the task using ACT-R cognitive architecture using the game theory of mind model, to have a cognitive-level, process-based explanation of the participants’ perspectives based on results found from the study. This model helps us better understand the factors affecting the decision-making strategies of the game participants.

In this paper, we discuss the beginning stages of building an experimental ACT-R GUI to model the feature engineering process, as well as the places in the feature engineering process where representations of the biocentric Man may impact behavior in important, even if subtle, ways. The biocentric Man can be thought of as a caricature with exaggerated socioculturally constructed features that represent the hegemonic in Western societies (white, cis-gender, man, etc.). In the feature engineering process, we argue there are 10 stages: task assignment, literature review, exploratory data analysis, data cleaning, human-domain expert collaboration, feature extraction, feature generation, feature selection, and feature evaluation. In this paper, we begin to model the literature review process. We use literature from the Scopus database to represent literature entering the pipeline, we then use Python frameworks and holographic declarative memory structures to process the literature, next we will perform principal component analysis in order to determine word associations made during scatter/gather. We also begin to build two agents, one Blackness concerned and one Blackness unconcerned, with the intention of comparing their performance or resulting holographic declarative memory vector space representations. This is important because we could uncover the influence of racial concern or racial awareness on feature engineering performance—which could inform us of the importance or relevance of making race-based concerns and racialized information known to feature engineers. In the future, we’ll create principal component analysis charts to visualize the relationships between certain race-concerned words and words related to the feature engineering task.

Perceptual decision-making provides a framework for understanding how organisms translate sensory evidence into actions, but traditional models face challenges in explaining choice phenomena and motor integration. Despite evidence of both covert and overt motor processes during deliberation, most frameworks treat movement as merely implementing a completed decision. We explore the relationship between action and decision making by extending a proposed framework for embodied choice and independently varying the influence of motor feedback on internal choice variables and the contribution of evidence to action. This new model, Degenerate Embodied Choice (DEC), arbitrates between parallel and embodied theories of choice. We demonstrate that DEC replicates the speed-accuracy trade-off (SAT) degenerately, with embodiment proving both necessary and unique for trading speed and accuracy across urgent and accuracy-emphasised tasks. DEC emulates empirical data both qualitatively and quantitatively, with model-fitted parameters falling exclusively within the embodied set and producing congruent predictive SAT values within a narrow band. We then introduce the Optimality Framework for Embodied Choice (OFEC) as a lens for examining embodied choice through optimality principles. Our findings suggest that complex decision behaviours can emerge from simple underlying principles, whether through geometric properties of decision boundaries or motor-cognitive integration.

Dynamic Causal Modeling (DCM) in task-based fMRI traditionally compares different network configurations based on condition-dependent changes in brain signals, such as from fMRI. Regions are typically selected based on showing experimental effects according to a First- or Second-Level Model. The temporal dynamics of those changes are used as an indicator for the directionality of information flow within the network, and drive the estimated fit of a specific model. However, in theory-driven models which use a predefined set of regions not driven by experimental effects, some predefined regions of interest (ROIs) may not exhibit a significant signal change in response to experimental conditions which would be indicative of the temporal dynamics. Yet, they might be involved in processing and relevant to the network. This can occur when a region is both involved in the processing within and outside task conditions, making the experimental inputs insufficient to capture its temporal dynamics. To still capture the neural dynamics of different regions, we explore the use of other inputs which might relate to the underlying connectivity, such as Low-Frequency Oscillations (LFOs), and could be used alongside traditional task-related inputs. LFOs have been shown to reflect synchronized activity across brain regions in resting-state fMRI, where experimental inputs are unavailable, and are established as driving inputs in resting-state DCM. Preliminary results indicate that the addition of LFOs in theory-driven Dynamic Causal Models can improve model fit based on a number of metrics. The source of these improvements is still an open question and requires further investigation.

Open source is a moving force in scientific research. The Cognitive Modeling community largely benefits from this mutualistic relationship, but how large is the actual gap between state-of-the-art methods and practical use in research settings? This study exemplifies complementary approaches for using open source to model cognitive mental demand from three conceptual and practical perspectives: subjective reports, peripheral biosignals, and psychochronometric behavior. Participants answered the NASA-TLX to report their perceived mental demand related to the task. Heart rate variability and pupillometry were the selected biosignals to assess the mental demand of the N-Back task. A demonstration of two uses of ML in the process of feature engineering and exploratory analysis showed that self-report alone is enough to predict a participant's trial level with a Cohen's Kappa of 0.9187 (SD = 0.02), and up to .9290 (SD = 0.02) when combined with the response times during the trial. Linear Mixture Models confirmed the effect of the N-Back level on participant's mental demand. Both the HRV and pupillometry mixture models showed a significant effect of the N-Back level on the features extracted from the biosignals, but HRV features demonstrated to be cost-effective and reliable. The selected feature for HRV: the proportion of low-frequency power over the total power of the HRV predicted the N-Back level with a Cohen’s Kappa of 0.8571 (SD = 0.01) in our best models.

Drs. Claudia González Vallejo, Director of Decision, Risk, and Management Sciences, DRMS, Program at the National Science Foundation, will be available to answer questions about the program and other NSF initiatives and priorities. DRMS supports scientific research directed at increasing understanding and effectiveness of decision making by individuals, groups, organizations, and society. DRMS supports research with solid foundations in theories and methods of the social and behavioral sciences, including mathematical models of cognitive processes in decision making; risk assessment and communication; decision and judgment sciences.

This is an in-person presentation on July 26, 2025 (17:30 ~ 19:30 UTC).