Poster Presentations
Mr. Zac Coates
Climate change is a collective problem, yet many mitigation efforts rely on voluntary contributions by individuals to accrue collective benefits. In this study, we conduct a contingent valuation experiment with 1,518 Australian participants to determine how willingness to pay for greenhouse gas abatement changes when the choice environment is framed as either an individual choice or policy preference. We study three domains, electricity consumption, long haul flights, and food. Across all three domains, participants who made decisions on public policy had more than double the mean willingness to pay than those who made decisions on voluntary action. This effect increased with respondent belief in human-caused climate change and government intervention in markets. These findings partly explain why there is often a significant mismatch between stated willingness to pay in surveys and revealed willingness to pay in carbon abatement markets. They also demonstrate that support for decarbonisation projects significantly increases when the costs are shared collectively. Finally, the study finds evidence that the relative difference in WTP for methane abatement and CO2 abatement is less than the 28x difference in greenhouse potency.
This is an in-person presentation on November 19, 2026 (13:00 ~ 14:00 PST).
Adaptive cognition requires using experience to regulate subsequent decisions. Outcomes can change learned associations, beliefs about control, and the processes recruited to act. We present two studies that examine how these adjustments unfold and interact across timescales, combining computational modeling, EEG, and large behavioral samples. The first study examines how rapid, capacity-limited working memory (WM) interacts with slower reinforcement learning (RL). Participants with major depressive disorder, bipolar disorder, schizophrenia, or no diagnosis (N=255) performed an associative learning task during EEG recording. Model-based EEG analyses extracted trial-wise neural markers of WM load and RL value. In controls, neural WM markers increased as reward history accumulated, consistent with RL supporting the management of WM resources. This adjustment was absent in depression, whereas bipolar disorder and schizophrenia showed reduced WM recruitment and weaker WM-RL cooperation. Similar behavioral impairments therefore arose from distinct failures in the coordination of fast and slow learning systems. The second study examines how beliefs about control are learned and used over longer timescales. In a novel persistence-escape paradigm (N=457), participants chose on each trial between effortful persistence and disengagement across contexts that varied incentives, effort, and outcome controllability. Threat reduced escape when outcomes remained controllable but increased escape when control was removed. Standard Q-learning could not reproduce these patterns. The Meta-Arbitration of Control and Agency Q-learning (MACA-Q) model embeds value learning within a meta-control architecture. Inferred agency gates whether outcomes update value, and meta-control regulates how learned values and contextual cues guide engagement. Simulations show that agency built in safe contexts carries over to threat and supports selective rather than indiscriminate persistence. Together, the two studies show that distinct failures update distinct internal states, including learned value, WM allocation, and inferred agency, and that each update changes how control is allocated on subsequent decisions. In both studies, similar behavior arose from different latent trajectories that became dissociable only across contexts. This approach connects the dynamics of learning with the dynamics of decision making to explain cognitive adaptation under limited capacity and partial control.
This is an in-person presentation on November 19, 2026 (13:00 ~ 14:00 PST).
Joachim Vandekerckhove
Model discovery techniques can help discover plausible candidate models from data and can aid theory-building in areas of behavioral and social sciences where empirical work has not yet led to formal specification. In this project, we employ genetic algorithms to explore the space of nomothetic mean structures that could describe hierarchical data. The mean structure is nomothetic in the sense that the variables and parameters are assumed to be mathematically arranged in the same way across individuals. Differences among individuals are exhibited through the observed (measured) variables and the unobserved (latent) parameters. The genetic algorithms are modified to accommodate the wide (rather than long) data formats common in the behavioral and social sciences. We discuss the computational set up of the approach and establish a proof of concept by demonstrating the general approach through numerical examples.
This is an in-person presentation on November 19, 2026 (13:00 ~ 14:00 PST).
Dr. Xiaozhi Yang
People tend to choose the options they look at more. Prior work explains this effect by assuming that visual attention increases preference for the attended option. However, it remains unclear why looking at an option increases its preference and when this relationship is weakened. Using an experimental paradigm that combines eye-tracking and a think-aloud protocol, we provide evidence that visual attention shapes preference by activating attributes associated with the attended option. Critically, the content of these activated attributes depends on contextual factors such as stimulus desirability and behavioral goals. In standard choice settings, where the goal is to choose the preferred option, visual attention primarily activates positive attributes. In contrast, when these contextual factors are incongruent, activated attributes become less positive, reducing the influence of visual attention on choice. To account for these patterns, we propose a computational model of gaze-driven attribute activation. Our model subsumes existing theories as special cases, and uniquely predicts which attributes are brought to mind by gaze and how they shape downstream decisions. Together, our work offers a mechanistic explanation of a fundamental driver of choice as well as its boundary conditions.
This is an in-person presentation on November 19, 2026 (13:00 ~ 14:00 PST).
Ms. Yiming Wang
Dr. Robby Ralston
Ms. Agnes Zhang
Hanrui Mei
Dr. Brandon Turner
How do people represent partial information when making decisions? We distinguish between two types of partial information: unknown features, whose values are hidden from the decision maker, and incompatible features, which are known to be absent rather than merely mismatched. We use similarity judgments to investigate how these two types of partial information are represented and to identify constraints on formal models of similarity. We designed an experiment using amoeba-like stimuli with three features located on their arms. To manipulate uncertainty, we obscured features by covering them with dirt. To manipulate incompatibility, we removed features entirely, resulting in two-dimensional amoebas. Using a comprehensive set of pairwise similarity judgments, we find four key results: (i) similarity decreases as the number of mismatching feature pairs increases; (ii) unknown features yield intermediate similarity judgments between known matches and known mismatches; (iii) stimuli with unknown features are not judged to be perfectly similar to themselves, challenging similarity models that represent uncertainty using a single point in feature space and compute similarity solely from distances between points; and (iv) incompatible features produce greater dissimilarity than known feature mismatches, suggesting that incompatibility is represented differently from certain mismatch and uncertainty. Together, these findings provide empirical constraints on formal theories of similarity under partial information. In particular, they motivate representations that capture uncertainty over feature values while distinguishing incompatibility from ordinary feature differences. Beyond static similarity judgments, such representations may also provide a foundation for dynamic models of decision making, in which individuals sequentially sample information to resolve uncertainty.
This is an in-person presentation on November 19, 2026 (13:00 ~ 14:00 PST).
Dr. Brandon Turner
Peter Kvam
Vladimir Sloutsky
Attention during learning is jointly shaped by task relevance through top-down mechanisms and perceptual salience through bottom-up mechanisms. However, their respective influences on different attention-related processes – information sampling and decision weighting – remain difficult to disentangle. Decisions reflect not only the utility weight assigned to information but also how that information is sampled, making performance alone an ambiguous measure of attentional weighting. To address this, we orthogonally manipulated information relevance and salience in a category-learning task, measured overt sampling through eye tracking, and fitted a family of gaze-conditioned generalized context models (GCMs) with parameters for decision weighting to test whether changes in overt sampling alone are sufficient to account for learning performance. We observed three distinct phases of overt sampling across learning. Early in learning, salient information attracted more sampling regardless of its relevance. As learning progressed, people increasingly redirected attention away from salient but irrelevant information within trials, suggesting reactive inhibition. Late in learning, sampling became predominantly relevance-driven, with minimal initial capture by salient but irrelevant information, suggesting proactive attentional control. Models incorporating salience-specific gaze sampling while assuming salience-invariant decision weights captured the learning curves when salience and relevance were aligned, but they failed to fully reproduce the accuracy change when they were misaligned. Allowing decision weights to vary with salience more closely reproduced the observed learning patterns. In conclusion, these findings suggest a potential dissociation between overt information sampling and decision weighting: learning progressively shifts overt sampling from salience-driven capture toward relevance-guided selection, while estimated decision weights remain sensitive to salience over the course of learning.
This is an in-person presentation on November 19, 2026 (13:00 ~ 14:00 PST).
Dr. Eeshan Hasan
Dr. Eungik Lee
Andrew Caplin
Jennifer Trueblood
Well-calibrated probabilistic thinking is essential for effective human-AI collaboration, helping people use AI advice selectively and appropriately. We test whether calibration can be improved through trial-by-trial learning from AI-predicted probabilities in a medical decision making task. Novice participants (N=300) judged the probability that white blood cells were blast (cancer) cells across three phases: training with ground truth feedback, practice with AI exposure, and test without feedback. A single CNN was shown in three ways, holding accuracy fixed at ~94% while varying the certainty it expressed: its original probabilities, which were conservative and poorly calibrated for the model's accuracy; a recalibrated version, well calibrated for the model and correspondingly certain; or categorical predictions conveying no explicit certainty. Exposure to the original conservative AI substantially reduced characteristic over extremity at test, improving calibration relative to other conditions. At the same time, minimal condition-specific improvement in discrimination suggests that calibration improved through superficial mimicry of the original AI’s uncertainty, which, although miscalibrated relative to its own high accuracy, was better suited to the lower accuracy of novices. AI-assisted training of calibration and other decision skills may face an alignment problem: the certainty signal that most appropriately describes a model need not be the signal that best supports a learner.
This is an in-person presentation on November 19, 2026 (13:00 ~ 14:00 PST).
Sean Conway
Models of decision-making are tasked with determining an appropriate structure for the noisiness, or stochasticity, while also measuring other psychological constructs. To do so, many models incorporate a response scaling parameter which transforms relative utility values into a response distribution to account for decision stochasticity. In categorization, the Generalized Context Model (GCM; Nosofsky, 1986) uses a scaling parameter to account for highly deterministic responding, but this parameter often overlaps with a parameter governing perceptual sensitivity. We explore the nature of stochasticity in the GCM through formal analyses. It is also shown how experimental manipulation of response noise generates an interpretable and generalizable noise ratio, a ratio which may aid the interpretability of other parameters. We report the results of an empirical study with a discussion of application to other decision tasks and models.
This is an in-person presentation on November 19, 2026 (13:00 ~ 14:00 PST).
Joachim Vandekerckhove
Adaptive behavior requires organisms to adjust their choices when environmental constraints change. The matching law is a widely established model of choice behavior, describing how organisms allocate their responses across different alternatives as a function of the relative rewards obtained from them. However, while matching describes behavior at equilibrium, there are still open questions regarding how animals reach such equilibrium and how they adjust and redistribute their choices when the relative payoff from each alternative suddenly changes. In this study, we examine the behavior of six pigeons responding to dynamic concurrent VI-VI schedules, in which the scheduled relative rates of reinforcement abruptly change within sessions. After such a change occurred, the birds started adjusting to the new equilibrium relatively quickly, but fully reaching the new matching prediction took them considerably longer. To account for these results, we compare different models of choice that consider different mechanisms, including the influence of previous reinforcers, the passage of time, and the tendency to repeat previous choices. We use a Bayesian approach to compare these models, with particular attention to whether the observed adjustments reflect gradual changes in learned preferences, persistence in responding, or some combination of both. We explore whether these mechanisms produce different predictions regarding the nature of the post-change adjustment and how well they account for the initial rapid adjustment and the slower convergence toward the new matching equilibrium. Our goal is to understand the processes underlying adaptation to changing environments and exploit the usefulness of the Bayesian framework in separating latent learning dynamics from observed choice behavior.
This is an in-person presentation on November 19, 2026 (13:00 ~ 14:00 PST).
Dr. Ami Eidels
Guy Hawkins
Dr. Scott Brown
Life requires us to undertake any number of tasks within a day, many of which cannot be completed at the same time. This requires us to consider the order in which to apply our limited resources to complete these tasks and achieve our goal(s). While the order is often left up to us, it is often constrained or dictated by external factors, such as when following a recipe. Task Scheduling refers to the order in which we assign our limited resources to process tasks, a common issue across individuals, teams and organisations. Despite its commonality, scheduling of multiple tasks is an area overlooked in decision making research. We aim to compare scheduling performance and behavior in individuals and teams, and how this changes under limited time and when externally constrained. Participants were required to order, select and complete a number of perceptual tasks of varying difficulty under limited time, both by themselves and with another agent. This experiment introduces a unique, extensible format to study performance and behavior when scheduling multiple tasks and under different constraints.
This is an in-person presentation on November 19, 2026 (13:00 ~ 14:00 PST).
The capacity coefficient traditionally assumed high enough accuracy to compare performance to the unlimited capacity, independent, parallel baseline based on response times alone. When estimating the capacity coefficient, this approach usually translates to ignoring error responses. However, this corresponds to the assumption that error response times are uninformative about correct response times. One alternative is to consider incorrect response time as censoring events. Treating incorrect response times as right-censoring corresponds the assumption that the correct response time would have been longer, but nothing else is known. This later assumption exactly fits race models such as the Linear Ballistic Accumulator Model and improves on the independence assumption for capturing the Diffusion Decision Model. Right-censoring works well when estimating the first-terminating capacity coefficient, which is based on the cumulative hazard function. The exhaustive capacity coefficient, however, is based on the cumulative reverse hazard function, which is more straightforward to estimate assuming left-censoring. Left-censoring corresponds to the assumption that a correct response time would have been faster if the incorrect response had not occurred, which makes little sense. In this work, I will demonstrate the biasing effects of left-censoring and ignoring incorrect response times relative to right-censoring for first-terminating models. I also propose that an alternative capacity measure for exhaustive models, based on cumulative hazard functions, be used when right-censoring is appropriate. Paralleling the examination of the first-terminating capacity coefficient with censoring, I will demonstrate the bias with left-censoring and ignoring incorrect response times for exhaustive processes.
This is an in-person presentation on November 19, 2026 (13:00 ~ 14:00 PST).
Submitting author
Author