Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers
Metacontrol & flexibility in inference and decision-making
Contributed Talk Session: Thursday, August 14, 10:00 – 11:00 am, Room C1.04
Confidence in absence as confidence in counterfactual visibility
Talk 1, 10:00 am – Maya Schipper1, Matan Mazor1; 1University of Oxford
Presenter: Maya Schipper
When things are perceived clearly they can be detected with confidence. But when can one be confident that something is absent? Here, we used a meta-perceptual illusion to show that confidence in absence scales the belief that a target would have been visible if present. We manipulated stimulus size in two near-threshold detection tasks with confidence ratings. While participants believed that detection was easier for large stimuli (measured with prospective confidence ratings and post-experiment debriefing), their perceptual sensitivity was in fact higher for small stimuli. Accordingly, while confidence in presence scaled with true visibility, confidence in absence scaled with beliefs about visibility. Moreover, the effect of size on confidence in absence, but not in presence, correlated with a meta- perceptual parameter from an ideal observer model. Overall, we conclude that confidence in absence tracked model-derived expectations about the visibility of counterfactual stimuli.
Probability Distortions Reflect Boundary Repulsions in Noisy Inference
Talk 2, 10:10 am – Saurabh Bedi1, Gilles de Hollander1, Christian C. Ruff1; 1University of Zurich
Presenter: Saurabh Bedi
Probability distortions—the apparent overweighting of small probabilities and underweighting of large ones—is central to decision-making under risk, but its normative and mechanistic origins remain unclear. Traditionally seen as irrational, we propose that probability distortion instead emerges from optimal but noisy inference on bounded quantities. In our proposed account, repulsions arise at natural boundaries of probabilities (0 and 1) due to both resource-rational efficient encoding and Bayesian optimal decoding. Our account predicts that experimental manipulations of boundaries and noise should systematically reshape both probability distortions and behavioral variability, in both risky choice and probability perception. We confirm these predictions in three pre-registered experiments. Our findings reframe probability distortion as a normative consequence of bounded noisy inference and offer a unified mechanistic explanation for its presence across valuation and perception.
Metacognitive Judgements of Confidence in Effortful Task Are Susceptible to Momentary Fluctuations of Fatigue
Talk 3, 10:20 am – Katarzyna Dudzikowska1, Nikita Mehta, Matthew Apps; 1University of Birmingham
Presenter: Katarzyna Dudzikowska
Classical theories describe prospective confidence as a readout of the probability of success based on reinforcement history. However, research suggests that confidence may be susceptible to fatigue despite continued success. Here, we test if confidence changes throughout an effortful task and whether the moment-to-moment fluctuations in fatigue can account for these changes. Participants exerted physical effort (30, 30, 48% of maximum voluntary contraction - MVC) and reported confidence in their ability to succeed. Across three studies, we show that confidence declines over time and fluctuates on a trial-by-trial basis, despite consistently successful performance. Decreases in confidence and increases in fatigue ratings (Study II) were both related to exerted effort. We introduce a novel computational model in which latent fluctuations in fatigue drive changes in confidence and show that it outperforms models based solely on past performance or time on task. We also show that the relationship between confidence and fatigue is distinct from boredom (Study III). Thus, computations of confidence are susceptible to fluctuating levels of fatigue, even when the probability of success remains high.
Proactive Counterfactual Inference in Flexible Decision Making
Talk 4, 10:30 am – Peiyue Liu1, Weiwen Lu, Xiaohong Wan1; 1Beijing Normal University
Presenter: Peiyue Liu
In complex, uncertain environments, individuals must flexibly integrate multiple sources of information to adapt to changing task demands. While prior research has primarily focused on confidence formation and rule inference within a single task, less is known about how information across multiple tasks is integrated. Here,we designed an experiment to address this question by asking participants to infer task rules while switching between two tasks. We found that participants were able to maintain cognitive control in the face of task-irrelevant information, ensuring smooth task performance. However, when such irrelevant information could potentially support task rule inference, individuals can flexibly adjust their strategies, leveraging this information to optimize the decision-making process. Participants' beliefs about the current task rule (rule belief) modulated this cognitive flexibility, influencing how they prioritized, processed, and integrated information. Neural data revealed that the dorsal anterior cingulate cortex (dACC) plays a central role in these processes, specifically in: (1) encoding both task-relevant and task-irrelevant evidence; (2) updating rule beliefs and (3) modulating functional connectivity with the human fourth visual area and middle temporal area (hV4/MT). To probe the underlying mechanisms, we trained a recurrent neural network (RNN) model. We showed that within a trial, these neurons operate under an attention bottleneck, which serves as a constraint and mimics the potential attention-splitting process observed in humans. As with human participants, the effect of task-irrelevant information on rule belief updating was observed, but with a stronger effect. Together, these findings reveal a neural process in the human brain, particularly in the dACC, for integrating and updating beliefs about tasks, and how individuals flexibly adjust their strategies based on both relevant and irrelevant information.
A single general strategy supports flexible behavior in response to both gradual and sudden changes
Talk 5, 10:40 am – Jessica Passlack1, Maria K Eckstein2; 1University of Edinburgh, University of Edinburgh, 2Google DeepMind
Presenter: Jessica Passlack
Most decisions we make occur in an ever-changing world, where flexibility is crucial for making intelligent decisions. However, despite a large array of different behavioral tasks being used to investigate flexible behavior, existing research has focused primarily on individual tasks in isolation, preventing understanding of whether there is a general strategy underlying flexible behavior. Here, we designed a set of tasks that spanned two hallmarks of behaviors that require flexible decision making: gradual and sudden changes. We found that behavioral correlates were similar across both types of changes and that a neural network could learn a general strategy to support behavior. Our results indicate that a neural network approach in combination with across-feature task design allows for determining whether a general underlying strategy supports different aspects of flexible behavior. This opens the door for interrogating the building blocks of the general strategy underlying flexible behavior.