Decision Making

Talk Session: Tuesday, May 19, 2026, 10:45 am – 12:15 pm, Talk Room 2
Moderator: Alan Stocker, University of Pennsylvania

Schedule of Events | Search Abstracts | Symposia | Talk Sessions | Poster Sessions

Talk 1, 10:45 am, 52.21

A re-analysis of 55 experiments reveals two factors that drive confidence for error trials

Herrick Fung¹ (), Xinyue Li¹, Dobromir Rahnev¹; ¹School of Psychology, Georgia Institute of Technology

Confidence on correct trials reliably increases as tasks become easier. On error trials, however, empirical results diverge: some studies report higher confidence for easier stimuli, others report the opposite. This inconsistency suggests the presence of overlooked factors that shape confidence computations. Here, we reviewed and reanalyzed 55 experiments (N = 2,804), yielding 77 independent analyses of how error-trial confidence varies with difficulty. We found 41 decreases, 26 increases, and 10 flat patterns. Two factors fully explained this heterogeneity: (1) type of difficulty manipulation and (2) availability of post-decisional information. First, manipulations that increase stimulus reliability (e.g., using higher contrast or lower noise) tend to produce confidence increases for easier stimuli. In contrast, manipulations that vary the judged stimulus feature (e.g., manipulating tilt in an orientation discrimination task) tend to produce confidence decreases. Further analyses suggest that this dichotomy is mediated by whether the difficulty of the manipulation is perceptually visible: visible manipulations provide implicit cues to trial-by-trial difficulty, which are then used to adjust the final confidence rating. Second, the presence of post-decisional sensory evidence strongly drives confidence downward for easier stimuli on error trials, regardless of the difficulty manipulation. This is likely because post-decision evidence typically contradicts the initial choice on errors; since this conflict is stronger for easier stimuli, confidence decreases more steeply for those stimuli. Together, these two factors account for all 77 observed patterns. Our results reveal that a core signature of metacognition, namely how confidence changes with task difficulty on error trials, is not fixed but is determined by factors that are not accounted for by most models and often ignored in experimental design. Finally, we show that standard computational models can be extended to incorporate both factors and include simulations that reproduce the effects observed across all analyses.

This work was supported by the National Institute of Health (award: R01MH119189).

Talk 2, 11:00 am, 52.22

Sequential evidence accumulation adapts to changing levels of evidence reliability

Mengting Fang¹ (), Alan Stocker; ¹University of Pennsylvania

Perceptual decision making often involves integrating evidence over time. Our prior work showed that this process is governed by a resource-rational principle where the brain actively balances task performance with cognitive effort. For a fixed level of evidence reliability, participants maintained a stable trade-off strategy. However, our previous study did not test whether participants adapt their strategy when evidence reliability changes. Given the brain’s ability to actively control the accumulation process, we considered two competing hypotheses: (1) the system increases encoding precision when evidence is reliable and reduces it when evidence is less reliable, or (2) the opposite, the system increases encoding precision to compensate for low-quality evidence and lowers it when reliability is high. To test these hypotheses, we conducted a visual estimation task where participants inferred the angular position of an unknown source from 8 sequentially presented, normally distributed stimulus samples. We manipulated sampling noise across three levels. The testing phase included 9 blocks across 3 visits in a balanced Latin square design, with trial-by-trial feedback and a performance-dependent bonus. Noise levels were not disclosed in each block. Results showed that participants' estimation performance decreased as sampling noise increased. Fits with our normative resource-rational model showed that encoding precision scaled with sampling noise, supporting our first hypothesis that the brain actively adjusts encoding precision to match external sensory reliability. Circular regression analysis further revealed that the variance of normalized temporal weights decreased as noise increased, reflecting shifts in temporal weighting strategies (e.g., from recency to uniform/primacy). Changes in trade-off parameters also showed greater tolerance for effort costs under higher noise, partly offsetting the reduced encoding precision. Together, these findings demonstrate that evidence accumulation is governed by a performance-effort trade-off strategy that adapts sensory encoding precision to the reliability of the evidence.

This work was supported by the NSF CRCNS grant IIS-1912232 to A.A.S

Talk 3, 11:15 am, 52.23

How history manifests in the macaque cortical hierarchy

Jongmin Moon¹, Zoe M. Boundy-Singer², Julie A. Charlton³, Robbe L. T. Goris¹; ¹The University of Texas at Austin, ²Massachusetts Institute of Technology, ³Princeton University

Perceptual judgments of ambiguous stimuli exhibit intricate history-dependent biases. These effects arise from a sophisticated computational process that is both efficient and effective and is hypothesized to involve sensory as well as decision-making circuits. To test this hypothesis, we studied neural population activity in the visual and prefrontal cortices of macaque monkeys while they performed a perceptual decision-making task. The animals judged whether a visual stimulus was oriented clockwise or counterclockwise relative to vertical and communicated their decision with a saccade. As expected, perceptual choices of each monkey exhibited a complex history-dependency consisting of a mixture of stimulus repulsion and choice attraction. How does this behavioral pattern emerge from the underlying neural activity in cortical circuits? To address this question, we analyzed the one-dimensional linear component of neural population activity in V1 (two monkeys) and FEF (two monkeys) that pertains to the formation of a perceptual decision. Neural activity in both areas was strongly influenced by the current sensory stimulus. In contrast, preceding sensory and choice events had a prominent effect on neural activity in FEF, but only a restricted effect in V1. Analysis of the dynamic evolution of FEF activity further revealed a temporal distinction between stimulus and choice history effects. The preceding choice already influenced FEF activity before the stimulus onset, while the influence of the preceding stimulus only manifested after the stimulus onset. This temporal distinction is consistent with theoretical proposals that attribute choice attraction and stimulus repulsion to two distinct computational components of decision-making: perceptual expectation and sensory adaptation. Together, our findings illuminate the physiological basis of the influence that recent experience exerts on current perceptual impressions. In doing so, our work illustrates how a targeted analysis of neural population activity in different brain regions can reveal the neurobiological basis of fundamental brain computations.

NIH R01 EY032999

Talk 4, 11:30 am, 52.24

Introspecting visual biases

Noa Perlmutter¹, Chaz Firestone¹, Ian Phillips¹; ¹Johns Hopkins University

We are aware of the world and its properties; for example, we can see the number, size, and motion of objects around us. Are we also aware of the internal processes that generate these percepts? Whereas it is controversial whether higher-level cognitive processes are accessible to introspection (e.g., knowing why we make various choices), basic perceptual biases are widely assumed to be closed off from introspection. In contrast to this consensus, here we reveal successful introspection of three classic effects in vision and visual memory: numeric underestimation, size contrast, and representational momentum. Experiment 1 investigated numeric underestimation: Subjects briefly saw an array of 11-100 dots, and estimated their numerosity; on a subset of trials, subjects were then asked whether they thought they had underestimated or overestimated. Experiment 2 investigated the Ebbinghaus illusion: Subjects saw a target circle surrounded by flankers of variable size, and adjusted a second circle to match the target; subjects were then asked whether they thought they had underestimated or overestimated. Experiment 3 investigated representational momentum: Subjects saw a fish glide across the screen and disappear, and estimated its final seen location; subjects were then asked whether they thought their estimate was too far left or right. We replicated all three biases: Subjects underestimated numerosity, were influenced by flankers, and extrapolated motion. Strikingly, however, all three experiments also revealed awareness of the effects themselves. For example, in Experiment 1, subjects correctly answered that they underestimated numerosity. And in Experiments 2 and 3, they showed trial-by-trial awareness of their biases, as computed by detection-theoretic statistics. These findings go beyond successful metacognition of performance (i.e., knowing whether one is performing well or poorly) to awareness of directional effects themselves, suggesting deeper and subtler access to internal mental processes than traditionally assumed.

Talk 5, 11:45 am, 52.25

Disentangling perceptual and analytical processes in solving Euclidean and non-metric traveling salesperson problems

Arman Baradaran¹, Jacob VanDrunen², Sébastien Hélie³, Zygmunt Pizlo¹; ¹University of California, Irvine, ²Westminster Seminary California, ³Purdue University

Human performance on classic traveling salesperson problems (TSPs) depends primarily on perceptual processes. Participants produce near-optimal tours by relying on hierarchical clustering followed by global-to-local tour refinement, though it is unclear to what extent analytical skills like forward planning are involved. We investigated this question using non-metric TSPs in which half of the cities were assigned one of two colors, and switching colors incurred a cost equal to twice the Euclidean distance between cities. Performance on 30 TSPs with 50 cities each was evaluated through two dependent variables: (a) error with respect to optimal tours and (b) number of color switches. Performance varied significantly more in non-metric TSPs than in Euclidean TSPs, with percent errors ranging from 3% to 32%. Several participants adopted a minimal-switch heuristic, producing tours that emphasized clustering of same-colored cities and making only 2 color switches. Others made deviations from color-based clustering that discounted the switching cost, with one participant even making 22 color switches on average. We then compared participants’ performance to two versions of a multiresolution graph pyramid model. The first version, which operated on spatial clustering and used cheapest insertion to consider switching costs, underperformed compared to most humans and overestimated the number of switches. The second version had a front-end that used a 3D Euclidean approximation produced by applying multidimensional scaling (MDS) to the non-metric pairwise distances. This latter version mimicked the behavior of the minimal-switch participants, and after parameter adjustments in the MDS process, it also replicated the performance of more analytically-inclined participants. These findings suggest that combining a multiresolution pyramid representation with 3D MDS can capture human behavior in non-metric TSPs. More broadly, non-metric TSPs may serve as a useful proxy for studying how humans integrate perceptual and analytical processes in complex problem solving domains such as chess.

Talk 6, 12:00 pm, 52.26

Hierarchical representations in visual structure learning interact with stimulus-dependent metacognitive dynamics

Rochelle Kaper^1,2,3, Megan A.K. Peters^1,2,3,4,5,6; ¹Department of Cognitive Sciences, University of California Irvine, ²Center for Theoretical Behavioral Sciences, University of California Irvine, ³Center for the Neurobiology of Learning & Memory, University of California Irvine, ⁴Department of Logic & Philosophy of Science, University of California Irvine, ⁵Department of Experimental Psychology, University College London, ⁶Program in Brain, Mind, & Consciousness, Canadian Institute for Advanced Research

Our daily lives afford the ability to construct internal models of multiple spatiotemporal patterns without feedback, in order to predict incoming information and generalize to novel contexts; this process is called visual structure learning. Along with prediction errors, metacognitive error signals guide learning in the absence of external feedback. It is unknown how metacognitive capacity may change over time throughout learning, and how such dynamics would affect learning across structures and stimuli differing in abstractness under uncertainty. To address these gaps, we developed a novel probabilistic visual structure learning paradigm where we probed the dynamics and representation of learning and metacognition under uncertainty. Participants learned probabilistic co-occurrences of item pairs (AA’, BB’…) and a probabilistic sequence of those co-occurrences (AA’→BB’...). After each learning phase where participants viewed probabilistic streams of item pairs, participants completed tests of explicit structure learning (forced-choice accuracy) and metacognitive capacity (accuracy-confidence correspondence). Participants completed 2 sessions in counterbalanced order: one with low-level perceptual fractals (n=56), and the other with familiar line drawings (n=57). Mixed effects linear models revealed significantly higher sequence accuracy compared to co-occurrences for fractals, and vice versa for line drawings. We also observed significant metacognitive capacity increases across learning for fractals but not line drawings, and metacognitive capacity was higher for sequence structure despite greater task accuracy in fractals. We modeled the representation of hierarchical structure information using analytical successor representations, revealing a potential tradeoff between metacognitive capacity, representation, and usage of sequential information in learning the co-occurrences that differ among stimuli. Our results reveal distinct learning rates, metacognitive capacity dynamics, and representations of hierarchical information across structures and stimuli varying in abstractness and familiarity, motivating novel task designs that can assess how different stimuli and structures alter how hierarchical information is represented and varying metacognitive capacity dynamics throughout learning.

Funding Source: Canadian Institute for Advanced Research / Research Corporation for Science Advancement