Modeling decision dynamics disentangles working memory, cognitive control and reinforcement learning and reveals clinical differences
8 affiliations
- 1Department of Cognitive & Psychological Sciences, Carney Institute for Brain Science, Brown University, Providence, Rhode Island, United States of America
- 2Department of Psychology, University of California, Santa Cruz, California, United States of America
- 3Department of Psychology, University of California, Irvine, California, United States of America
- 4Department of Psychiatry & Behavioral Neuroscience, University of Chicago, Chicago, Illinois, United States of America
- 5Department of Psychiatry, University of Maryland School of Medicine, Baltimore, Maryland, United States of America
- 6Department of Psychology, University of California, Davis, California, United States of America
- 7Department of Psychology, University of Minnesota, Minneapolis, Minnesota, United States of America
- 8Department of Psychology, Washington University in St. Louis, St. Louis, Missouri, United States of America
Abstract
Various neurocognitive processes contribute to reinforcement learning (RL) and decision making, requiring careful task designs and computational modeling to disentangle them. People rely on capacity-limited working memory (WM) to rapidly and flexibly adapt behavior during learning, together with slower incremental RL processes that support robust long-term retention. The RLWM paradigm was developed to isolate these contributions by manipulating WM demands during learning. Empirical and computational advances support the separability of such processes, but only imperfectly, because a given choice can often equally be attributed to use of WM vs. RL. Here, we show that jointly modeling decision dynamics – accounting not only for choices but also response time distributions, together with hierarchical Bayesian parameter estimation – improves the ability to disentangle separable learning and cognitive control processes. Despite added complexity, models with decision dynamics improved parameter recovery and yielded accurate out-of-sample prediction for choices in a held-out test phase after learning. In contrast, models fit only to choices produced an inflation of RL learning rates and failed the out-of-sample test. Moreover, joint modeling also revealed a novel neurocognitive process by which participants proactively widen decision boundaries to increase response caution under increased WM load. Applying our model to patients with schizophrenia revealed a deficit in proactive control mechanisms as well as slowed incremental RL, which were masked by previous choice-only models, while replicating previously established WM deficits. In summary, we show how joint modeling enables accurate, model-based and mechanism-oriented computational phenotyping of psychopathologies. Author summary: Humans learn from feedback using multiple brain systems in parallel: reinforcement learning, which gradually strengthens successful behaviors through trial and error, and working memory, which supports rapid learning by holding recent information in mind. These systems often drive the same outward behavior, making their contributions notoriously difficult to disentangle. Here, we show that analyzing not only what people choose but how quickly they respond fundamentally changes what we can learn about these processes. We introduce a computational framework that captures both learning and decision dynamics in a single model, and discover that people proactively slow down their responses when they anticipate greater memory demands, a cognitive control strategy not previously identified in this context. As an independent check, we used the model to predict a later memory-free phase it had never been fit to; it succeeded, while choice-only predictions diverged from people’s behavior. This model also reveals that individuals with schizophrenia show both a blunting of gradual learning and a failure of this proactive adjustment, deficits that were entirely hidden from prior analyses based on choices alone. Our approach is broadly extensible, allowing researchers to formulate and compare theories that jointly characterize learning processes and decision dynamics within a unified model.
Related
- Multi-parametric spectral modeling disentangles synaptic kinetics and neural excitability from non-invasive electrophysiological recordings
- Assessing reinforcement learning and vigor with an online decision-making battery
- Impaired Reinforcement Learning Underlying Explore-Exploit Decision Making in Theft Recidivists
- Effort and substance use: differentiating tobacco use through reinforcement learning of effort based decision making
- Individual differences reveal shared computational structure in perceptual decision-making
BibTeX
@article{bera2026modeling,
title = {{Modeling decision dynamics disentangles working memory, cognitive control and reinforcement learning and reveals clinical differences}},
author = {Krishn Bera and Alexander Fengler and Megan A. Boudewyn and Cameron S. Carter and Molly A. Erickson and James M. Gold and Steven J. Luck and J. Daniel Ragland and Andrew P. Yonelinas and Angus W. MacDonald III and Deanna M. Barch and Michael J. Frank},
journal = {PLOS Computational Biology},
year = {2026},
month = sep,
doi = {10.1371/journal.pcbi.1014796},
url = {https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1014796},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Modeling decision dynamics disentangles working memory, cognitive control and reinforcement learning and reveals clinical differences
AU - Krishn Bera
AU - Alexander Fengler
AU - Megan A. Boudewyn
AU - Cameron S. Carter
AU - Molly A. Erickson
AU - James M. Gold
AU - Steven J. Luck
AU - J. Daniel Ragland
AU - Andrew P. Yonelinas
AU - Angus W. MacDonald III
AU - Deanna M. Barch
AU - Michael J. Frank
PY - 2026
DA - 2026/09/30
JO - PLOS Computational Biology
DO - 10.1371/journal.pcbi.1014796
UR - https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1014796
AB - Various neurocognitive processes contribute to reinforcement learning (RL) and decision making, requiring careful task designs and computational modeling to disentangle them. People rely on capacity-limited working memory (WM) to rapidly and flexibly adapt behavior during learning, together with slower incremental RL processes that support robust long-term retention. The RLWM paradigm was developed to isolate these contributions by manipulating WM demands during learning. Empirical and computational advances support the separability of such processes, but only imperfectly, because a given choice can often equally be attributed to use of WM vs. RL. Here, we show that jointly modeling decision dynamics – accounting not only for choices but also response time distributions, together with hierarchical Bayesian parameter estimation – improves the ability to disentangle separable learning and cognitive control processes. Despite added complexity, models with decision dynamics improved parameter recovery and yielded accurate out-of-sample prediction for choices in a held-out test phase after learning. In contrast, models fit only to choices produced an inflation of RL learning rates and failed the out-of-sample test. Moreover, joint modeling also revealed a novel neurocognitive process by which participants proactively widen decision boundaries to increase response caution under increased WM load. Applying our model to patients with schizophrenia revealed a deficit in proactive control mechanisms as well as slowed incremental RL, which were masked by previous choice-only models, while replicating previously established WM deficits. In summary, we show how joint modeling enables accurate, model-based and mechanism-oriented computational phenotyping of psychopathologies. Author summary: Humans learn from feedback using multiple brain systems in parallel: reinforcement learning, which gradually strengthens successful behaviors through trial and error, and working memory, which supports rapid learning by holding recent information in mind. These systems often drive the same outward behavior, making their contributions notoriously difficult to disentangle. Here, we show that analyzing not only what people choose but how quickly they respond fundamentally changes what we can learn about these processes. We introduce a computational framework that captures both learning and decision dynamics in a single model, and discover that people proactively slow down their responses when they anticipate greater memory demands, a cognitive control strategy not previously identified in this context. As an independent check, we used the model to predict a later memory-free phase it had never been fit to; it succeeded, while choice-only predictions diverged from people’s behavior. This model also reveals that individuals with schizophrenia show both a blunting of gradual learning and a failure of this proactive adjustment, deficits that were entirely hidden from prior analyses based on choices alone. Our approach is broadly extensible, allowing researchers to formulate and compare theories that jointly characterize learning processes and decision dynamics within a unified model.
SN - 1553-7358
ER -