Impaired Reinforcement Learning Underlying Explore-Exploit Decision Making in Theft Recidivists
- 1Kyoto University
Abstract
Larceny imposes profound societal and economic burdens; however, punitive judicial measures frequently fail to deter recidivism. The neurobehavioral mechanisms driving habitual offending, whether instrumental or kleptomanic, in theft recidivists remain poorly understood. In this study, we investigated explore-exploit decision-making and underlying reinforcement learning architectures in theft recidivists with a 4-arm bandit task while prefrontal cortex (PFC) hemodynamics were continuously monitored using functional near-infrared spectroscopy (fNIRS). Model-free behavioral analyses revealed that non-kleptomanic (TR-K) but not kleptomanic (TR+K) theft recidivists accumulated significantly higher cumulative regret and made fewer optimal choices compared to control individuals without criminal records (CT). Model-based analyses through Variational Bayesian Analysis identified Q-learning with decay model as the optimal computational fit. Parameter extraction using the model demonstrated that the TR-K group exhibited a lower learning rate than CT and TR+K groups, indicating a learning deficit in updating action values following environmental feedback. fNIRS tracking of trial-by-trial latent reinforcement variables revealed that while PFC activity was modulated by these variables, group differences were characterized by static baseline hemodynamic shifts rather than rewirings of value-tracking neural circuits. These findings suggest that non-kleptomanic recurrent thefts may be associated with impaired reinforcement learning mechanisms, challenging current punitive deterrence models.
Related
- Impaired value estimation over space explains turn-induced freezing of gait in Parkinson's disease: A cortico-basal ganglia-inspired reinforcement learning model
- Virtual reality, machine learning, and statistical modeling reveal an association between multiple concussion history and impaired perceptual decision-making
- Assessing reinforcement learning and vigor with an online decision-making battery
- Effort and substance use: differentiating tobacco use through reinforcement learning of effort based decision making
- Modeling decision dynamics disentangles working memory, cognitive control and reinforcement learning and reveals clinical differences
BibTeX
@misc{goto2026impaired,
title = {{Impaired Reinforcement Learning Underlying Explore-Exploit Decision Making in Theft Recidivists}},
author = {Y. Goto and S. Yoshino and C. Kita and M. Won and Y.-A. Lee},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.11.750836},
url = {https://www.biorxiv.org/content/10.64898/2026.09.11.750836v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Impaired Reinforcement Learning Underlying Explore-Exploit Decision Making in Theft Recidivists
AU - Y. Goto
AU - S. Yoshino
AU - C. Kita
AU - M. Won
AU - Y.-A. Lee
PY - 2026
DA - 2026/09/11
PB - bioRxiv
DO - 10.64898/2026.09.11.750836
UR - https://www.biorxiv.org/content/10.64898/2026.09.11.750836v1
AB - Larceny imposes profound societal and economic burdens; however, punitive judicial measures frequently fail to deter recidivism. The neurobehavioral mechanisms driving habitual offending, whether instrumental or kleptomanic, in theft recidivists remain poorly understood. In this study, we investigated explore-exploit decision-making and underlying reinforcement learning architectures in theft recidivists with a 4-arm bandit task while prefrontal cortex (PFC) hemodynamics were continuously monitored using functional near-infrared spectroscopy (fNIRS). Model-free behavioral analyses revealed that non-kleptomanic (TR-K) but not kleptomanic (TR+K) theft recidivists accumulated significantly higher cumulative regret and made fewer optimal choices compared to control individuals without criminal records (CT). Model-based analyses through Variational Bayesian Analysis identified Q-learning with decay model as the optimal computational fit. Parameter extraction using the model demonstrated that the TR-K group exhibited a lower learning rate than CT and TR+K groups, indicating a learning deficit in updating action values following environmental feedback. fNIRS tracking of trial-by-trial latent reinforcement variables revealed that while PFC activity was modulated by these variables, group differences were characterized by static baseline hemodynamic shifts rather than rewirings of value-tracking neural circuits. These findings suggest that non-kleptomanic recurrent thefts may be associated with impaired reinforcement learning mechanisms, challenging current punitive deterrence models.
ER -