Factorized Prefrontal Geometry of Goal and Uncertainty Explains Flexible yet Stable Human Goal Pursuit

Yoondo Sung, Mattia Rigotti, Sang Wan Lee

Keywords: Representational geometry, cognitive flexibility, stability, uncertainty, LPFC, MVPA, fMRI, reinforcement learning

Abstract: A central challenge for adaptive agents is achieving behavioral flexibility without losing stability, especially during goal-directed learning in uncertain environments. Flexibility enables rapid adjustment when goal changes, while stability prevents overreaction to noisy outcomes--- these properties often trade off. To examine how the brain resolves this dilemma, we combined reinforcement learning simulations, human behavior, and fMRI study during a sequential decision task with varying goals and uncertainty. Simulations showed that model-free agents suffered a flexibility–stability trade-off, while model-based agents were flexible but with a range of stability. Human participants, in contrast, displayed both high flexibility and high stability. fMRI analyses revealed the underlying representational mechanism: goals and uncertainty were encoded as partly independent dimensions in lateral prefrontal and orbitofrontal cortex. Importantly, the separability and robustness of goal representations in these regions correlated with individual behavioral flexibility and stability, pointing to a geometric account of robust goal pursuit beyond value-learning models.

Submission Number: 145