Bayesian Modeling of Context-Specific Goal-Directed and Habitual Control in Sequential Decision Making

It is an open question how context-specific habits develop alongside goal-directed planning. We examined action sequence repetition, contingency change, and context switching using a three-step optimization task with interleaved planning and habit contexts. During an initial acquisition phase, optimal sequences varied in the planning context but remained constant in the habit context. When contingencies changed, the original habit sequence became suboptimal and another became optimal. Action sequence repetition improved response speed and optimal-choice performance in the habit context relative to the planning context. Following the contingency change, both advantages initially reversed. Optimal-choice performance in the habit context subsequently recovered but did not significantly exceed that in the planning context, while responses remained slower in the habit context. Context switches reduced optimal-choice performance in both directions, with additional slowing when switching from habit to planning. The Bayesian contextual control model captured choice patterns through learned action sequence priors, goal-directed evaluation, and context inference. Allowing a context-stability bias to vary across participants did not improve model performance. These findings suggest that sequence repetition facilitates performance under stable contingencies while impairing adaptation to change.