Continuous and compositional mixing of goals in continuous choice

Animals readily make value-based decisions between discrete and static items, with decades of work elucidating neural circuits involved in such choices (Rangel et al., 2015). Much less is known about naturalistic decision-making contexts such as goal pursuit (e.g., hunting). Dynamic choices require determining the value of blending different goal-pursuit strategies, switching goals, and continuously monitoring the actions and goals most valuable (Yoo et al., 2021). Despite its prevalence, there is a lack of methods to identify continuous decision processes and thus a dearth of understanding their neural implementation. Using a task where monkeys continuously pursue several evading prey, we develop a novel control-theoretic decomposition of subjects’ moment-to-moment pursuit behavior that can infer subjects’ continuous decision variable based on the evolving relative value of goals. The model revealed most pursuit strategies involved a blending of goals, with subjects often exhibiting changes of mind between goals. Neurons in dorsal anterior cingulate cortex (dACC) encoded this model-based decision variable independent of other task variables (e.g., target distances). To elucidate the population coding of the decision variable, we examined changes of mind between goals. dACC exhibited low-dimensional coding of the decision variable, with the same code being used to read-out goal blending across switch directions. We also found a decision signal that ramps during change of mind initiation, exhibiting hallmarks of an evidence accumulation signal. Our modeling opens a new avenue for studying naturalistic decision-making from a theoretical and model-based perspective, with our behavioral results showing how continuous choice is composed from a previously unidentified mixtures of control policies. We show identification of internal control processes are necessary to understanding decision making, allowing us here to reveal previously