A prevailing theory in goal-directed decision-making is that the brain accumulates evidence over time to make perceptual decisions. However, even after two decades of studies on evidence accumulation, it is still unclear how neuronal populations cooperate to dynamically represent decision variables. This is because of a lack of tools for gaining experimental control over variables of interest to test their relationship to behavior. To address this gap, we develop a state-space control framework to causally interact with decision-related latent variables. To expose decision variables to target for control, we trained mice to perform a 2AFC accumulation of evidence task and recorded from rostrolateral (RL) area of posterior parietal cortex (PPC). We modeled population activity with recurrent switching linear dynamical systems (rSLDS) and found that competing 2D latent accumulation dynamics explained both electrophysiological and behavioral data. From the identified latent dynamics during stimulus presentation, we were able to infer the animals’ eventual choice on a trial-by-trial basis, including when the choice was incorrect. Leveraging the piecewise-linear structure of rSLDS models, we developed an input-constrained, adaptive model predictive control (MPC) framework for stabilizing latent variables about target trajectories. In simulations, we were able to control decision variables moment-by-moment under realistic experimental constraints using one excitatory and one inhibitory exogeneous channel, representing a pair of opsins, even in the presence of additional obscuring dynamics. Additionally, we determined the reachable set of states under input constraints given model parameters, which provides a method of evaluating experimental design of input channels. Our control framework enables direct interaction with latent decision variables to test their causal link to behavior, as well as a method for choosing experimental inputs with enough complexity to drive desired trajectories. More generally, this control framework can be applied in various neural systems to test computation-through-dynamics hypotheses.