Todd Hudson^1,2, Mahya Beheshti^1,2, JR Rizzo^1,2; ¹NYU Grossman School of Medicine, ²Tandon NYU School of Engineering

From a control-theoretic perspective, smooth pursuit is an inherently unstable tracking problem: even small temporal errors accumulate until the continuous controller fails and generates a discrete corrective reset. Adaptive catch-up and backtracking saccades are therefore computational consequences of imperfect tracking. In healthy observers, they reveal visuomotor prediction, delay compensation, and uncertainty; in clinical populations they become exaggerated, delayed, or misdirected, providing a functional biomarker of impaired controller dynamics, processing, and/or models. Quantifying these events, however, requires solving a difficult inverse problem: the observed data trace reflects the sum of outputs from the continuous pursuit controller and saccadic corrective mechanism - a combination that is not identifiable without a generative model. We develop a generative Bayesian framework for joint inference of latent pursuit dynamics and saccadic correction. Pursuit is represented as a 3-state linear dynamical system (position, velocity, acceleration), capturing the smooth but uncertain internal trajectory. A heteroskedastic observation model increases variance during saccadic intervals, enabling the forward estimator to preserve the continuity of latent pursuit while avoiding contamination from large saccadic deviations. Saccades arise from a mechanistic phase-error policy: instantaneous pursuit–target phase discrepancy drives a probabilistic hazard for generating adaptive saccades, yielding a complete generative account of pursuit failure. We conducted a stress test of the model at the frequency of our control and patient data (0.25 Hz), with simulations spanning observation noise from near-zero to a significant proportion of pursuit amplitude along with a range of positive/negative imposed phase errors. For each condition, we computed timing error distributions, timing SD, amplitude gain, and detection fidelity. Results reveal large identifiability regions where Bayesian inference reliably reconstructs pursuit and recovers saccade timing. We will present these computational results alongside human pursuit data, demonstrating how this generative-inference approach yields calibration-robust, model-based biomarkers of visuomotor control in both healthy control and clinical populations.