Hiroichi Yoshida^1,2,3,5, Zhongxuan Wu^1,3,4,5, Xue-Xin Wei^1,2,3,4,5, Eyal Seidemann^1,2,3,5; ¹Department of Neuroscience, The University of Texas at Austin, ²Department of Psychology, The University of Texas at Austin, ³Center for Perceptual Systems, The University of Texas at Austin, ⁴Center for Learning and Memory, The University of Texas at Austin, ⁵Center for Theoretical and Computational Neuroscience, The University of Texas at Austin

Introduction: Perceptual decisions are thought to arise from neural dynamics in decision-related areas such as the frontal eye field (FEF) that transform sensory evidence into choices. These dynamics have been studied using stochastic stimuli that promote evidence accumulation. However, it remains unclear whether static stimuli presented on the timescale of a single natural fixation also engage an integrative, continuous process (e.g., drift-diffusion), which may benefit performance in the presence of internal noise, or a near instantaneous, discrete state transition (stepping). Methods: Using Neuropixels probes, we recorded spiking activity from 244-392 single- and multi-units in the FEF of two monkeys performing a coarse orientation discrimination task, in which they reported a binary saccadic choice based on the orientation of a briefly presented Gabor patch at varying contrasts. We analyzed the response properties of these neurons at the level of individual units and neural populations. Results: Our analysis of individual units revealed three subpopulations: choice-selective, stimulus-selective, and units selective for both stimulus and choice. We next trained linear classifiers on the population activity to predict choice. Trial-averaged decision variable (DV) traces for the two choices gradually diverged over time. While this result suggests that choice-predictive information increases as trials progress towards saccade onset, it can be consistent with both drift-diffusion and stepping dynamics. To further adjudicate between the two, we developed a computational approach based on Hidden Markov Models. From single trial DV traces, we inferred the number of latent states and the transition probabilities between them. Our results show that the data are best explained by models with a substantial number of states and smooth state-to-state transitions. Conclusions: Our current results suggest that neural population responses in FEF during decisions about brief static stimuli are more consistent with drift-diffusion dynamics than stepping dynamics.