Fig. 1: Rats accumulate and discount evidence in a dynamic accumulation task.

a Schematic showing task events and timing. The center port is illuminated by an LED. The rat pokes its nose into the port to initiate playback of randomly timed auditory clicks from speakers on either side. Clicks on each side are generated with different underlying Poisson rate parameters that depend on a hidden environmental state. The stimulus duration is drawn from a uniform distribution between 500 and 2000 ms. During that time the hidden state changes stochastically at a fixed hazard rate, h = 1Hz. At the end of the stimulus presentation, the center LED turns off and reward is baited in the side port corresponding to the final state. b Schematic of the evolution of the accumulation model on an example trial. Three example accumulation traces are shown for different instantiations of the noise applied at each time point (σ_a) and the noise applied to each click (σ_s). Neighboring clicks can either depress or facilitate each other according to the adaptation parameters (ϕ and τ_ϕ). The evidence discounting rate (λ) determines how quickly the decision variable a decays back to zero. At the end of the trial, a choice is made by comparing the decision variable to the decision boundary parameter B. c Frequency of state changes per trial across all rats' datasets. d Example psychometric plot showing the probability that the rat chooses “go right” as a function of the ideal observer log-odds supporting a “go right” choice. Rat data (black points) is overlaid on predictions of the accumulation model with parameters fit to this rat (red traces). Errorbars for rat data represent 95% binomial confidence intervals around the mean (n = 92,468 trials from 252 sessions). e Example final state chronometric plot for the same dataset as in (d). Accuracy (mean with 95% binomial confidence intervals) is plotted as a function of the duration of a trial’s final state and the number of state changes in a given trial. f Psychophysical reverse correlation kernel for the same dataset as in (d) and (e). Green and blue patches indicate strength (mean ± s.d.) of evidence favoring rightward choice as a function of time until the trial ends for rightward and leftward choices, respectively. The red patches are corresponding predictions from the accumulation model. g Discounting parameters for each rat in this study (red points) compared to each rat in a previously published stationary environment (lilac points; Brunton et al.⁸). Group medians are plotted as black horizontal lines. Source data are provided as a Source Data file.