Fig. 1: Attractor models of decision-making were tested by recording from the rat frontal cortex and striatum.

a, Rats were trained to accumulate auditory pulsatile evidence over time. While keeping its head stationary, the rat listened to randomly timed clicks played from loudspeakers on its left (L) and right (R). At the end of the stimulus, the rat received a water reward for turning to the side with more clicks. The earliest time when a rat could respond was fixed at 1.5 s relative to the moment of inserting its nose in the centre port (that is, not a reaction time paradigm). b, Behavioural performance in an example recording session. Dashed reference lines at abscissa = 0 and ordinate = 0.5. c, The decision process is modelled as a dynamical system. Right, the blue and red arrows represent the change in the decision variable in the presence of a left or right, respectively. z₁, z dimension 1; z₂, z dimension 2. d, Autonomous dynamics illustrated using the bistable attractor hypothesis. In the velocity vector field (that is, flow field; left), the arrow at each value of the decision variable z indicates how the instantaneous change depends on z itself. The orientation of the arrow represents the direction of the change, and its size represents the speed, also quantified using a heat map (right). e, Changes in z driven solely by external sensory inputs (example of bistable attractors). f, Bistable attractor hypothesis of decision-making, with directions of the input dynamics (based on ref. ¹). g, A hypothesis supposing a line attractor in the autonomous dynamics on the basis of the DDM of decision behaviour (based on ref. ²³). h, Recurrent neural networks can be trained to make perceptual decisions using a line attractor that is not aligned to the input dynamics (non-normal; based on ref. ²). i, Unsupervised discovery (this study) of dynamics that have not been previously considered. j, Six interconnected frontal cortical and striatal regions are examined here. vStr, ventral striatum. k, Neuropixels recordings (318 ± 147 neurons per session per probe, mean ± s.d.) from 12 rats in total (two to three regions per rat). AP, anteroposterior; ML, mediolateral.