Behavioural tasks requiring context-dependent selection and accumulation of noisy sensory evidence are a powerful tool to study flexible decision-making (Mante et al., 2013), and recent work revealed a surprisingly degree of variability in the underlying neural mechanisms across individual animals (Pagan et al., 2022). Here we study how this individual variability develops during learning in rats and recurrent neural networks (RNNs). First, we show that the number of hidden units in RNNs plays a crucial role in the resulting amount of variability across networks initialized with different random parameters: 100-unit RNNs show stereotyped learning trajectories, while 20-unit RNNs match the variability observed in rats, in terms of both learning speed and diversity of learned solutions. Second, we discovered that almost all RNNs undergo the same two-phase learning process: in the first phase, networks learn to form a line attractor (accumulation of noisy evidence); in the second phase, networks learn to use context to differentially integrate relevant and irrelevant pulses of evidence (contextual feature selection). Remarkably, this two-phase process closely matches learning trajectories in rats, even though RNNs are exposed to the full task from the beginning of training. Finally, we show that it is possible to speed up learning in RNNs by breaking down the task into simpler stages, similar to the stepwise training procedure used in rats. For example, we find that for same random initial parameters, 20-unit RNNs are unable to learn the task because they cannot form a stable line attractor. Exposing these networks to a “pre-training” stage requiring only evidence accumulation (with no context information) greatly accelerates the formation of the line attractor, allowing networks to then learn the full task. In summary, our work lays the foundation for the study of individual variability in learning flexible tasks, and for the development of optimal training curricula.