Animals must exploit environmental regularities to make adaptive decisions, yet the learning algorithms that enabels this flexibility remain unclear. A central question across neuroscience, cognitive science, and machine learning, is whether learning relies on generative or discriminative strategies. Generative learners build internal models the sensory world itself, capturing its statistical structure; discriminative learners map stimuli directly onto choices, ignoring input statistics. These strategies rely on fundamentally different internal representations and entail distinct computational trade-offs: generative learning supports flexible generalisation and transfer, whereas discriminative learning is efficient but task-specific. We compared humans, rats, and mice performing the same auditory categorisation task, where category boundaries and rewards were fixed but sensory statistics varied. All species adapted their behaviour near-optimally, consistent with a normative observer constrained by sensory and decision noise. Yet their underlying algorithms diverged: humans predominantly relied on generative representations, mice on discriminative boundary-tracking, and rats spanned both regimes. Crucially, end-point performance concealed these differences, only learning trajectories and trial-to-trial updates revealed the divergence. These results show that similar near-optimal behaviour can mask fundamentally different internal representations, establishing a comparative framework for uncovering the hidden strategies that support statistical learning. ### Competing Interest Statement The authors have declared no competing interest. Wellcome Trust, https://ror.org/029chgv08, 219880/Z/19/Z, 225438/Z/22/Z, 219627/Z/19/Z Gatsby Charitable Foundation, GAT3755 UK Research and Innovation, https://ror.org/001aqnf71, EP/Z000599/1