An Evolutionary Epistemology of Post-Training

The capacity of frontier models to produce new knowledge is actively contested, and recent results, including GPT-5.4 Pro's solutions to several Erdős Problems, have brought greater attention to the debate. Past work has tended to grant novelty only in a thin sense, attributing model outputs either to recombination of latent training-data elements or to reversion to modal outcomes within the training distribution. We argue that Campbell's evolutionary epistemology, and in particular Blind Variation Selective Retention (BV-SR), supports the stronger claim that post-training techniques such as Reinforcement Learning from Verified Rewards (RLVR) instantiate a genuine selectionist process, and that the resulting knowledge is properly attributed to the system rather than redistributed back across the authors of the pre-training corpus. We further consider where this framing predicts current approaches will succeed, where they will stall, and what it implies for extending these methods beyond mathematics into the wider sciences.