“Advances from fine-tuning general models for specialized domains like mathematics can create a synergistic loop, where the resulting data and reward models are used to improve the core general model.”