The paper notes that frontier models, both general-purpose and domain-specific, now include a midtraining stage between pretraining and post-training. However, little has been published about how this stage is actually constructed. The authors aim to fill that gap.

They present an open midtraining recipe, presumably describing the data, methods, and steps needed to prime base models for agentic post-training. Because the abstract is brief, the specific details of the recipe are not yet available, but the contribution is positioned as an open resource for a stage that has largely remained opaque.