Frontier post-training recipe review with Finbarr Timbers

Post-training recipes for frontier AI models have evolved from simple three-stage supervised fine-tuning and RLHF to complex, industrial-scale pipelines. Current state-of-the-art approaches, exemplified by models like DeepSeek R1 and various multi-teacher distillation methods, prioritize reasoning-focused reinforcement learning and synthetic data generation over traditional human-labeled datasets. This shift reflects a move toward on-policy distillation, where models learn from multiple domain-specific expert teachers to refine reasoning behaviors. Building these systems requires significant organizational capacity to manage compute and data workflows, creating a divide between resource-heavy frontier labs and smaller academic efforts. While DPO remains a useful tool for smaller-scale efficiency, the industry is increasingly converging on complex, multi-teacher frameworks that integrate reasoning and tool-use capabilities to maintain performance at the frontier.