Keep pulling the thread on Nathan Lambert.
By 2026, post-training recipes like that of MiMo Flash V2 fragmented into training many specialist models which are then merged into a single generalist model.
Multi-teacher On-Policy Distillation (MOPD) is an emerging pattern in frontier model post-training recipes in 2026.
The Nemotron 3 Ultra recipe uses two iterations of multi-teacher on-policy distillation with more than 10 teachers covering reasoning, code, math, and agentic domains.
The field of LLM post-training is currently converging on the technique of multi-teacher on-policy distillation.
NVIDIA's research found that teacher models trained with substantially different pipelines cannot be effectively combined through on-policy distillation, leading to suboptimal performance.
NVIDIA hypothesizes that a mismatch in SFT data between teacher and student models causes different reasoning behaviors and output distributions, reducing the quality of supervision signals during on-policy distillation.
The structure of post-training recipes for large language models changed more in the last year than in the preceding three years.
In 2024, open-source recipes like Llama 3 and Tülu 3 formalized a pipeline of Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Learning with verifiable rewards (RLVR).
The DeepSeek R1 model established a recipe where large-scale, reasoning-focused Reinforcement Learning (RL) became the central component of post-training.
The Multi-teacher On-Policy Distillation (MOPD) technique was introduced by MiMo Flash v2 and later scaled to over 10 teacher models by DeepSeek V4 and Nemotron 3 Ultra.
The emergence of Multi-teacher On-Policy Distillation (MOPD) is partly because mixing math, code, and agentic RL in a single training run leads to capability trade-offs.
The use of specialist models in post-training is organizationally scalable because training a specialist with SFT and RL on a single domain is a well-understood and parallelizable task.