Rofly: Large-scale Models Enter the Era of Post-Training, Top Teams Achieve a 1:1 Pre-Training to Post-Training Computing Power Ratio

According to Dongzuo Beating monitoring, Luofuli, Head of Xiaomi's Large Model Team, pointed out that the era of large models has shifted from the pre-training-dominated Chat era to the post-train-dominated Agent era. The current key point is "how to scale reinforcement learning (RL) on an Agent."

This paradigm shift has directly led to a restructuring of compute allocation. Luofuli revealed that in the Chat era, the ratio of compute power for research, pre-training, and post-training was about 3:5:1; whereas in today's Agent era, the reasonable compute allocation ratio has changed to 3:1:1, meaning the investment in pre-training and post-training compute power is now roughly equal. Currently, top model teams have achieved a 1:1 investment ratio in these two areas.

At the same time, there have been significant changes in system architecture requirements. The previous RL infrastructure was mainly centered around a "model inference engine" that dealt with plain text algorithms; whereas the current infrastructure must be centered around an "Agent" to support heterogeneous cluster scheduling and tolerate the ambiguity of Agent interruptions in complex workflows due to various uncontrollable factors.