Applied Compute introduces production RL training recipe

The authors introduce production RL (prod RL), a training paradigm that creates a continuous feedback loop for models to learn from real-world inference deployments without a model harness or synthetic environment. The training recipe combines exact token recording, score centering to handle train-rollout mismatch, and sampled KL penalties to maintain run stability. Experiments demonstrate that score centering outperforms importance sampling while asynchronous execution improves data efficiency. Open article →