SageMaker AI multi-turn RL lifts gpt-oss-20b solve rate to 84% for $339

The guide comes from a practical gap AWS describes: supervised fine-tuning needs labels for each step, and single-turn reinforcement learning cannot credit a decision made several steps before the outcome. MTRL instead trains on one reward reported when an episode ends, with the agent hosted on Amazon Bedrock AgentCore and the policy model living in the SageMaker AI training job. In the Wordle example, the dataset carries only the secret word, so the model never sees a target trajectory.

The reported numbers show why the evaluation setting matters. The training job's own validation curve peaked at 48 percent and ended at 37 percent at temperature 0, but at medium reasoning effort and temperature 1.0 the trained model solved 84 percent of held-out words against 61 percent for the base model. AWS notes the low-effort, temperature-0 pair corresponds to the training job's validation at steps 0 and 100, and that the adapter added 14–54 points across settings. The post also shares a reward-design revision: summing per-guess scores produced inversions where a lost game paid positive reward and a first-guess solve scored below a third-guess solve, which the final outcome-dominant formula corrects.