Multi-Turn Reinforcement Learning on Amazon SageMaker AI: A hands-on guide

Amazon SageMaker AI now supports Multi-Turn Reinforcement Learning (MTRL) to train models on sequential decision tasks using feedback only at the end of episodes. Using Wordle as an example, MTRL improved the gpt-oss-20b model's solve rate significantly with a cost of about $339 in training tokens. The process integrates with Amazon Bedrock AgentCore and AWS infrastructure, enabling custom multi-step task training within users' AWS accounts.