Amazon SageMaker AI supports an approach known as Multi-Turn Reinforcement Learning, or MTRL. This training technique trains models on the outcome of an entire multi-step task instead of relying on individually labeled steps. AWS outlines the implementation of this methodology in a hands-on guide that uses the game Wordle as its practical example.
The walkthrough explains the full end-to-end pipeline required to set up the system. It covers deploying the agent and reward components on Amazon Bedrock AgentCore while preparing the necessary dataset. In addition, the guide shows practitioners how to handle training and evaluate the model under serving settings.