Multi-Turn Reinforcement Learning on Amazon SageMaker AI: A Hands-On Guide
Multi-Turn Reinforcement Learning on Amazon SageMaker AI: A hands-on guide
- Who
- Daniel Quang and coauthors
- What
- publish a hands-on guide training models with multi-turn reinforcement learning on Amazon SageMaker AI
- Quote
In this post, we walk through the pipeline end-to-end: deploying the agent and reward to AgentCore, preparing the dataset, training, monitoring in MLflow, and evaluating at serving settings.
What happened
AWS provides an end-to-end guide to multi-turn reinforcement learning on Amazon SageMaker AI using a Wordle agent hosted on Amazon Bedrock AgentCore. Training improved gpt-oss-20b’s held-out solve rate from 61 to 84 percent at serving settings, costing about $339 in training tokens and $79 in runtime usage. The walkthrough covers deployment, reward design, monitoring, evaluation, and failures encountered during training.
Why it matters
The gap between training validation and serving results illustrates why multi-turn RL evaluation should match deployment settings, with decoding choices substantially changing the measured success rate.
Read next
See the full storyAWS Publishes Multi-Turn Reinforcement Learning Guide With Wordle Training ResultsThe story so far
Daniel Quang and coauthors published a hands-on guide to multi-turn reinforcement learning on Amazon SageMaker AI, reporting that their Wordle training example improved gpt-oss-20b’s held-out solve rate from 61% to 84% when evaluated at serving settings. The authors reported costs of about $339 in training tokens and $79 in runtime usage for the example. The walkthrough uses a Wordle agent hosted on Amazon Bedrock AgentCore and covers deploying the agent and reward, preparing the dataset, training, monitoring in MLflow, and evaluating the trained model. It also describes reward design and failures encountered during training, giving readers a worked example of the pipeline and its practical challenges.
Keep up with AI, one email a day
Model releases, price changes and compute developments, with original sources. Free, confirmed by email, and easy to unsubscribe. Read a past issue · 2026-10-23