Multi-Turn Reinforcement Learning on Amazon SageMaker AI: A hands-on guide
Multi-Turn Reinforcement Learning (MTRL) on Amazon SageMaker AI trains models on the outcome of a whole multi-step task rather than labeled steps. This hands-on guide walks the full pipeline using Wordle: deploying the agent and reward on Amazon Bedrock AgentCore, preparing the dataset, training, and evaluating at serving settings.
First seen by Oriel on .
AI summary
The guide demonstrates training an agent for multi-step tasks with Multi-Turn Reinforcement Learning on Amazon SageMaker AI, reporting improved Wordle solve rates using end-of-episode rewards.
- An AgentCore-hosted Strands agent plays up to six Wordle turns, calls the policy model through SageMaker Job Runtime, and reports one reward when each game ends.
- GRPO compares rewards across rollouts of the same word to update a LoRA adapter; training rows contain secret words but no target responses or trajectory labels.
- On held-out words, gpt-oss-20b improved from 61% to 84% solve rate at default reasoning effort and from 15% to 69% at low effort.
- The 100-step run used 105,486 rollouts over 12 hours 40 minutes, costing about $339 in training tokens plus about $79 for AgentCore runtime.
- The walkthrough covers IAM roles, agent deployment, dataset preparation, training and MLflow monitoring; MTRL supports four listed base models in specified AWS Regions.
#reinforcement learning #amazon sagemaker #amazon bedrock agentcore #llm agents
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| AWS Machine Learning Blog | d4a0021627e124e1cbf32bd2a1426e1abfda6b33 | https://aws.amazon.com/blogs/machine-learning/multi-turn-reinforcement-learning-on-amazon-sagemaker-ai-a-hands-on-guide/ Read discussion |