AWS SageMaker AI adds multi-turn reinforcement learning for agentic tasks
A new SageMaker AI capability trains models on whole-task outcomes rather than step-by-step labels, using Wordle as a teaching example and targeting internal workflows.
1 source · cross-referenced
- SageMaker AI now supports multi-turn reinforcement learning (MTRL) to train models on the outcome of entire multi-step tasks rather than labeled steps.
- The approach uses Amazon Bedrock AgentCore to host the agent and reward function, with training run in SageMaker AI.
- In a Wordle example, MTRL improved gpt-oss-20b’s solve rate from 61% to 84% on held-out words at default reasoning effort.
- Training cost about $339 in training tokens; datasets, agent, and model remain in the customer’s AWS account.
- Code and end-to-end pipeline are provided in an aws-samples repository.
Amazon SageMaker AI now supports multi-turn reinforcement learning (MTRL), a capability that trains models on the outcome of a complete multi-step task rather than on labeled steps. This contrasts with supervised fine-tuning, which targets single-turn responses, and single-turn reinforcement learning, which cannot credit decisions made early in a sequence based on a delayed reward.
The MTRL pipeline in SageMaker AI sends prompts to an agent hosted on Amazon Bedrock AgentCore or on customer-managed infrastructure such as Amazon Elastic Kubernetes Service, Amazon EC2, or AWS Fargate via an AWS Lambda forwarder. The agent calls the policy model being trained, executes actions in the environment over as many turns as needed, and returns a single reward at the end of the episode. Only the final outcome is scored; the training process does not require predefined trajectories.
The example in the announcement uses Wordle, a game where a player has six attempts to guess a five-letter word. Each guess receives feedback—green for correct position, yellow for correct letter in the wrong position, and gray for absent letters—making the game suitable for MTRL because the rules are simple, each action depends on prior feedback, and the outcome is objectively scored. While frontier models already perform well at Wordle, the authors emphasize that MTRL is intended for proprietary, multi-step business workflows that public models were not trained to handle.
In the Wordle demonstration, MTRL improved the solve rate of the gpt-oss-20b model from 61% to 84% on held-out words at the model’s default reasoning effort. At the low reasoning effort level used during training, the solve rate increased from 15% to 69%. The training run consumed approximately $339 in training tokens. The authors note that the datasets, agent, and trained model remain within the customer’s AWS account and are not used to train the base models.
The post provides an end-to-end walkthrough of the MTRL pipeline, including deploying the agent and reward to AgentCore, preparing the dataset, training with monitoring in MLflow, and evaluating at serving settings. The authors also share lessons learned from reward-design mistakes and training failures encountered during development. All code is available in the accompanying aws-samples repository.
- Nov 11, 2026 · Hugging Face
Hugging Face proposes token pooling-aware training to improve compressed multi-vector retrieval
Trust79 - Nov 11, 2026 · TechCrunch — AI
Fractile raises $720M Series C to develop AI chips and compiler stack
Trust79 - Nov 11, 2026 · The Verge — AI
Google’s Tabletop AI experiment lets users build custom collaborative apps via waitlist
Trust79