What changed

Amazon SageMaker AI has introduced Multi-Turn Reinforcement Learning (MTRL), a framework that trains models to learn from sequences of actions and their outcomes. This approach is designed for tasks that require multiple steps, such as filing an expense report or triaging a support ticket, where each action depends on the previous one.

What this means in practice

MTRL allows models to be trained using a reward-based system that evaluates the entire sequence of actions rather than individual steps. The system uses an environment, such as the Wordle game, to demonstrate how the model learns from feedback over multiple turns. The training process involves deploying an agent on Amazon Bedrock AgentCore or on-premises infrastructure, and the model is trained using a Low-Rank Adaptation (LoRA) adapter. The reward function is outcome-dominant, with higher rewards for successful completion and penalties for incorrect actions.

Limitations and availability

MTRL is available as a serverless service, where users select a base model, and the service provisions the training and sampling infrastructure. It supports models like Amazon Nova 2 Lite, GPT-OSS-20B, Gemma 4 31B, and Qwen 3.6 27B, all available in specific AWS regions. The pricing model is based on token usage across three dimensions: prefill, sample, and train. The solution requires specific IAM roles and an S3 bucket in the same region as the agent. The training process involves deploying the agent, preparing the dataset, training, monitoring in MLflow, and evaluating at serving settings. The code and datasets remain within the user's AWS account and are not used to train the base models.

Source facts for Amazon SageMaker AI Introduces Multi-Turn Reinforcement Learning for Complex Tasks: An agent that files an expense report or triages a support ticket takes a sequence of actions, and each step depends on what came before. There’s rarely one correct sequence to imitate, and the only reliable feedback is whether the whole sequence succeeded. Supervised fine-tuning has no labels for that, and single-turn reinforcement learning cannot credit a decision made three steps before the outcome.
Original AINewsTrend explainer based on the cited source. This graphic summarizes source statements; it is not an independent benchmark.

Sources and further reading

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