Source time Edition date
ToolsAWS Machine Learning Blog
Multi-Turn Reinforcement Learning on Amazon SageMaker AI: A hands-on guide
Worth reading to see how Amazon SageMaker AI handles multi-turn RL: training on whole-task outcomes instead of labeled steps shifts the reward-design burden onto the builder, as the Wordle walkthrough shows.
Judgment
Teams choosing between supervised fine-tuning and multi-turn RL on SageMaker AI must budget for reward design and whole-task evaluation rather than step labels; adopt only if they can score episode-level success reliably.
Basis
AWS Machine Learning Blog authors Daniel Quang, Hadrien Almela, and Vinay Bhat (23 OCT 2026) frame the split: supervised fine-tuning teaches a better answer to a single prompt, but agentic tasks like filing an expense report or triaging a ticket take a sequence of actions where each step depends on what came before, and there is rarely one correct sequence to imitate. SFT has no labels for whole-sequence success, and single-turn RL cannot credit a decision made three steps before the outcome. The mechanism they describe: SageMaker AI sends prompts to an agent you host on Amazon Bedrock AgentCore or your own infrastructure (Amazon EKS, Amazon EC2, or AWS Fargate) through an AWS Lambda forwarder; the agent calls the policy model being trained, acts in your environment over as many turns as the task needs, and reports one reward when the episode ends. That moves the burden onto the builder: the training signal is now a reward you define and an episode outcome you must judge, so the gating cost shifts from labeling steps to designing and validating a reward that reflects task success.
Unclear The source does not quantify reward-design effort, sample counts, training cost, or any success-rate comparison between MTRL and supervised fine-tuning.
Next Check whether the published post reports concrete Wordle episode success rates or reward curves for the MTRL run versus a baseline, and whether it names the reward function used.