MTRL on Amazon SageMaker AI trains models for multi-turn agentic tasks
Current overview (not historical)
Amazon SageMaker AI has launched a Multi-Turn Reinforcement Learning (MTRL) service designed for training policy models on agentic tasks. In a 100-step Wordle training run utilizing GPT-OSS-20B, MTRL required 491.6 million prefill, 159.7 million sample, and 643.3 million train tokens across 105,486 rollouts, completed in 12 hours and 40 minutes at a cost of approximately $339. The resulting trained model demonstrated significant improvement, increasing the solve rate from 61 percent to 84 percent on held-out words at default reasoning effort, and from 15 percent to 69 percent at low effort.
Written with AI from the reporting · updated 10/23/2026
Expand the detailed fact route and sources
Story route
A station-by-station route from the first public evidence onward; every station carries its own evidence and uncertainty.
- 1Unclassified fact stationOct 24, 00:00recorded Oct 24, 02:15single source
MTRL on Amazon SageMaker AI trains models on multi-turn agen
Read original source ↗ AWS Machine Learning BlogReplay this moment in the explorerEvidence & revisions · 1 source records
- Subject：
- Amazon SageMaker AI
- Action：
- launches Multi-Turn Reinforcement Learning (MTRL) service
- Object：
- policy model training for agentic tasks
- Occurred (fact side)：
- Oct 24, 00:00
- Published (valid side)：
- Oct 24, 01:48
- First seen (known side)：
- Oct 24, 02:15
Fact stationSingle source · one-sided claimCopy may have been revised later; time replay is bounded. Layer labels group reading only, not causation between layers.
Uncertainty marks flag where to check the sources; they are not proven causation.
Report timeline
Explore every side through dated source reports.
- AWS Machine Learning BlogFirst-party sourceRelated fact: single sourceAWS demonstrates Multi-Turn Reinforcement Learning on Amazon SageMaker AI using Wordle
Report first recorded Oct 24, 02:15
Multi-Turn Reinforcement Learning (MTRL) trains models on the success or failure of an entire sequence of actions, which is necessary for agentic tasks like filing expense reports, unlike single-turn fine-tuning. Using Wordle as an example, the system connects SageMaker AI for training the policy model with Amazon Bedrock AgentCore for hosting the agent and reward function. MTRL improved gpt-oss-20b's solve rate from 61% to 84% at default reasoning effort, costing about $339 in training tokens.
Read original source ↗
Event discussion
Reader opinions are hosted on GitHub Discussions; posting requires a GitHub account. Comments are not verified facts.