AWS Announces Multi-Turn Reinforcement Learning on Amazon SageMaker AI
According to AWS. New serverless capability enables training models for complex, multi-step agentic workflows using reinforcement learning.
- • Enables training for models requiring sequential decision-making.
- • Serverless architecture removes the need for manual cluster sizing.
- • Supports LoRA adapter training for specific business workflows.
- • Ensures user data remains within the customer's AWS account.
Addressing Multi-Step Agentic Tasks
AWS has introduced Multi-Turn Reinforcement Learning (MTRL) on Amazon SageMaker AI to address the limitations of supervised fine-tuning in agentic workflows. By allowing models to receive feedback only after a full sequence of actions is completed, MTRL enables agents to learn complex tasks such as triaging support tickets or managing internal business processes.
Technical Implementation and Availability
The MTRL service is serverless, allowing users to select a base model without managing underlying infrastructure. It currently supports models including Amazon Nova 2 Lite, GPT-OSS-20B, Gemma 4 31B, and Qwen 3.6 27B. The system utilizes Group Relative Policy Optimization (GRPO) to update LoRA adapters, ensuring that training remains efficient and specific to the user's environment.
- 1. aws.amazon.comOpen source