From the source
Amazon SageMaker AI now supports multi-turn reinforcement learning (MTRL) for fine-tuning LLMs in agentic tasks.
- The service offers modular agent-environment interfaces, serverless execution, and a native algorithm library including PPO and CISPO.
- Amazon used MTRL to fine-tune a Qwen3.6-27B model for a search agent with BM25 and vector search tools, using nDCG@10 as the reward metric.





