Fine-tune Search Agents Using Amazon SageMaker MTRL

Amazon SageMaker AI now supports multi-turn reinforcement learning (MTRL) for fine-tuning LLM-powered search agents. Unlike supervised fine-tuning or single-turn RL, MTRL optimizes agent behavior across full multi-turn trajectories, using a reward signal tied to final retrieval quality. The approach trains smaller models to match frontier model reliability at lower cost and latency. SageMaker MTRL offers modular agent-environment interfaces, serverless execution, asynchronous rollout collection, and built-in algorithm support including PPO and GRPO. The post demonstrates fine-tuning a Qwen3-27B model with BM25 and vector search tools in an enterprise search setting.