Amazon Launches Search Agent Fine-Tuning with Multi-Turn RL on SageMaker AI

It is a training approach that optimizes an agent's performance across a sequence of interacting steps or turns, rather than scoring a single answer, which fits search tasks that require sequential decisions.

Traditional fine-tuning teaches a model input-output examples, while multi-turn RL teaches the agent how to act inside your environment and tools through rewards on outcomes, improving reliability in search tasks.

Yes, because the agent is tuned to your tools and data sources, reducing latency and cost compared with relying on a general frontier model, which matters for enterprises and government entities handling sensitive data.

Amazon reported measured gains in retrieval quality and reliability, but the published summary did not include detailed figures, so check the original link.