# AWS shows multi-turn RL fine-tuning for search agents

> AWS says fine-tuning Qwen3.6-27B with SageMaker AI improved results on three of four held-out search benchmarks and cut failures on BrowseComp-Plus.

Oossa · 2026-10-02 · https://oossa.com/en/aws-shows-multi-turn-rl-fine-tuning-for-search-agents

AWS published a guide to fine-tuning a search agent with multi-turn reinforcement learning on Amazon SageMaker AI. The method trains the agent across a whole sequence of searches, using the quality of its final results as the reward.

In AWS’s evaluation, the fine-tuned Qwen3.6-27B improved on three of four held-out benchmarks. On BrowseComp-Plus, its nDCG@10 score rose by 23.7% and its failure rate fell from 22.89% to 0.68%. AWS says the training used one epoch and a batch size of 128; the post does not give a complete set of results in the supplied text.

## The facts

- AWS reports a 23.7% nDCG@10 improvement on BrowseComp-Plus.
- BrowseComp-Plus failure rate fell from 22.89% to 0.68% in AWS’s evaluation.

## Why it matters

For teams building search agents on SageMaker AI, AWS’s results suggest multi-turn fine-tuning may reduce task failures, though the supplied post excerpt does not include full benchmark results or training costs.

## Sources & references

1. [Fine-tune a search agent with multi-turn RL on Amazon SageMaker AI](https://aws.amazon.com/blogs/machine-learning/fine-tune-a-search-agent-with-multi-turn-rl-on-amazon-sagemaker-ai/) – AWS, 2026-10-02

Last updated: 2026-10-02