# AWS reports fewer search-agent failures after multi-turn training

_Published Friday, October 2, 2026 at 12:01 PM EDT · AI · Latest · Tier 2 — Notable_

![AWS reports fewer search-agent failures after multi-turn training — Primary](https://d2908q01vomqb2.cloudfront.net/f1f836cb4ea6efb2a0b1b99f41ad8b103eff4b59/2026/10/01/ML-21889-featured-image.png)

AWS reports that reinforcement learning across complete search sequences improved a Qwen3.6-27B agent on three of four held-out benchmarks. On BrowseComp-Plus, its failure rate fell from 22.89 percent to 0.68 percent, while its document-ranking score improved by 23.7 percent.

The SageMaker AI training rewarded the relevance of the final retrieved documents and penalized agents that exceeded turn or token limits. The agent used keyword and vector search tools. Results improved on WixQA and Wands but regressed slightly on FreshStack.

## Sources

- [Artificial Intelligence](https://aws.amazon.com/blogs/machine-learning/fine-tune-a-search-agent-with-multi-turn-rl-on-amazon-sagemaker-ai/)

---
Canonical: https://techandbusiness.org/newswire/XDeSEOpMyMqmQ5xQY7yr6I
Published: 2026-10-02T16:01:30.730Z
Story chronology: 2026-10-02T15:44:20.000Z
Retrieved: 2026-10-02T18:44:30.908Z
Publisher: Tech & Business (techandbusiness.org)