AWS shows multi-turn RL fine-tuning for search agents on SageMaker AI
AWS published a tutorial on using Amazon SageMaker AI's multi-turn reinforcement learning (MTRL) to fine-tune a Qwen3.6-27B model into a search agent. The agent uses BM25 and vector search tools across multiple interaction turns to retrieve information. Training used nDCG@10 as a trajectory-level reward, with a -1 penalty for hitting turn or token limits.
Key points
- Fine-tuned Qwen3.6-27B with SageMaker AI MTRL for multi-turn search agent
- nDCG@10 improved up to 23.7% on BrowseComp-Plus, failure rate fell to 0.68%
- Serverless training with default PPO/CISPO algorithms, per-token pricing
The fine-tuned model improved nDCG@10 on three of four benchmarks: +23.7% on BrowseComp-Plus, +18.4% on WixQA, and +6% on Wands, with a slight regression on FreshStack. Failure rates dropped sharply, from 22.89% to 0.68% on BrowseComp-Plus. Training ran with maxepochs=1, globalbatchsize=128, and rolloutmaxconcurrency=32 on serverless infrastructure in the US West (Oregon) region.
The story so far
2 episodes →- AWS shows multi-turn RL fine-tuning for search agents on SageMaker AIthis story
Fine-tune a search agent with multi-turn RL on Amazon SageMaker AI
AWS Machine Learning Blog · 2 October 2026
Loading the full article…
This text was published by AWS Machine Learning Blog and written by Huibin Shen. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- OpenAI launches Dots, always-on AI agents in ChatGPT · 63 src
- Llama.cpp adds decision models for routing and scoring tasks · 1 src
- Anthropic introduces mods for Claude Code · 1 src
- Microsoft adds Copilot Home, Code, and Autopilot to automate work tasks · 2 src
- Earendil releases Pi 1.0 agent harness with default MCP support · 2 src
Comments
via GitHub Discussions