Fine-tune a search agent with multi-turn RL on Amazon SageMaker AI
aws ·
This article details how AWS researchers fine-tuned a Qwen3.6-27B search agent using Amazon SageMaker AI’s multi-turn reinforcement learning (MTRL) capabilities. Unlike traditional supervised fine-tuning or single-turn RL, MTRL optimizes the model across full interaction sequences, allowing it to learn complex retrieval strategies autonomously. The team configured the agent with lexical and vector search tools, using nDCG@10 as the primary reward signal and penalizing failure modes like exceeding turn limits. Results demonstrated significant improvements in retrieval quality, with nDCG@10 gains of up to 23.7% on the BrowseComp-Plus benchmark. Additionally, the fine-tuned agent exhibited remarkable reliability, reducing its failure rate from nearly 23% to less than 1% on complex queries. The serverless nature of SageMaker AI MTRL allowed for efficient training without managing GPU clusters, while default hyperparameters proved sufficient for substantial performance gains. This approach offers enterprises a cost-effective way to achieve frontier-model-like reliability with smaller, faster models.
aws