Amazon SageMaker AI MTRL: A Breakthrough in Fine-Tuning Search Agents with Multi-Turn Reinforcement Learning

Summary:AWS has introduced a novel approach using Amazon SageMaker AI's Multi-Turn Reinforcement Learning (MTRL) to fine-tune search agents, significantly enhancing their performance in retrieval quality, reliability, and efficiency. Fine-tuning the Qwen3.6-27B model with MTRL led to an average improvement of over 15% in nDCG@10 scores across multiple benchmarks, with a substantial reduction in failure rates and lower computational costs. This innovation provides enterprises with a more efficient and co

---

Core Breakthrough

Amazon SageMaker AI's Multi-Turn Reinforcement Learning (MTRL) technology marks a revolutionary advancement in optimizing search agents. By fine-tuning the Qwen3.6-27B model, this technology achieves the following core breakthroughs:

• Multi-Turn Interaction Optimization: MTRL enables holistic optimization of the agent's multi-turn interaction behavior, rather than just improving single-step decisions. This results in more reliable performance in complex tasks.
• Efficient Training Framework: MTRL offers a modular agent-environment interface, serverless execution environment, and asynchronous trajectory collection, significantly reducing the complexity of training.
• Significant Performance Improvement: In multiple benchmark tests, the nDCG@10 score improved by an average of over 15%, with a substantial reduction in failure rates. For example, in the BrowseComp-Plus test, the failure rate dropped from 22.89% to 0.68%.

Technical Highlights

1. Modular Design: The MTRL framework features a modular design that allows developers to customize reward functions, tool loops, and multi-turn conversation structures to suit different application scenarios.
2. Serverless Architecture: There is no need to manually manage GPU clusters, as the training process is based on a pay-as-you-go model, reducing computational costs.
3. Resumable Training: The training process supports checkpointing, allowing training to resume from the last checkpoint even after interruptions during long training sessions.
4. Comprehensive Observability: Through the MLflow platform of Amazon SageMaker AI, developers can inspect the behavior trajectory of the agent turn by turn, gaining a deeper understanding of its learning process.

Application Scenarios

• Enterprise Search: In corporate knowledge bases, the agent can retrieve and integrate information more efficiently, enhancing employee productivity.
• E-commerce: In product search, the agent can better understand user intent and provide more accurate recommendations.
• Customer Service: In customer service scenarios, the agent can handle more complex queries and provide higher quality service.

Developer Recommendations

1. Start with Default Configurations: The default configurations of MTRL perform well in most cases, and developers do not need a deep background in reinforcement learning to get started.
2. Prepare High-Quality Datasets: The training effect is highly dependent on the quality of the dataset, so it is recommended to use diverse datasets for training.
3. Define Clear Reward Functions: The reward function should be closely related to the task goal to ensure that the agent learns the correct behavior.
4. Continuous Evaluation and Iteration: Regularly evaluate the performance of the agent and iterate based on the evaluation results.

Industry Impact

This technology provides enterprises with a more efficient and cost-effective solution for complex information retrieval scenarios, promoting the further adoption of AI in enterprise applications. Additionally, the modular design and serverless architecture of MTRL also lower the barriers to AI development, providing developers with more powerful tools.

---

Source: AWS Machine Learning Blog (2026-10-02)

— END —

Tags: #Amazon SageMaker #Reinforcement Learning #Search Agents #Multi-Turn Interaction #Fine-Tuning