A new Google DeepMind paper demonstrates that training Gemini models with multi-turn reinforcement learning on simulated patients significantly improves psychiatric intake performance. The method reduced missed acute risk escalations from 30.5% to 12.9% and outperformed the base model in 76.7% of matched cases.
Multi-turn RL
多轮强化学习
An algorithm optimizing cumulative rewards for agents across multiple interaction steps in sequential decision processes.
Curated Intelligence
Amazon releases a hands-on guide for Multi-Turn Reinforcement Learning (MTRL) on SageMaker AI to address the limitations of supervised fine-tuning in agent tasks. The solution trains models using sparse reward signals from final task outcomes by connecting to Bedrock AgentCore or custom infrastructure, enabling effective learning for complex multi-step interactions.
Amazon SageMaker AI introduces a fine-tuning workflow using multi-turn reinforcement learning to enhance search agent reliability. By defining task-specific reward functions, developers can train small models to achieve frontier-level multi-step reasoning and tool-use capabilities while significantly reducing latency and cost.