Youtu-Agent: Automated Agent Generation and Hybrid Policy Optimization

This paper introduces Youtu-Agent, a modular framework designed to address the challenges of LLM agent configuration and adaptability. It tackles the high costs of manual tool integration and prompt engineering by automating agent generation. Furthermore, it improves agent adaptability through a hybrid policy optimization system, including in-context optimization and reinforcement learning. The results demonstrate state-of-the-art performance and significant improvements in tool synthesis, performance on specific benchmarks, and training speed.

• •Youtu-Agent automates agent generation, reducing manual effort in tool integration and prompt engineering.
• •The framework uses a hybrid policy optimization system, including in-context optimization and reinforcement learning, to improve agent adaptability.
• •Experiments show state-of-the-art performance on WebWalkerQA and GAIA benchmarks.
• •The automated generation pipeline achieves a high tool synthesis success rate.
• •The Agent Practice module improves performance on AIME benchmarks.
• •Agent RL training achieves significant speedup and performance improvements on coding/reasoning and searching tasks.