Hybrid Open-Ended Tri-Evolution Makes Better Deep Researcher

The Hybrid Open-Ended Tri-Evolution (HOTE) framework has been proposed to enhance the capabilities of AI agents in deep research tasks by integrating autonomous information retrieval and model evolution. Utilizing an 8B model, HOTE employs hybrid-mode reinforcement learning to evolve a proposer, solver, and judge collaboratively, achieving superior performance on long-form deep research benchmarks compared to static models ranging from 8B to 32B. This advancement is significant for practitioners as it addresses the limitations of static models in open-ended research environments, enabling more effective and efficient AI agent development.