Hybrid Open-Ended Tri-Evolution Makes Better Deep Researcher
Researchers have developed a new framework called Hybrid Open-Ended Tri-Evolution (HOTE) to improve AI agents' capabilities in deep research and autonomous evolution. HOTE utilizes hybrid-mode reinforcement learning to foster collaborative evolution among proposer, solver, and judge modules, drawing upon web-scale knowledge. Experiments show that an 8B model trained with HOTE outperforms larger static models and existing deep research methods on long-form research tasks, with all three HOTE modules proving essential for its effectiveness. AI
IMPACT This framework could enable more autonomous and effective AI agents for complex, open-ended research tasks.