Hybrid Open-Ended Tri-Evolution Improves AI for Deep Research Tasks

The Hybrid Open-Ended Tri-Evolution (HOTE) framework has been introduced to advance AI agents in deep research and autonomous evolution, crucial steps towards artificial general intelligence. Traditional deep research methods are often limited by static parametric capabilities, while existing agent evolution techniques are primarily validated on verifiable tasks with clear answers, leaving a gap for open-ended research. HOTE bridges these two areas by employing hybrid-mode reinforcement learning. This approach facilitates the collaborative evolution of three distinct modules: a proposer, a solver, and a judge. These modules learn and adapt based on web-scale knowledge, enabling agents to autonomously evolve in open-ended tasks and environments. Extensive experiments were conducted on three long-form deep research benchmarks. The results demonstrated that an 8B model trained using HOTE surpassed the performance of stronger static open models (8-32B parameters) and also outperformed models trained with other state-of-the-art deep research methods, all while requiring less training time. The study further confirmed that the evolution of all three modules within HOTE is essential for its superior performance.