Han Zheng

Introduced METIS, a self-evolving framework that internalizes curriculum judgment as a native LLM capability for RFT, letting the policy learn what to learn next and accelerating convergence by up to 67%.

A decentralized multi-agent system inspired by Hayek's economics, where agents compete via auctions and evolve through economic selection. Simple market signals turn populations of weak partial agents into collective intelligence that outperforms stronger monolithic baselines across five agentic tasks.

First framework for autonomous multi-agent evolution on open-ended problems. CORAL uses long-running LLM agents that explore, reflect, and collaborate through shared persistent memory and asynchronous execution, achieving 3–10× higher improvement rates over fixed evolutionary baselines across mathematical, algorithmic, and systems optimization tasks.

Developed RL-RH-PP, combining deep reinforcement learning with search-based prioritized planning for lifelong MAPF. Demonstrated superior throughput and strong generalization across agent densities and planning horizons.

Self-improving LLM framework that builds an experience library for formulating optimization programs across diverse problem domains.

Integrated spatial optimization with LLMs for open-domain urban itinerary planning, generating personalized routes from natural-language user requests.

Pioneered BK-PBS, integrating offline-trained behavior prediction with multi-agent pathfinding to coordinate connected and human-driven vehicles under realistic kinematic constraints on highways.

Proposed OBS-KATS for signal-free intersection coordination. Proved soundness, completeness, and optimality of the crossing order search with significant throughput improvements.