Agentic RL · Game AI · World Model
Former StarCraft II pro (KR Grandmaster) turned AI researcher
Tip
Open to work — looking for research or engineering roles in Agentic RL, LLM/VLM agents and game AI. Reach me at sc2meisah@gmail.com.
正在求职:Agentic RL、LLM/VLM Agent 与游戏 AI 方向的研究 / 算法岗位,欢迎邮件联系。
My research asks how AI can perceive, decide and keep learning from interaction the way people do. I use games and multi-step agent tasks as testbeds for LLM/VLM agents, Agentic RL, world models and multi-agent decision-making.
| Work | What it is | Venue | Links |
|---|---|---|---|
| ⭐ LLM Play SC2 (TextStarCraft II) | The world's first work to put LLMs in charge of real-time decision-making in a complex environment — pioneering LLM research on complex real-time decision-making, with StarCraft II as the benchmark. 10+ LLMs benchmarked; >50% win rate vs. Harder (Lv5) AI | NeurIPS 2024 | Paper · Code |
| AVA | The world's first VLM-based multi-agent system for complex real-time decision-making — VLM micro from raw frames; CCF-A | ACL 2026 Findings | Paper · Code |
| EvoCurr | The world's first recursive self-improvement (RSI) in complex decision-making — 72 h unattended self-evolution; 100% win rate with 100+ units, beyond human Grandmasters | EMNLP 2026 Findings | Paper · Code |
| LLM-SMAC | The world's first Agentic RL in complex multi-agent decision-making (StarCraft II) — far deeper multi-turn interaction than Search-R1-style tool use; 7B model beats DeepSeek-V3 | AAMAS 2026 | Paper · Code |
| FreshPER | The world's first general replay-buffer framework for LLM post-training — NQ Search +20 pts, Sokoban +39 pts, MathVista +15 pts (absolute) | Under review | Paper · Code |
| JEV-Star | SOTA for direct LLM control of StarCraft II — 9/10 real-time wins vs. Lv7, the strongest non-cheating AI | Preprint 2026 | Paper · Code |
| eSports World Model Dataset | The world's first open-source, multi-game dataset for long-horizon, multi-agent world modeling in eSports — complete matches with synchronized per-player views (SC2 / Dota 2 / CS2) | Released 2026.10 | Dataset · Code |
More papers, projects and demos → weiyu.page
我关注一个核心问题:如何让 AI 像人一样感知、决策,并从交互中持续学习。我以游戏和多步智能体任务为试验场,研究 LLM/VLM Agent、Agentic RL、World Model 与多智能体决策。本科期间是星际争霸 II 职业选手(韩服宗师)。
代表作
- ⭐ LLM Play SC2(TextStarCraft II,NeurIPS 2024,引用 170+):全球首个将大模型用于复杂场景实时决策的工作,开创了以星际争霸 II 为标杆的大模型复杂实时决策方向;系统评测 10+ 主流大模型,对战困难(Lv5)AI 胜率 >50%
- AVA(ACL 2026 Findings):全球首个基于 VLM 的多智能体复杂实时决策系统,CCF-A
- EvoCurr(EMNLP 2026 Findings):全球首个在复杂决策场景实现递归自我改进(RSI)的工作,无人干预自主进化 72 小时,百余单位协同 100% 胜率、超越人类宗师
- LLM-SMAC(AAMAS 2026):全球首个在星际争霸 II 这类复杂多智能体决策场景下的 Agentic RL 工作,多轮交互深度远超 Search-R1 类工作,7B 模型超越 DeepSeek-V3
- FreshPER:全球首个通用的大模型后训练经验回放(Replay Buffer)框架,NQ Search、Sokoban、MathVista 分别提升 20、39、15 个百分点
- JEV-Star:大模型直接决策在星际争霸 II 上的 SOTA,实时对战最高非作弊难度 Lv7 AI 10 局 9 胜
- 电子竞技世界模型数据集:全球首个开源的多游戏、长时程、多智能体电竞世界模型数据集,以完整对局为单位提供逐玩家同步视角
完整论文、项目与演示视频见个人主页:weiyu.page