北京大学 · 计算机科学与技术 · 2024—2028
目前主要做 RL for LLM / VLA Systems，也关注 Coding Agent、AI Self-Training 和 AI Infrastructure。
Peking University · Computer Science · 2024—2028
I work mainly on RL for LLM and VLA systems, with broader interests in coding agents, AI self-training, and AI infrastructure.
Beijing, 2026
SELECTED PROJECTS / 项目 SELECTED PROJECTS
01 · RL POST-TRAINING
一套完整的 RL post-training 流程。
面向 scientific coding tasks，覆盖数据生成、reward 设计、GRPO 训练、checkpoint evaluation 和 official-style evaluation。
A complete RL post-training pipeline.
Built for scientific coding tasks, covering data generation, reward design, GRPO training, checkpoint evaluation, and official-style evaluation.
02 · MULTI-AGENT GAME
一个完整、可本地运行的多人 LLM 德州扑克系统。
支持真人与多个 LLM 同桌对局，包含独立规则引擎、合法动作校验、牌局日志和复盘功能；也可以用来研究 LLM 在不完美信息博弈中的行为。
A complete, locally runnable multiplayer LLM poker system.
It supports one human and several LLM players, with a dedicated rules engine, legal-action validation, hand histories, and post-game review. It can also serve as a testbed for LLM behavior in imperfect-information games.
03 · IOS PRODUCT
一个可运行的完整 MVP，正在准备上架。
面向具体关系和具体对话的 iOS 沟通辅助工具。用户可以输入、口述或从聊天截图中提取消息，App 会结合联系人关系、个人表达习惯和可编辑记忆，给出三种回复建议；AI 只提供建议，不读取、也不代发消息。
A complete, working MVP currently being prepared for release.
An iOS communication assistant built around a specific relationship and conversation. Users can type, dictate, or extract a message from a screenshot; the app uses contact context, personal preferences, and editable memories to generate three reply suggestions. It never reads or sends messages on the user's behalf.
04 · COURSE PROJECT
一门课的大作业，做完觉得还挺酷。
北京大学《软件设计实践》课程项目：使用 C++ 实现的 Mini-Lisp 解释器，支持 tokenizer / parser、词法作用域、闭包、宏、REPL 和脚本执行。
A course final project that turned out pretty cool.
Built for PKU's Software Design Practice course: a Mini-Lisp interpreter in C++ with a tokenizer and parser, lexical scoping, closures, macros, a REPL, and script execution.
05 · CAMPUS RPG
一个完整可通关的七日校园生活微 RPG。
玩家在有限时段里选择去哪里、见谁以及是否兑现承诺；关系、记忆、成长和结局由游戏规则结算。无需 API Key 也能完整游玩，接入模型后会扩展自由对话。
A complete seven-day campus-life micro-RPG.
The player chooses where to go, whom to meet, and which promises to keep. Game rules own relationships, memories, progression, and endings. It is fully playable without an API key; an optional model expands free-form dialogue.