北京大学前沿计算研究中心

本科生科研助理 · 导师：董豪

参与具身模型的 post-training 与评测。手上的工作比较具体：多源机器人数据清洗和对齐、SFT / CoT 样本构造、RL 候选筛选、评测脚本维护，以及错误案例回流。

这段工作让我真正意识到，模型“失败”本身并不自动成为有用数据。只有把输入、轨迹、判分和目标都记录清楚，失败才有可能进入下一轮训练。

Center for Frontier Computing Research, PKU

Undergraduate Research Assistant · Advisor: Hao Dong

I work on post-training and evaluation for embodied models: cleaning and aligning robot data, building SFT and CoT samples, filtering RL candidates, maintaining evaluation scripts, and feeding failure cases into the next round.

The work made one thing clear: a model failure does not automatically become useful data. Inputs, trajectories, scoring, and targets all need to be recorded well enough for the failure to teach the next run anything.

北京大学王选计算机研究所

拔尖班科研轮转 · 导师：刘洋

做过 multi-agent debating 框架下的视频时空定位实验，主要参与定位模块调试、模型评测、结果记录和 bad case 分析。

最有用的收获不是某个模型名字，而是学会把 temporal failure 和 spatial failure 分开看。两个问题混在一个总分里，很多错误其实解释不清。

Wangxuan Institute of Computer Technology, PKU

Research Rotation · Advisor: Yang Liu

I worked on video spatio-temporal grounding in a multi-agent debating setup, helping debug localization modules, evaluate models, log results, and inspect bad cases.

The lasting lesson was simple: temporal and spatial failures need to be inspected separately. Once they are collapsed into one score, many errors become hard to explain.

北京大学

计算机科学与技术 · 拔尖班

本科在读。课程之外，我更愿意用项目把一个概念做实：写过 C++ Mini-Lisp 解释器，也做过叙事 Agent、多人世界模拟和本地 LLM 德州扑克桌。

Peking University

B.S. in Computer Science and Technology · Elite Program

Undergraduate study. Outside coursework, I tend to make concepts concrete through projects: a C++ Mini-Lisp interpreter, narrative agents, a multi-agent world simulation, and a local LLM poker table.

PLIB / PKUHub

学生 AI 社群与校内产品

参与学生 AI 分享和产品工作。PKUHub 是面向校内的笔记共享平台，注册用户超过 5,000。它让我第一次直接面对“有人真的每天会用”的软件，而不只是自己能跑通的 demo。

PLIB / PKUHub

Student AI community and campus product

I helped with student AI talks and product work. PKUHub is a campus note-sharing platform with more than 5,000 registered users. It was my first experience with software people actually returned to, not merely a demo that ran once.