Experience 经历
This page is a web-native overview of my education, research, selected projects, and toolbox. It is intentionally more narrative than a full CV: the goal is to show what I worked on, what my role was, and what evidence is public.
Education
- Peking University, B.S. in Computer Science and Technology (Elite Program), 2024.09 - 2028.07
Research Experience
Center for Frontier Computing Research, Peking University
Advisor: Hao Dong
Role: Undergraduate research assistant
Time: 2025.12 - Present
Problem: Multimodal embodied models often fail in ways that are hard to reuse. A useful training loop needs to turn those failures into cleaned data, training samples, reward signals, evaluation scripts, and the next round of diagnosis.
My role: I participate in the post-training and evaluation loop: cleaning and aligning multi-source robot datasets, constructing SFT/CoT samples, filtering RL candidates, maintaining evaluation scripts, and analyzing error cases.
Methods: SFT/CoT sample construction, GRPO/RL candidate filtering, benchmark-style evaluation, error-case feedback, and real-world navigation evaluation.
Artifacts / evidence: Internal experiment records and evaluation scripts; public-facing related evidence is CritPT-RL, where I made a smaller open post-training/eval loop to study reward/eval mismatch.
Status / lesson: Ongoing. The strongest lesson so far is that failure cases are only useful when their data path and evaluation target are explicit enough to reproduce.
Wangxuan Institute of Computer Technology, Peking University
Advisor: Yang Liu
Role: Research rotation
Problem: Video spatio-temporal grounding needs robust localization across language descriptions, video clips, temporal intervals, and candidate spatial regions.
My role: I worked on experiments around a multi-agent debating framework for video grounding, including localization module debugging, model evaluation, result logging, and error-case analysis.
Methods: Temporal grounding, spatial grounding, multi-agent debating, evaluation workflow construction, and localization result inspection.
Artifacts / evidence: Research rotation work; public details are limited.
Status / lesson: Completed rotation. The useful lesson was that grounding errors are easier to study when temporal and spatial failures are logged separately.
Selected Projects
- CritPT-RL: a public Qwen-style GRPO post-training and evaluation lab for scientific Python-answer tasks. The key result is a negative finding: better format and cleaner
answer()structure did not automatically improve official70 accuracy. - Texas-Poker-Agents: a public local Hold’em environment for one human and multiple LLM seats, with strict visible-state prompts, rules-engine validation, logs, fallback actions, and replay.
- PKUHub / PLIB: campus AI community and product work, including student AI talks and PKUHub, a note-sharing platform with 5,000+ registered users.
Toolbox
- Post-training / eval: SFT, GRPO/RL, CoT sample construction, reward design, failure mining, official-style evaluation, LLM-as-a-judge.
- Data pipelines: robot data cleaning, filtering, synthesis, annotation alignment, quality checks, evaluation-set construction, automated processing.
- Systems / product: Python, JavaScript, C++, Flask, Django, FastAPI, SQL/ORM, Git, browser UIs, JSON/JSONL logs.
- Interests: embodied AI, multimodal agents, post-training, benchmark transfer, inspectable agent systems.
这个页面是网页版经历概览，覆盖教育背景、科研经历、代表项目和工具箱。它不是完整 CV 的逐条搬运，而是更强调：我做过什么、我的角色是什么、哪些证据可以公开看到。
教育背景
- 北京大学，计算机科学与技术（拔尖班），本科，2024.09 - 2028.07
科研经历
北京大学前沿计算研究中心
导师：董豪
角色：本科生科研助理
时间：2025.12 - 至今
Problem：多模态具身模型在真实任务中失败后，需要把失败案例转化为清洗后的数据、训练样本、reward 信号、评测脚本和下一轮诊断依据。
My role：我参与 post-training 与评测闭环，包括多源机器人数据清洗与对齐、SFT/CoT 样本构造、RL 候选样本筛选、评测脚本维护和错误案例分析。
Methods：SFT/CoT 样本构造、GRPO/RL 候选筛选、benchmark-style evaluation、错误案例回流和真实环境导航评测。
Artifacts / evidence：内部实验记录和评测脚本；公开可看的相关证据是 CritPT-RL，我在其中做了一个更小的开放后训练/评测闭环，用来研究 reward/eval mismatch。
Status / lesson：进行中。到目前为止最重要的体会是：失败案例只有在数据路径和评测目标都足够明确时，才真的能被复现和利用。
北京大学王选计算机研究所
导师：刘洋
角色：拔尖班科研轮转
Problem：视频时空定位需要同时处理语言描述、视频片段、时间区间和候选空间区域，鲁棒性很容易被细粒度定位误差影响。
My role：我参与 multi-agent debating 框架下的视频 grounding 实验，包括定位模块调试、模型评测、结果记录和错误案例分析。
Methods：temporal grounding、spatial grounding、multi-agent debating、评测流程整理和定位结果检查。
Artifacts / evidence：科研轮转工作；公开细节有限。
Status / lesson：轮转已完成。比较有用的体会是：如果 temporal failure 和 spatial failure 能分开记录，grounding 错误会更容易分析。
代表项目
- CritPT-RL：公开的 Qwen-style GRPO 后训练与评测实验，面向 scientific Python-answer 任务。关键结论是一个负结果：格式更干净、
answer()结构更像样，并不自动提升 official70 accuracy。 - Texas-Poker-Agents：公开的本地德州扑克实验环境，真人对战多个 LLM 座位，包含严格可见状态 prompt、规则引擎校验、日志、fallback 行为和复盘。
- PKUHub / PLIB：校内 AI 社群与产品工作，包括学生 AI 讲座，以及注册用户超过 5,000 的校内笔记共享平台 PKUHub。
工具箱
- Post-training / eval：SFT、GRPO/RL、CoT 样本构造、reward 设计、failure mining、official-style evaluation、LLM-as-a-judge。
- 数据管线：机器人数据清洗、筛选、合成、标注对齐、质量检查、评测集构建和自动化处理。
- 系统 / 产品：Python、JavaScript、C++、Flask、Django、FastAPI、SQL/ORM、Git、浏览器 UI、JSON/JSONL 日志。
- 兴趣方向：具身智能、多模态 agent、后训练、benchmark transfer、可检查的 agent 系统。