Ryder Lu路梦晨
Photo of Mengchen (Ryder) Lu

Mengchen (Ryder) Lu路梦晨路梦晨Ryder Lu

PhD Researcher in Computer Science · University of Exeter, UK 计算机科学博士研究生 · 英国埃克塞特大学

01About简介

I am a PhD researcher in Computer Science at the University of Exeter, supervised by Dr Shiqiang Wang and Dr George De Ath. I started in 2026 on the MPhil route, which transfers to the PhD after the progression review.

Before Exeter I received my master's in Communication Engineering (IoT) from Nanyang Technological University (NTU), Singapore, and earned bachelor's degrees in Communication Engineering from Beijing Jiaotong University and in Electronics and Communications Engineering from Lancaster University. I also worked as an engineering intern on wireless network planning and optimisation at China Mobile and Beijing ZX-CE Technology.

我是英国埃克塞特大学计算机科学专业的博士研究生,导师是 Shiqiang Wang 博士和 George De Ath 博士。我于 2026 年入学,目前在 MPhil 阶段, 通过阶段考核后转为 PhD。

此前,我在新加坡南洋理工大学(NTU)获得通信工程(物联网方向)硕士学位;本科就读于北京交通大学,同时获得北京交通大学通信工程学士学位和英国兰卡斯特大学电子与通信工程学士学位。我还曾在中国移动和北京 ZX-CE 科技有限公司担任电子信息工程实习生,从事无线网络规划与优化工作。

02Research研究方向

I study LLM-based agentic AI, and in particular the harness around a model: the prompts, tools, control flow and context management that turn a frozen LLM into an agent. The same model can perform very differently depending on its harness, and I want to understand and optimise that layer.

我研究基于大语言模型的智能体(agentic AI),尤其关注模型外面的那一层 harness: 也就是把一个固定的大模型变成智能体所需要的 prompt、工具、控制流程和上下文管理。同一个模型配上不同的 harness, 表现可能差别很大,我希望理解并优化这一层。

Harness designHarness 设计

Agent harness optimisation智能体 harness 优化

Which tools, prompts and feedback loops actually make an agent better, and how to improve them automatically. 哪些工具、prompt 和反馈循环真正让智能体变得更好,以及如何自动地改进它们。

Efficiency效率

Accuracy vs. cost准确率与成本的权衡

Agents can spend many times more tokens than a single model call. I look at where that cost goes and when it pays off. 智能体消耗的 token 可能是单次调用的许多倍。我关注这些成本花在哪里,以及什么时候值得。

Evaluation评测

Reliable agent evaluation可靠的智能体评测

Controlled comparisons on coding benchmarks with locally served open-source models. 在本地部署的开源模型上,用编程基准做可控的对比实验。

Currently: comparing a minimal and a full coding-agent harness with local open-source models, and studying where the extra tokens of multi-turn agents go. 目前在做:用本地开源模型对比最小 harness 与完整 coding agent harness, 并分析多轮智能体多消耗的 token 都花在了哪里。

03Publications论文发表

04CV履历

Education教育经历

Experience实习经历

The full CV, including projects, leadership roles and awards, is available as a PDF. 完整简历(包括项目、学生工作和获奖情况)可下载 PDF。

Download CV下载简历