SOP解析：3篇拿到麻省理工博士录取的神级范文

写在前面：SOP 到底要解决什么问题

Statement of Purpose 不是自传，也不是获奖清单。它要回答的核心问题只有一个：为什么这个项目应该录取你，并且为什么是这个项目、而不是别的项目。下面这份写作总纲，是我们从大量成功案例中反复验证过的四条基本要求。

• 说服 faculty committee，证明你具备入读该项目的资格，并且你与该项目的培养重点和目标高度契合。
• 展示一系列精选的技能与经历，用以体现你的科研能力、技术成就与学术兴趣。
• 包含具体且可量化的经历、成就与资质，而不是形容词的堆砌。
• 严格遵守你所申请项目的篇幅与格式要求。

下面三篇范文均来自 MIT 博士录取者。为便于阅读，我们在每一段前标注了该段承担的功能——这也是招生委员会读你材料时，脑子里真正在做的事。

第一篇 · 机器人与人工智能方向

功能：简要介绍研究经历，并概括你的主要贡献。对每个项目，清楚说明研究问题、具体贡献（要够技术）与研究成果，以证明你在该领域的知识深度。

Research in Robotics

I have been working in XX Lab since my second semester in EECS at UC Berkeley. The lab focuses on artificial intelligence for personal robots in households. My work in the lab is in hierarchical task and motion planning (TAMP), which seeks to unify two types of reasoning: logical and geometric. Logical reasoning involves specifying the various actions a robot can take within its environment, using an abstract representation. Geometric reasoning considers constraints such as obstructing objects and reachability. Efficient integration of these two types of reasoning is a major open problem that our work seeks to address. Our approach allows robots to execute long task sequences that require many individual arm and base motions, such as doing laundry, picking an object out of clutter, and setting a dinner table.

In my first few semesters, I worked on implementing our TAMP system. My main contribution was building an interface between our planner and the robot's perception system, so that we could move beyond simulation and run real-world experiments on our lab's PR2 robot. I also helped implement the dinner table task, where the robot had to transport cups and bowls from one table to another using a tray. In our experiments, the robot was able to plan for and execute long task sequences with high success rates, a breakthrough in robot capabilities. The work led to a conference publication at ICRA 2014.

In real-world situations, there is often uncertainty about the environment a robot interacts with. Since the ICRA 2014 system did not address planning in these situations, I worked with a graduate student and another undergraduate on a novel algorithm for TAMP in belief space. Here, the exact locations of objects are unknown; instead, a distribution over possible locations is maintained for each object. My contributions included implementation of a Monte Carlo approach to computing collisions in belief space and a randomized algorithm for sampling robot trajectories used in motion planning. I also co-implemented the detection algorithms. We were able to apply our belief space planner to challenging experimental tasks, such as navigating a narrow corridor containing obstructions with unknown locations and finding a key located in one of several possible drawers. The work led to a conference publication at IROS 2015.

A core issue with TAMP systems is often their reliance on hand-coded heuristics to guide search. As a result, these systems did not easily generalize or scale robustly to more challenging problems. I saw an opportunity here to integrate machine learning into TAMP. Under the guidance of a graduate student, I took the lead on developing a reinforcement learning approach to learn how to bias the search spaces. Compared to the original system, my system significantly improved robustness and success rate (e.g., 42% to 86% on a task of picking up a particular can from a cluttered table containing 30 cans, a standard benchmark task). The work led to a first-author workshop publication at MLPC in IROS 2015 and a first-author conference submission to ICRA 2016.

Research in Natural Language Processing

I believe research at the junction of natural language processing and robotics will be critical to bringing robots into our everyday environment. This belief motivated me to work with Professor XXX in natural language processing. Our project seeks to improve the performance of neural machine translation systems. A major limitation of these systems is that the vocabulary size must be small due to expensive computation of a normalization term. We observed that performance could be improved by compressing the training data into a smaller symbol set using invertible Huffman codes, thus circumventing the vocabulary size issue. I developed and ran evaluations for this system on my own. Against a baseline, our approach yields improvements in translation quality of up to 1.7 BLEU points on English-French translation for a widely-used dataset. The work led to a first-author conference publication at EMNLP 2015.

STEM Outreach

功能：用具体例子展示你对 STEM 公共参与的热情与投入。

Doing impactful work to push the frontier in robotics is only one aspect of what I hope to accomplish in graduate school — I also want to place a strong emphasis on outreach. For the past two semesters, I have been the lead of, and point of contact for, my robotics lab's outreach efforts. This is a critical aspect of our lab, as it represents the communication layer between the technical work we do and the public, from children in elementary school to visiting investors to journalists from Bloomberg. Additionally, I have been a Teaching Assistant for five semesters across three courses: artificial intelligence, introductory computer science, and computer architecture. Student feedback on my teaching averages 4.7/5.0, leading me to receive a 2015 UC Berkeley Outstanding Graduate Student Instructor Award, given to the top 10% of Teaching Assistants across the university.

Future Work

功能：证明你确实做过功课，了解该校的研究强项。

I believe MIT would cater well to my aspirations because of the diverse strength of its faculty and students. Specifically, I would be excited to work with Professors Leslie Kaelbling, Tomás Lozano-Pérez, and Russ Tedrake. As I have spent my undergraduate research career gaining a deep understanding of hierarchical planning, I would certainly be interested in continuing this line of work in graduate school. I hope to integrate machine learning and natural language processing into TAMP. Professors Kaelbling and Lozano-Pérez have done astounding work together in manipulation planning using TAMP methods, such as Hierarchical TAMP in the Now, the FFRob heuristic, and hybrid backward-forward planning. Another major benefit would be that I could bring the perspective and experience I have accrued from working on hierarchical planning at Berkeley to MIT.

功能：回应文书题目中关于“你的具体兴趣，以及 MIT 如何支持这些兴趣”的提问，并交代长期职业目标。

In a more general sense, I am very excited by applications of machine learning to AI in robotics, such as deep reinforcement learning. Professor Tedrake's work in optimization-based control for locomotive robots with constrained dynamics is particularly fascinating to me, as recently deep learning approaches have led to breakthroughs in autonomously learned locomotive capabilities; I would be very interested in contributing. After graduate school, I hope to work in an academic setting toward professorship.

第二篇 · 量子信息方向

功能：简要介绍研究兴趣与背景；对每段经历说清研究问题、具体贡献与可量化成果；明确展示你对目标项目所在研究方向的热情与投入。

My research interest lies in quantum information science and technologies, especially quantum information processing. Building on a strong background from my undergraduate research in this field and the related field of photonics, I am applying to MIT's PhD in EECS to conduct research with the top experts in this field.

Research in Photonics

My interest began with studying light at the nanoscale. In the summer of my first year, I started as a research assistant to Prof. XXX in University of Toronto's XXX Group. I developed expertise in modelling, optimization, and fabrication of photonic nanomaterials and devices, especially light-matter interactions involving plasmons. I thrived in the research environment and received another fellowship to continue my research with the group part-time during the school year. During my second year, I presented my modelling and optimization work as a research poster at the 8th Symposium on Functional Coatings and Surface Engineering. Over the next year, I co-authored a device patent and a journal article on photocatalysis. Furthermore, with the trust and support of the group, I self-initiated a study that led to the surprising discovery of localized surface plasmon resonance in optical films from interface diffusion. Working independently during my third school year, I formulated a theoretical model that accounted for the observed optical phenomenon and made critical predictions about the necessary material morphology that were subsequently validated by additional experiments. After months of exacting revisions for scientific rigour as well as clarity and brevity of writing, the year-long effort culminated in my first-author manuscript currently in review at Nature Photonics.

Research in Quantum Information

The two years of experience in modelling nanoscale light-matter interactions led me naturally to the booming field of quantum information processing. I was fascinated by the potential of a powerful quantum computer built from light-matter interactions between microwave photons and superconducting circuits. In the summer of my third year, I shifted focus and joined Prof. XXX's group at Université de Sherbrooke. With an Institut Quantique Fellowship, I undertook a new project on improving the resonance-induced phase gate with our experimental collaborators at the Engineering Quantum Systems group at MIT. It was an extremely enriching experience to join a starting project and develop the full spectrum of skills in designing superconducting circuits, making and validating approximations, modelling interactions, and optimizing device parameters. By the end of the summer, I was fluent in basic Quantum Electrodynamics, the various phase space methods, analytics with Mathematica, and numerics with QuTiP. Armed with these research tools, I was able to present my research, within and across disciplines, at lab meetings and at the XXX Workshop.

Teaching and Outreach

In addition to research, I strive to be a reliable source of knowledge in any field I can. I served as a TA for "Algorithms" and "Decidability and Tractability," both early major requirements with over a hundred students from diverse backgrounds. Additionally, I was invited to help redesign and TA a computer graphics class, for which I created course notes and an assignment on bump mapping and mesh data structures. Outside of coursework, I was a lead developer on a project adding accessibility features to Minecraft: my team added 3D acoustics, text-to-speech, and keyboard shortcuts to allow Minecraft to be used by a much broader audience, and the project won 2nd place in XXX. My past experiences teaching and presenting my research have made me strong at communicating complex topics to audiences of varying backgrounds and levels.

第三篇 · 计算机图形学与几何处理方向

功能：简要介绍研究兴趣与背景；对每段经历说清研究问题、具体贡献与可量化成果。

Throughout my life, I have been compelled by a desire to understand what fundamentally drives seemingly complex systems. In high school, I took a class that dissected the ideas in Hofstadter's Gödel, Escher, Bach. One idea that resonated strongly with me is that large observable phenomena are usually made of smaller, elusive yet simple components. This idea was reinforced when I discovered the applications of discrete differential geometry (DDG) in computer graphics. The DDG framework beautifully constructs fundamental building blocks in discrete geometry, leading to simple and intuitive descriptions of the complex behaviors of hair, cloth, water, and more. I want to continue studying simulation and geometry processing as a university professor. Earning a Ph.D. in Computer Science at MIT would allow me to pursue my research interests and give me new opportunities to contribute to the field.

My past research experiences have helped me discover new ways of thinking and pinpoint my interests. In Spring and Fall of 2014, I joined Prof. XXX and Dr. XXX in studying the control of decision making. An earlier paper by the lab provided compelling evidence that the inferior lateral prefrontal and frontopolar cortex act as a hub of arbitration. We sought to further validate this result by studying the arbitration between model-based and model-free learning using fMRI data and Dynamic Causal Modeling (DCM). I computed the exceedance probability for hypothetical biological networks and used similarity scores with known brain networks to determine the most likely models. We then perturbed the remaining models and iterated DCM analysis to obtain the best model. I learned how it is possible to incrementally gain insight into the behavior of something as complicated as the human brain. While I greatly enjoyed the analysis, ultimately I did not have the necessary background in neurobiology to continue.

In the summer after my junior year, I participated in the XX Program. I had the honor of studying quantum secret sharing (QSS) with Professor XXX. The classical (n, k, L) secret sharing scheme encodes L symbols into n shares such that any k of the shares can be used to recreate the original secret. We expanded QSS by defining quantum strong security, a far more nuanced condition than its classical analog. Furthermore, I created a QSS protocol and proved that it holds the integrity of quantum strong security. Strong quantum security is critical to the eventual use and safety of quantum computation. This work was published in XXX. During the weekly lab meetings and discussions with my mentor, I gained an international perspective of higher education and was exposed to many exciting mysteries in quantum information.

My interests took a turn in senior year during Professor XXX's class on discrete differential geometry. The class involved reproducing theoretical results from differential geometry and placing differential forms on mesh locations in order to create mesmerizing displays of heat flow, mesh fairing, parametrization, and vector field design. I was immediately enthralled by the simple and efficient algorithms that could model complex behaviors through the theoretical elegance of DDG. I sought to dive deeper and was fortunate to be advised for the next two terms on preserving vortical structures in fluid simulation. I surveyed a large range of fluid simulation techniques such as Steinhoff's vorticity confinement, Mullen's energy-preserving method, and Elcott's circulation-preserving algorithm. I then derived a unique method as an extension to Dupont's back-and-forth error compensation and correction (BFECC), implemented using circulation-preserving, semi-Lagrangian advection. I presented my results at the XXX undergraduate research seminar, showing that increasing recursion depth preserves vorticity and kinetic energy better than BFECC.

三篇范文的共同点

把三篇放在一起看，你会发现它们的写法高度一致：

• 每段研究经历都遵循同一结构：研究问题 → 我具体做了什么（技术细节） → 产出了什么（论文、数据、指标）。
• 用数字说话：42% → 86% 的成功率、1.7 BLEU、4.7/5.0 的教学评分。可量化的成果永远比“取得了显著成效”更有说服力。
• 敢于写“我不擅长的部分”：第三篇坦诚交代自己因缺乏神经生物学背景而转向，反而让研究品味的演进显得真实可信。
• fit 段落点名到具体的人和具体的工作：不是“贵校师资雄厚”，而是说得出某位教授的某篇工作，以及自己能接上哪一段。

如果你正在准备博士申请的 SOP，欢迎把你的研究方向和目标项目发给我们。卡方留学，懂科研、懂求职、更懂留学！