Hello👋, I’m Jinning Zhang, from Nanjing, Jiangsu. I’m a master’s student in the 2024 cohort at Beijing Sport University. I’m currently looking for PhD opportunities starting in 2027 or 2028. My research area is computational neuroscience, specifically neural signal processing/electrophysiology.

Here is a look back at the first two years of my master’s:

• In my first semester, under the guidance of a senior labmate (I am sincerely grateful to her), I learned surgical procedures for in vivo electrophysiology and basic neural data analysis using software.
• In my second semester, I began picking up the Python I had learned as an undergraduate, because the data analysis I could do with software was too limited. That semester, I also went to a biotechnology company in Nanjing to learn how to make multichannel electrodes, and successfully recorded neural signals using electrodes I had made myself.
• Also that semester, I happened to come across a computational neuroscience paper and began to fall in love with the field. Computational neuroscience requires a lot of mathematics, so I began reviewing the mathematics I had learned as an undergraduate and trying math problems from the postgraduate entrance examination. I kept this up until the end of the summer break and had largely mastered the mathematical fundamentals. (The following summer, I started reviewing mathematics again. Because I forget it so quickly, I decided to review the fundamentals every summer.)
• Over the summer, I attended two summer programs: the 2025 Neuromatch Academy Computational Neuroscience summer school and the 2025 Fourth Training Course on Computational Neural Modeling and Programming. I met a group of like-minded friends, some of whom I am still in touch with.
• In the middle of the summer break, I also entered a machine learning-related postgraduate competition (the National Innovation Cup Big Data Competition) to test what I had learned from studying Python on my own. Although it was a team competition, I couldn’t find anyone to join me, so I entered alone, stayed up late for several nights (mainly because writing the paper was very difficult for me at the time), and won third prize.

• Encouraged by the award and influenced by the atmosphere at the time (there was era-defining news every day, such as the release of GPT-5), I began exploring AI further and made it a secondary line of research during my master’s.
• After about a month of introductory study, I had an initial idea for my first paper when the semester began. I wanted to develop a domain-specific RAG project for sports medicine.
• Throughout my third semester, I worked on this idea. Every day, I read papers, ran programs, and kept trying things and learning from errors. I also asked members of my research group to help collect publicly available text corpora. At the end of 2025, I finally completed the project. My supervisor helped me contact doctors for a real-world evaluation, and the results were excellent. A statistics PhD student at the University of Hong Kong helped me improve one of the paper’s core algorithms. Thanks to their help, I finally released the preprint on January 1, 2026.
• My plan was to submit to a computer science conference first, then expand the paper for a journal after acceptance. However, the submission process brought repeated setbacks. Several reviewers raised a similar concern: the paper relied too heavily on LLM-as-a-judge evaluation.
• I decided to study LLM-as-a-judge. During the winter break, I kept thinking about a question: Why is LLM-as-a-judge considered unreliable? What criteria establish this? Influenced by the mathematical modeling books I was reading at the time, I realized that the degree of its “unreliability” could only truly be quantified by explicitly modeling LLM-as-a-judge as a form of information compression.
• With this idea in mind, I began the following semester’s work. To cut a long story short, I collaborated with computer science PhD students from FDU and SJTU to model and quantify distortion in LLM-as-a-judge. Our modeling approach was clearly influenced by ideas from source analysis in signal processing. The paper was submitted in May and accepted to the EMNLP main conference. I had never dreamed that my first publication would be at a top conference in NLP.
• The run of setbacks with my first paper seems to have come to an end recently. The ARR review scores were released in September, and I think the paper has a very good chance of being accepted at EACL. This semester, I’m planning to expand it (expanding and revising a manuscript from almost a year ago is a curious experience), hoping to submit an extended version to a journal after acceptance.
• I think I will focus more on computational neuroscience this academic year—everyone has their own area of expertise.

• Returning to my main line of research: at the beginning of this year, I began systematically studying signal processing, alongside physics topics such as neurodynamics. However, I found that I preferred the various algorithms and concepts in neural signal processing. I also enjoy reading signal processing papers from other fields, such as seismology and geophysics, and mechanical fault diagnosis.
• I had just finished recording my experimental data, so I used them to try out many interesting projects. When the semester began in March, I successfully applied to the university’s program for innovation and outstanding talent with a project titled “Effects of Aerobic Exercise on Pathological Dynamical Patterns in Lateral Habenula Neural Networks in a Mouse Model of Depression.” I wanted to identify the specific targets and mechanisms through which exercise interventions affect mice in a model of depression, and try to explain these effects through neurodynamics.
• One day that semester, I happened to come across a signal processing paper related to geophysics. Its main contribution was an algorithm combining minimum-norm least squares with wavelet filtering to process time series with gaps. Since I was also reading papers on brain–computer interfaces at the time, I immediately thought of EEG artifact estimation in neural signal processing. After reviewing the literature, I settled on the following research question: In the presence of localized missing data, how can we prioritize preserving the physiological quantities most important for MI-BCI—mu/beta ERD? Two months later, I submitted the paper to BIBM. The results will be out in 10 days.
• This summer, my main work was processing an in vivo electrophysiology dataset from 10 years ago. It was exceptionally large (1 TB) and messy (multiple brain regions, tasks, and age groups). I spent a considerable amount of time processing it, though most of my time went into choosing a research question, because the topic involved AD and I didn’t know much about the disease.

• There is one other project I am most proud of. In the last few months of 2025, I built a research website for sports science entirely on my own—sports-matter.com. Details of its features can be found in the introduction on the WeChat official account. In any case, I believe the features I designed were very comprehensive and ahead of their time. After its release in January this year, I “completed the platform’s launch and promotion, managed a platform community of over 500 members, and continuously iterated on the platform based on feedback from real users; total registered users exceeded 1,000, with profits of over RMB 6,000 before it was open-sourced.” Recently, I have been preparing to turn it into a desktop application and focus on AI4S capabilities.

Now in my final academic year, I’m planning to apply for PhD programs starting next year or the year after. My motto is a song lyric—Somehow I will try. So, if given an opportunity, I will try, whatever the odds. I firmly believe that something must be tried before it can be declared unworkable, and I enjoy trial and error. So—all I need are some opportunities.