Why keep an open research notebook?
Research rarely moves in a straight line. A question becomes an experiment. The experiment reveals a faulty assumption. An approach that looked promising stops working, and a small observation opens another direction.
This notebook is a place to keep that journey visible.
I’m Jiazheng Li, a Computer Science Ph.D. student at Tsinghua University, advised by Jingzhao Zhang. I study machine learning, particularly LLM optimization, training dynamics and reinforcement learning. Read more about my background or explore LucidGPT, What-s-up and the AI4Game notebook.
The story between the results
A formal paper needs a clear argument and a focused set of results. Those constraints leave out much of the process: the alternatives considered, the evaluation that changed, or the failure that made the problem easier to understand.
Here, those parts can stand on their own. A useful note might be an early research question, a failed attempt with a careful explanation, or a successful direction with its limitations still attached.
An experiment does not need a positive result to be worth documenting. It needs a clear question and an honest account of what happened.
What belongs here
- Research ideas: questions, hypotheses, assumptions, and connections worth exploring.
- Experiments: the setup, observations, interpretation, and remaining uncertainty.
- Failed attempts: what was tried, why it seemed reasonable, and what the outcome changed.
- Project development: showcases and workflows behind LucidGPT, What-s-up, AI4Game and future projects.
The aim is to distinguish evidence from interpretation. A surprising result is a starting point for investigation; a working prototype is a starting point for evaluation.
A notebook that can change
These posts can be smaller and more provisional than a paper. When an idea changes, a follow-up note can explain why. When an experiment becomes reproducible, its code, configuration, and artifacts can accompany the write-up.
The first entries provide a project outline and writing templates. They establish a place to begin; they do not claim completed experiments or research findings.
The project showcases introduce work described in my CV. The visualization gallery and LucidGPT article show how interactive figures can explain a process; their numerical examples are synthetic.
The question behind this notebook is simple: what can be learned when the process is shared alongside the outcome?