How I Am Learning To Do Research

I am a PKU mathematics student trying to turn the behavior of large models into evidence I can actually inspect.

Right now, the question that keeps pulling me back is large-model interpretability, especially mathematical reasoning. When a model reaches an answer, what is really carrying the work: a representation, a circuit-like path, a training signal, or a shortcut that only looks convincing from the outside? I care less about making a fluent story after the fact than about finding handles that survive ablation, attribution, and patient counterexamples.

Most of my work so far is project-shaped rather than paper-shaped: course reports, challenge entries, small agent systems, and tools I built while teaching or helping others build. I keep that boundary visible because it is part of how I am learning research taste. I want this site to show the making of questions: where an idea came from, what evidence I trusted, what I had to revise, and where the claim should still stay modest.

The personal side is not decoration to me. I play Texas Hold'em for the discipline of incomplete information, travel with a camera because places change how I notice structure, and write because a good research life also needs a voice. The formal record is still at /cv/; this page is closer to the person learning how to make that record mean something.