Zhuowei Li

second-year cs student. currently: do self-improving agents really get better?

fast learner · curious · interest-driven

Oct 13–20, 2026
After seeing Kimi’s No Genre event, I quickly chose a direction I find interesting, recursive self-improvement, and learned the field. Then I ran a study on whether self-improving agents’ gains are real, and built an open-source toolkit for checking them.
a studyIn a self-improvement loop on Kimi, an agent revised its own playbook to raise an LLM judge’s score. Judge scores rose while accuracy against the answer key fell by 12 points. The pattern held across three judges (Kimi K2.6, Kimi K3, DeepSeek) and two real-world tasks, business-data analysis and contract review, with predictions pre-registered.
toolkitrealgain takes any self-improvement loop’s evaluation log and separates each claimed gain into real improvement, luck, fitting the selection tasks, a fooled judge, or the agent reading the answer key. GitHub ↗
Jun–Sep 2026
Tried running a cross-border e-commerce business on the independent-store model. Started hiking.
My second trek was already one of the harder routes in western Sichuan :)
Mar–May 2026
First research project, as sole lead: a small closed loop in which a diffusion model, a scorer and an RL agent tried to improve a medical-image classifier with generated data.
I learned more from what went wrong: the scorer’s rating rose 55% while the classifier didn’t improve, the RL had quietly become a grid search, and a “best” result over 3 runs vanished over 20. These became the questions behind my RSI project.Details → · GitHub ↗
Mar 2026
Went from a first neural network written by hand in NumPy to a diffusion model (DDPM) in ten days.
When something genuinely interests me, I focus on it completely. GitHub ↗
Nov 2025
Joined a lab and started learning deep learning on my own.
The lab was new and tiny, with almost no guidance. Over the next month or two, I taught myself the foundations of machine learning and deep learning, up through the Transformer.I was in a flow state every day, and for the first time, I truly enjoyed studying.Over winter break, I learned the basics of reinforcement learning.
Sep 2025
Started computer science at Northeastern University (Qinhuangdao), with almost no coding background.
That summer, while trying to spend less time on screens, I stumbled into an early agent beta from Zhipu. That was my first real look at AI agents.
2022
Captained my high school’s AI maker club to a provincial award. The same year, I was selected for China’s national Youth Talent Program in physics.
clubI joined because I like trying new things.
physicsA national program: universities select promising high-school students, and professors train them in research for free. Mine was at USTC.Since middle school, curiosity has drawn me to physics and philosophy, and I loved reading and thinking about both. In high school I also taught myself some university physics, just out of interest.I left the program soon after joining, on a gut feeling that I’d end up in computer science.