Xiaoyu Li

stylethemescene

Last login: 2026-10-21 on ttys001 What does it mean for AI to help a person? Type a command, or just scroll. "help" lists them all.

Who I am

xiaoyu@home:~$ whoami

Xiaoyu Li PhD student · UNSW Computer Science and Engineering I’m trying to understand myself and the world I find myself in. Mathematics, art, and philosophy are ways I explore it. I’m still asking what matters and what I want to do with my time here.

Research questions

xiaoyu@home:~$ cat questions.txt

  1. [1]Purpose — What should AI help people pursue? How can people express, revise, and refuse goals? Learning a preference is not permission to act. keywords: Goals · Revision · Authorization
  2. [2]Agency — When does help empower rather than replace? Does assistance increase a person’s ability to understand and act? People should be able to choose how much to delegate and take back control. keywords: Assistance · Capability · Control
  3. [3]Evidence — How do we know the help was real? What observations justify a claim of benefit? When do guarantees survive connections between models, tools, and people? keywords: Identification · Verification · Composition

-> research-statement.txt

Essays

xiaoyu@home:~$ ls -lt essays/

total 25

  1. The New Eight-Legged Essays and the Smiling, Proud Wanderer 新八股与笑傲江湖
  2. Why Is the Golden Snitch Worth 150 Points? 金色飞贼为什么值150分？
  3. Can't Learn It? Is It the Concept, or the Teaching? 学不会，是概念太难，还是讲法错了？
  4. Werewolf with a Million Players: 674 Wolves Are Enough 一百万人的狼人杀，674 只狼就够了
  5. Why Don’t We Design Slower Algorithms?
  6. You Didn't Learn Badly. You Were Taught Too Well. 你不是没学好，是被教得太好了
  7. Why You Always Lose at Tetris 俄罗斯方块，为什么注定会输
  8. Chess on an Infinite Board 无穷棋盘上的国际象棋
  9. Above Infinity 无穷之上
  10. A Question with No Answer 一个答不出的问题
  11. Counting Past ω 数到 ω 之后
  12. I See It, but I Don't Believe It 我看见了，但我不信
  13. Infinities Come in Sizes 无穷也分大小
  14. The Universe Did Not Include a Progress Bar
  15. Isn’t It Just Math? 不就是数学嘛？十个箭头，和箭头背后的东西
  16. 722 Math Papers Land on GitHub Overnight 722 篇数学论文，一夜上了 GitHub
  17. AI Can Prove Theorems Now. Why Are We Still Writing Math? AI 都会证明了，我们还写数学干嘛？
  18. Is English the Best Programming Language? 英语是最好的编程语言吗？画家、钟表匠和巫师
  19. Why I Asked AI to Build a PEFT Atlas
  20. What Changes When AI Helps Us Write?
  21. Does Dota 2 Armor Really Diminish? DOTA2 的护甲，到底递不递减？
  22. In 48,000 Tang Poems, What Did Poets Write About Most? 四万八千首唐诗里，诗人最常写的是什么？
  23. From the Four Color Theorem to Navier–Stokes 从四色定理到 NS 方程：机器证明的五十年
  24. Reason and Feeling: Really Opposites? 理性与感性，真的是对立面吗？
  25. Hofstadter’s ABCD: Between Creation and Discovery 侯世达的 ABCD：在创造与发现之间

-> essays/

Selected papers

xiaoyu@home:~$ ls papers/ --selected

  1. [1]A Categorical Atlas of Parameter-Efficient Fine-Tuning: Cones, Orbits, and Gauges Xiaoyu Li, Zhizhou Sha, Chiwun Yang, Dai Shi · Technical report · 2026 Hundreds of fine-tuning methods, viewed through one mathematical framework. [pdf] [companion]
  2. [2]Flood and Harvest: The Provable Necessity of Trivia for Generating Valuable Mathematics via the Lens of Language Generation in the Limit Xiaoyu Li, Andi Han, Dai Shi, Zheng Gao, Jiaojiao Jiang, Junbin Gao · arXiv:2606.14688 · 2026 Discovering unseen valuable mathematics requires infinitely many correct but worthless statements.
  3. [3]Watermark Forensics for Generative Models: An Information-Theoretic Perspective Xiaoyu Li, Zheng Gao, Xiaoyan Feng, Jiaojiao Jiang, Yulei Sui, Jiankun Hu · arXiv:2607.13003 · 2026 We quantify the cost of user attribution, payload recovery, and watermark localization.
  4. [4]When the Canonical Completion Is Wrong: Formalizing and Measuring the Jump in LLMs Dai Shi, Xiaoyu Li, José Miguel Hernández-Lobato · arXiv:2608.26187 · 2026 We formalize and test whether LLMs can leave a canonical default when constraints demand a different structure.
  5. [5]Feature Superposition in Neural Networks: From Theory to Practice Dai Shi, Xiaoyu Li, Andi Han, José Miguel Hernández-Lobato · arXiv:2609.06862 · 2026 We review how neural networks encode more features than dimensions and how to recover them.
  6. [6]What Pretraining and Midtraining Make Learnable from Rewards? Chiwun Yang, Xiaoyu Li · arXiv:2609.38446 · 2026 Task-independent prediction can teach computations that rewards alone leave undetermined.

-> papers/ papers/2026/

Works

xiaoyu@home:~$ ls -F works/

  1. -rw-r--r--A Categorical Atlas of Parameter-Efficient Fine-Tuning Interactive atlas
  2. -rw-r--r--A Categorical Atlas of PEFT: the film Film · 17 min · narrated
  3. -rw-r--r--The golden flower Instrument
  4. -rw-r--r--The Universe Did Not Include a Progress Bar Illustrated essay · Six drawings
  5. -rw-r--r--Essays from Chinese social media Essays in Chinese and English
  6. -r--------Video collections Video · coming soon
  7. ----------Not yet imagined Open · not yet imagined

-> works/

Contact

xiaoyu@home:~$ cat contact

剑心犹在。 other ways in: the arcade · the book · the game world · the deep sea · the theorem · the voyage