Purpose
What should AI help people pursue?
How can people express, revise, and refuse goals? Learning a preference is not permission to act.
Goals · Revision · Authorization
PhD student · UNSW Computer Science and Engineering
Xiaoyu Li
I’m trying to understand myself and the world I find myself in. Mathematics, art, and philosophy are ways I explore it.
Nigredothe black, where a question begins
What does it mean for AI to help a person?
I work on learning theory and explore formal models of goals, assistance, and evidence.
The research statement
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.
Assistance · Capability · Control
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?
Identification · Verification · Composition
Albedothe white, where thought is written clear
Essays & notes
The New Eight-Legged Essays and the Smiling, Proud Wanderer新八股与笑傲江湖
Why Is the Golden Snitch Worth 150 Points?金色飞贼为什么值150分？
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？
Werewolf with a Million Players: 674 Wolves Are Enough一百万人的狼人杀，674 只狼就够了
Why Don’t We Design Slower Algorithms?
You Didn't Learn Badly. You Were Taught Too Well.你不是没学好，是被教得太好了
Citrinitasthe yellow, the first light of a proof
Selected recent work
A Categorical Atlas of Parameter-Efficient Fine-Tuning: Cones, Orbits, and Gauges
Hundreds of fine-tuning methods, viewed through one mathematical framework.
Flood and Harvest: The Provable Necessity of Trivia for Generating Valuable Mathematics via the Lens of Language Generation in the Limit
Discovering unseen valuable mathematics requires infinitely many correct but worthless statements.
Watermark Forensics for Generative Models: An Information-Theoretic Perspective
We quantify the cost of user attribution, payload recovery, and watermark localization.
When the Canonical Completion Is Wrong: Formalizing and Measuring the Jump in LLMs
We formalize and test whether LLMs can leave a canonical default when constraints demand a different structure.
Feature Superposition in Neural Networks: From Theory to Practice
We review how neural networks encode more features than dimensions and how to recover them.
What Pretraining and Midtraining Make Learnable from Rewards?
Task-independent prediction can teach computations that rewards alone leave undetermined.
Rubedothe red, where the work is made real
Works
A paper is one way of knowing. Jung counted four: sensation tells us that a thing is, thinking what it is, feeling what it is worth, and intuition where it might lead. The work here is gathered the same way.
Film · 17 min · narratedA Categorical Atlas of PEFT: the filmWatch the filmFeeling
tells us what it is worth
Intuition
tells us where it might lead
- A Categorical Atlas of Parameter-Efficient Fine-Tuning
- Not yet imagined
The centrewhere every path returns
Hello, I’m Xiaoyu Li.
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.
Jung called the face we turn toward the world the persona, and the whole person behind it the self.