I am an independent AI researcher working on architectures for better reasoning, efficiency, memory, and continual learning. My background spans economics, computer science, and mathematics. I studied at Tsinghua University and Duke University. Outside research, I lift and play video games.
Research overview
At a high level, my current research areas are:
- New architectures for improved efficiency and inductive biases
- Memory, continual learning, and reinforcement learning
- Approaches inspired by symbolic methods
I approach research with:
- Ideation and experimentation as complements
- Simplicity as a prior in both research and engineering
- Ideas drawn across disciplines
I conduct my independent research through
Iter Labs (iter as in journey in Latin and iter() in Python).
I’m open to technical roles and collaborations.
Selected works
GM-2: Ultra-sparse 2–4/4096 dynamic routing replaces most Qwen3-0.6B layers, preserving pretraining loss over 15B tokens.
Paper
Two intuitions for Graph Machine (2), one algorithmic and one biological.
Blog
GM-1: Investigating whether explicit, dynamically updatable edges improve relational reasoning on Sudoku.
Paper