Hi! I'm Qiyao (Catherine) Liang!
PhD student at MIT EECS at Massachusetts Institute of Technology
I'm a fourth-year PhD student in the Electrical Engineering and Computer Science department at MIT, advised by Ila Fiete from the MIT Brain and Cognitive Science department.
My central interest is emergence—in particular, the emergence of intelligence, both artificial and natural:
- → To understand artificial intelligence, I investigate puzzling and intriguing deep learning phenomena, including how diffusion models learn compositionality and how evolution strategies can finetune large language models.
- → To understand biological intelligence, I build models investigating how computational properties emerge from structural patterns arising from cortical wiring rules.
- → Even more fundamentally, I seek to understand the emergence of self-organization and self-replication—the very foundation of life itself—using artificial life as a platform and evolution as a tool.
I completed my undergraduate studies at Duke University in physics and math, where I worked on controlling and denoising quantum computers.
Research Areas
Recent Publications
Google ScholarModular connectivity in neural networks emerges from Poisson noise-motivated regularisation, and promotes robustness and compositional generalisation
arXiv preprint • 2025
Under review at Physical Review X Life, by invitation. We show how modular connectivity in neural networks emerges from Poisson noise-motivated regularisation, and how it promotes robustness and compositional generalisation.
Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning
arXiv preprint • 2025
Under review. We explore evolution strategies at scale for LLM fine-tuning as an alternative to reinforcement learning approaches.
Compositional Generalization via Forced Rendering of Disentangled Latents
ICML 2025 • 2025
Standard networks given separate x/y cues still memorize and stitch examples rather than combine them. Forcing each cue into the output space—via low-rank embeddings or simple stripe data—enables true compositional generalization.