About Me

I am a Ph.D. student at University of Texas at Austin (UT Austin). I am honored to be advised by Prof. Qiang Liu.

I will be graduating in 2026 and am actively seeking full-time positions in industry.

I am interested in developing fundamental yet computationally feasible algorithms for the basic learning, inference, and optimization problems that underpin the cutting-edge AI/ML/statistical technologies. These days, I am mostly drawn by the training efficiency of large-scale models.

If you share common interests, are interested in potential collaboration, or simply want to connect for a chat, feel free to contact me. I'm always open to conversation :)

Educations

University of Texas at Austin, US

Ph.D student in Computer Science (2022.8 - ~)

Advisor: Prof. Qiang Liu.

Highlights

Developed Cautious Weight Decay, used in all modded-nanogpt world records since its publication. Karpathy on nanochat: "Worked great out of the box on nanochat too, beat standard weight decay in a solid sweep."

Core contributor to DeMo, the optimizer foundation behind Nous Research's distributed-training stack, used to train the Hermes model series.

Set the current world record on Slowrun, the leaderboard for language modeling in the fixed-data, unlimited-compute regime; my submission has held the best validation loss on the benchmark for over a month.

Andrej Karpathy on using Cautious Weight Decay in nanochat: "Worked great out of the box on nanochat too, beat standard weight decay in a solid sweep."

Andrej Karpathy@karpathy

Worked great out of the box on nanochat too, beat standard weight decay in a solid sweep.

Hugging Face timm adopted Cautious Optimizers. Ross Wightman: "One of the last minute papers I added support for that delayed this release was 'Cautious Optimizers'... Consider me impressed, this boost appears more consistent than some of the new optimizers."

Ross Wightman@wightmanr

One of the last minute papers I added support for that delayed this release was 'Cautious Optimizers'... Consider me impressed, this boost appears more consistent than some of the new optimizers.

Contributing to google-deepmind/simply, DeepMind's minimal and scalable JAX research codebase for frontier LLM research.