|
Frank Wu I'm a Master's student in Machine Learning at Carnegie Mellon University. I'm interested in exploring emerging paradigms for training neural networks, with an emphasis on reinforcement learning, continual learning, and generalization. Previously, I was a Machine Learning Engineer at Alibaba Group, where I built action recognition models and visual search engines serving 50M+ daily active users. I received my B.A. in Computer Science & Mathematics from New York University. |
|
News |
|
Publications |
|
Random Remasking Scales Masked Diffusion Inference
Zhaoyi Zhou, Frank Wu, Andrea Zanette arXiv preprint, 2026 arXiv We introduce Random Remask-Denoise (R2D), a finetuning- and reward-free test-time procedure, and Trajectory Correction Training, achieving near-log-linear inference scaling and strong gains on Sudoku, 3-SAT, and HumanEval+. |
|
Local Reinforcement Learning with Action-Conditioned Root Mean Squared Q-Functions
Frank Wu, Mengye Ren International Conference on Learning Representations (ICLR), 2026 website / arXiv / code We propose ARQ, which unifies Forward-Forward learning with TD updates for fully local reinforcement learning using action-conditioned, vector-valued value functions; we achieve 50% average return improvement and match backprop-based baselines on MinAtar and DeepMind Control Suite. |
|
Website design inspired by Jon Barron |