Wenlong Deng

My name is Deng Wenlong (邓文龙), I recently defended my Ph.D. in the Electrical and Computer Engineering department at the University of British Columbia, co-supervised by Prof. Xiaoxiao Li and Prof. Christos Thrampoulidis. My research focuses on long-horizon reinforcement learning and reasoning for large language models, particularly how to improve agents’ ability to solve complex multi-step tasks through better trajectory generation, reward design, credit assignment, and test-time search. I am especially interested in multi-turn reasoning, tool-using agents, and long-horizon data generation, with applications including coding, search, and medical diagnosis. I have also worked at Meta, Amazon, Google, and TikTok, where I applied these research ideas to practical, large-scale systems.

Previously: I obtained my master’s degree in Electrical Engineering at EPFL in 2019, where I was fortunated been supervised by Prof. Alexandre Alahi. I received my bachelor’s degree in Electronic and Information Engineering (Honors) at UESTC in 2017.

Personal Highlights

• Data-Centric AI: Scalable Data Valuation & Synthetic Data. Study how training data shapes model learning and develop scalable methods for data valuation, long-horizon reasoning data curation, and synthetic data. Representative works include Evolving Dynamics (arXiv 2026), For-Value (ACL 2026), GMValuator (ICLR 2025), CCD (arXiv 2024), and MedReason (ICML 2026 GFM Oral).

• Long-Horizon RL: Off-Policy Data, Reward Modeling & Policy Optimization. Study the learning dynamics of long-horizon reinforcement learning, with a focus on off-policy data, exploration–exploitation, and reward modeling. Representative works include NTHR (NeurIPS 2025), LLDS (ICML 2026), THR (ICLR 2026), Pass@K Surrogate (TMLR 2026), and Directional Reward Alignment (ICML 2026 AIWILD).

• Industry Deployment: End-to-End Long-Horizon Agent Systems. Built end-to-end agent pipelines spanning long-horizon trajectory generation, milestone-based reward labeling, and stable multi-turn policy training. Applied these systems at Amazon and Meta, translating research advances into scalable training infrastructure and production workflows.