Shan Yang

Researcher · RL Post-Training & Agentic Reasoning

I work on RL post-training — models that reason and use tools reliably. My focus is the unglamorous half of the stack: audited training data, hard evaluation benchmarks, reward design, and RL recipes that survive real-world deployment. Physics reasoning is where I proved it out (Physics-R1); I'm now extending it to agentic tool-use and reward hacking.

Currently Staff Applied Scientist at Adobe Foundry, building custom generative foundation models for enterprise customers (SFT, DPO, RL, with data curation as the main lever). Previously Tech Lead at Amazon Video Search (multimodal search + RL post-training, shipped to customers) and GenAI Live Action Studio, and Senior Research SDE at Google Research (multi-modal modeling, AIST++). PhD from UNC-Chapel Hill with Prof. Ming C. Lin on learning physical parameters from video.

Currently exploring: tool-use RL and reward hacking in agents (Physics-R2, in preparation).