I’m interested in methods that make AI systems better, both at learning and at running efficiently. I’ve been deep into [SFT + DPO on a 32B game agent in production at Tencent; SFT/RL data pipeline for DoorDash’s LLM recommender] and [SkillLearnBench, COLM ’26]. Recently, I’m especially curious about [how can an agent keep what’s useful from its experience and discard the noise, when the environment keeps changing?].
I write things down until they stop confusing me, and some of it ends up here. Feel free to connect with me via email at [email protected].