Daisy Xinlei Lin

I build zero-to-one SFT/RL recipes for frontier agents: identify failure modes, turn them into targeted data and environments, design evaluations, and debug production-scale training until it learns stably. I own this process for a 100B mixture-of-experts coding agent, raising Terminal-Bench 2.0 to 45% and reaching 70% on SWE-bench Verified, outperforming leading open-weight baselines. Previously, I built and shipped Nova Act’s zero-to-one RL recipe—including algorithm experiments, environment scoping, online data curation, and fixes for reward collapse and train–inference mismatch. Nova Act launched with state-of-the-art performance on REAL Bench and WorkArena. I enjoy driving a research agenda, from choosing the right problems to carrying long-running work through to impact. Across both projects, my favorite part is finding the failure no one can quite name yet, making it measurable, and building the training system that fixes it.

My PhD at NYU was about how humans plan and learn: I ran large-scale experiments to understand what makes people get better at complex tasks, and how that compares to AI.

I think AI should be humans' thought partners, adapting to collaborate with us better and better.

For fun, I explore new interaction paradigms for human-AI collaboration: tools that adapt to how you think, not the other way around.