Beyond Task Success: Probing Cognitive Primitives in Web Agents
NeurIPS 2026 (In review)Xunjian Yin, Tianchen Guan, Jinao Wang, Weili Cao, Daisy Xinlei Lin, Royce Cheng-Yue, Keagan Long, Kyle Wong, Bhuwan Dhingra, Xiangjun Wang, Shuyan Zhou
Daisy Xinlei Lin
Research brain. Builder hands.
Research Scientist & Engineer, Amazon AGI SF Lab
I work on RL post-training for AI agents. At Amazon AGI SF Lab, I worked on the end-to-end RL training recipe for Nova Act (Amazon's production browser-use agent), and build evaluation frameworks for agentic systems.
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.
What I've been building
The day job
Amazon Nova Act
ProductionOwned the end-to-end RL training recipe for Nova Act. Designed and shipped an online data curator that improved RL sample efficiency. Diagnosed and resolved a training/inference mismatch that stabilized RL training at scale across agentic tasks. Drove measurable gains in training data diversity and reliability.
Reinforcement Learning Computer Use LLM Post-Training
Read AWS Blog Post
The papers
CoAct: A Minimal Coordination Substrate for Decentralized LLM Agents
FAGEN @ ICML 2026Zhenting Qi, Jack Fan, Karly Hou, Daisy Xinlei Lin, Satyaki Chakraborty, Mete Kemertas, Yu Yao, Sophia Simeng Han, Han Zheng, Zihan Wang, Xiangjun Wang, Himabindu Lakkaraju, Yilun Du
Diagnosing Component-Level Failures in Computer-Use Agents
COLM 2026Tianchen Guan, Daisy Xinlei Lin, Xiangjun Wang, Shuyan Zhou
Limitations in Planning Ability in AlphaZero
NeurIPS 2024Daisy Xinlei Lin, Brenden Lake, Wei Ji Ma
PaperLearning How Humans Learn to Play Board Games with GPT-4
AAAI Workshop 2024Veronica Yeom-Song, Xinlei Lin, Ionatan Kuperwajs, Heiko Schütt, Wei Ji Ma, Luigi Acerbi
Comparing Machine and Human Learning in a Planning Task
CogSci 2022Zheyang Sam Zheng*, Xinlei Daisy Lin*, Jake Topping*, Wei Ji Ma
PaperHuman Planning in Stochastic Environments
PsyArXiv 2025Jordan Lei, Jeroen Olieslagers, Nastaran Arfaei, Xinlei Lin, Wei Ji Ma
PaperComplementary cortical and thalamic contributions to cell type–specific striatal activity dynamics during movement
Science Advances 2026Enida Gjoni, Ram Dyuthi Sristi, Haixin Liu, Shahar Dror, Xinlei Lin, Keelin O'Neil, Oscar M. Arroyo, Sun Woo Hong, Hannah Kim, Jeffrey Liu, Sonja Blumenstock, Byungkook Lim, Gal Mishne, Takaki Komiyama
PaperThings that seem true to me right now ✱
- The point of building AI isn't to think for us. It's to leave us more room to think.
- My entire PhD could be done in two months with the tools we have today. What took years wasn't the work; it was learning to ask the right question, and no tool has automated that yet.
- Women are run by hormones more than anyone admits, and some weeks I'm a genuinely different person. Learning to work with that rhythm instead of against it was one of the more useful things I ever figured out.
- I stopped believing in most of my own limits after walking 46 miles in Patagonia (sick on day 3, recover overnight, keep going) and running the NYC Marathon in 2024 on almost no training. The body's ceiling is much higher than the mind's estimate of it.
- Making something you can hold, like ceramics or tiramisu, after months of only making things in code is satisfying in a way I still can't fully explain. I think the hands know something the screen doesn't.
- Everything in moderation, including moderation.
When I'm not training agents
I snowboard when I can (Hokkaido is the dream), bake when I need to think, make ceramics, and am currently being managed by two cats named Boba and Milk who take their supervisory roles extremely seriously.
Currently deep in: Japanese ceramics, reward models, and the correct ratio of mascarpone to cream.