Research Interests
I’m interested in designing RL environments in which agents can genuinely learn. Designing a non-hackable environment is a lot like sculpting a work of art. As agents become more capable, the environments they require grow increasingly complex, and the distinction between a well-designed environment and a poorly designed one becomes ever more consequential. I am convinced that environments will become the most crucial component of post-training.
Previously, I worked on post-training LLMs for efficiency (reducing token and tool usage), safety (abstention), and general capability (math and question answering). Back then, I mainly designed rewards that shaped model behavior, along with tasks that probed how well those rewards worked.
Education
Experience
📫 Contact
Email: jinyansu6[@]gmail.com