My research asks: how do we move LLMs beyond pattern matching toward genuine understanding?

I think of it like Taylor expansion. Prompt engineering gives us first-order approximation—linear workflows. RL and fine-tuning add second-order terms—reasoning chains. But true creativity lives in higher-order terms.

My work focuses on mental models—structured ways of understanding that could give LLMs higher-order capabilities. Current focus: agentic deep research. End goal: machines that genuinely predict, not recall.