research interests
A compact map of the questions I want this site to signal first.
Multi-agent systems and agent swarms
My main interest is how multiple agents can divide work, preserve useful context, produce inspectable artifacts, and make progress without turning coordination into noise. I care especially about formats for roles, traces, handoffs, evaluations, and shared state.
Agent-system format and verification
A strong model is only one component of an agent system. I am drawn to the runtime layer: assignment manifests, execution boundaries, context pieces, judges, metrics, and failure attribution. The goal is to make long agent workflows easier to reproduce and debug.
Scientific swarms
In Prof. Dequan Wang's lab, I am thinking about agentic harnesses for research work: systems that can run experiments, compare evidence, and leave enough trace for a human to audit what happened. More lab context is available at dequan.wang .
Embodied agents as prior work
My earlier work around visual language navigation and embodied AI still shapes how I think about agents: they need local memory, task decomposition, changing observations, and evaluation that matches the actual environment. I now treat that experience as one source of design pressure for broader multi-agent systems.