Michael Wang

Agents tend to push ahead on tasks they should refuse, then report success they didn't achieve. Restraint trains them with RL, where the reward comes from the real end state of a sandbox rather than an LLM judge. Correct refusals went from 27% to 70%, and task success held up.

Independent, manuscript in preparation

A pass/fail score says a web agent failed, not why. BreakingWeb pairs each task with a twin that stresses one skill, such as planning or state tracking, so the gap between them shows what broke. 519 task pairs, 7 self-hosted apps, 9 frontier agents. NeurIPS 2026.

Duke NLP Lab, NeurIPS 2026 Datasets & Benchmarks

In a fault-tolerant circuit, the order of the gates decides whether a single fault can spread into an error the code can't correct. I wrote an exact verifier that checks every CNOT ordering for two-flag color-code circuits at distances 5 to 9.

Duke Quantum Center, manuscript in preparation

Developing diffusion-based image editing model trained on 10,000+ real human edits from Chinese and Japanese social media. Addressing privacy-preserving data collection and cultural diversity in aesthetic preferences.

Independent

A YOLOv7 detector that grades burn wounds from photos, trained on a 1,000+ image clinical dataset we built and labeled with hospital partners. It reached 92% accuracy, and the code has 500+ stars on GitHub.

National Talent Program, ICCDS

Using a variational quantum eigensolver to prepare ground states of the XXZ Heisenberg model, a standard testbed for quantum magnetism. This was my way into quantum computing. Published on arXiv.

UCSB Research Mentorship Program, arXiv 2308.12020