Unified steerability benchmark evaluating whether humanoid behavior foundation models faithfully follow diverse physical instructions across 14 tasks and multiple constraints.

Contribution: Second Author · Task Taxonomy & Multimodal Benchmark Curation

• Co-authored arXiv 2026 preprint (2nd author), advised by Prof. Juncheng Li.
• Formulated 3-level steerability hierarchy: Conditional, Constraint, Compositional.
• Standardized multimodal evaluation samples for 773,716 tasks from 38,522 source motions.

Investigating whether vision-language-action policies are vulnerable to semantic instruction paraphrasing in closed-loop manipulation. Evaluates Paired BC and Flow Consistency on LIBERO-Spatial.

Contribution: Lead Researcher & Engineer (Independent Project)

• Fine-tuned SmolVLA action expert with PyTorch/LeRobot on LIBERO-Spatial.
• Implemented paired BC and frozen-teacher flow consistency with matched training controls.
• Observed 94.1% canonical success (paired BC) vs. 70.6% (canonical FT); documented lack of transfer to unseen paraphrases (8% success in 75-rollout pilot).

Human-AI co-creation environment breaking complex app design into functional hierarchy, design system wireframes, and component interfaces to prevent black-box hallucination.

Contribution: Co-Author · GenUI System Implementation & Empirical Study

• Co-authored ICIC 2026 paper (Springer LNCS), advised by Prof. Liuqing Chen.
• Co-developed progressive GenUI web app with React, TypeScript, React Flow, and LLM prompting.
• 16-participant user study vs. Google Stitch showed +19.4% design consistency and -27.4% time per screen.