Research
Research on LLM agents for software engineering, agent evaluation, and applied machine learning.
CUA-SWE: When Computer-Use Agents Meet Visual Software Engineering
Co-first Author · FORTIS Lab, USC · Jun 2026 - Sep 2026 · Under review at ICLR 2027
CUA-SWE studies when computer-use agents (CUAs), which can see and operate a browser, outperform code-only agents on visual web bugs.
- Curated 33 web repair tasks through an adversarial curation loop, adapting real GitHub issues and observed UI failures into cases that challenge code-only agents.
- Built deterministic behavioral verifiers and analyzed paired code-only and CUA trajectories across screenshots, browser actions, patches, and test outcomes to diagnose partial repairs, protocol violations, and frontier failures.
SkillStack: Modular Skill Agent Framework
Project Lead · Jun 2026 - Present · GitHub
A modular Python framework for studying skill-based LLM agents. Replaceable skill acquisition, retrieval, execution, and governance components share common interfaces, and a reproducible ALFWorld pipeline measures task success, token cost, and component compatibility with full execution tracing.
Publications
Workshop papers in visuo-tactile manipulation and personalized federated learning.
FELT: Generating Tactile Signals from Vision for Visuo-Tactile Manipulation
Z. Li, Y. Ling, Y. Gu, B. Huang, Chenhao Liang, S. Islam, others
Tactile Sensing for Robotic Foundation Models Workshop, RSS 2026 2026
Research on generating tactile signals from visual observations for visuo-tactile manipulation.
Data Heterogeneity-Aware Personalized Federated Learning for Diagnosis
H. Lin, H. Li, H. Li, X. Yu, K. Yu, Chenhao Liang, others
OMIA Workshop, MICCAI 2024 2024
Personalized federated learning research addressing data heterogeneity in medical diagnosis.