Publications

Selected work on AI security, trustworthy language models and agents, and embodied intelligence. The full publication list is available on Google Scholar ↗.

SkillReducer description minimization and skill-body compression pipeline

SkillReducer: Optimizing LLM Agent Skills for Token Efficiency

Y. Gao, Z. Li, Y. Yuan, Z. Ji, P. Ma, S. Wang

ICSE 2027 · Accepted · CCF-A · First author

Optimizes LLM Agent skills for token efficiency. Across 600 skills and 3,000 tasks, reduces description and body tokens by 48% and 39%, while improving the evaluated task score from 0.722 to 0.742.

CertMoE input masking, routing-containment certification, induced sub-network certification, and final certified radius

CertMoE: Robustness Certification for Mixture-of-Experts LLMs via Routing-Aware Smoothing

Q. Wang, Y. Gao, L. Wang, R. Huang, Y. Yuan, Y. Song, S. Wang

arXiv preprint 2026 · Second author

Certifies MoE LLMs by combining expert-routing containment with prediction stability of the induced sub-network. Evaluated on 3 MoE models and 2 text-classification datasets, improving certificate reliability under adversarial perturbations while exposing the trade-off with certification coverage.

Six panoramic RGB-D viewpoints and their coverage of an indoor scene

CM-EVS: Sparse Panoramic RGB-D-Pose Data for Complete Scene Coverage

J. Liu, J. Li, J. Yu, X. Yu, et al., Y. Gao*, M. Cheng

NeurIPS 2026 · Accepted · CCF-A · * Corresponding author

A coverage-curated dataset of 36,373 RGB-D frames across 1,275 indoor scenes, connecting panoramic perception with compact, geometry-consistent 3D environments.

TriBA framework for spatial frequency and feature-domain stealth

A Triple Stealthy Backdoor: Hidden in Spatial, Frequency, and Feature Domains

Y. Gao, H. Chen, P. Sun, J. Li, Y. Yin, Z. Wang, W. Liu

IEEE TDSC 2025 · CCF-A · First author

Develops TriBA to hide backdoor triggers in the spatial, frequency, and feature domains, combining frequency-domain transformations with feature-space alignment. Evaluated on four datasets to study attack effectiveness and stealthiness.