About
👋 I am a Ph.D. student in Computer Science (Aug. 2025 – present) at the University of Illinois Chicago (UIC), advised by Prof. Philip S. Yu.
I received my M.Eng. from the Shanghai Advanced Research Institute, University of Chinese Academy of Sciences (UCAS), under the supervision of Prof. Jun Huang, and my B.Eng. from Soochow University. I am grateful to Prof. Lu Cheng for our past collaboration.
My research interests focus on two areas: Trustworthy LLMs, Agents & Post-Training (agent skill and security, LLM safety alignment, and RL post-training) and Graph Learning & Graph Foundation Models (graph learning, foundation models, and their connection to LLMs and agents).
Trustworthy AI
LLM Safety & Alignment
LLM Agents
RL Post-Training
Graph Foundation Models
Graph Learning
Education
Ph.D. in Computer Science
University of Illinois Chicago (UIC), Chicago, IL, USA
M.Eng. in Electronic & Information Engineering
Shanghai Advanced Research Institute, University of Chinese Academy of Sciences, Shanghai, China
Sept. 2022 – June 2025 | Advisor: Prof. Jun Huang
B.Eng. in Computer Science & Technology
Soochow University, Suzhou, China
Sept. 2018 – June 2022
Publications
ICLR 2027Submitted
Compile Locally, Query Globally: Learning Reusable Graph Programs
H. Hao, W. Zhang, J. Yang, S. Fan, W. Liu, Z. Shen, L. Sun, L. Wei
Submitted to ICLR 2027 (2026)
bioRxivPreprint
ConfDock: Atom-specific Uncertainty Quantification for Molecular Docking via Conformal Prediction
H. Hao, N. Elhendawy, Y. Wang, L. Cheng
bioRxiv preprint (2026)
KDD 2026Accepted
SafeCRS: Personalized Safety Alignment for LLM-Based Conversational Recommender Systems
H. Hao, Y. Xu, X. Li, Y. Ge, L. Cheng
Accepted at ACM SIGKDD 2026, Research Track | arXiv:2603.03536 (2026)
AAAI 2027Submitted
Poise: Position-Aware One-Instruction Skill Injection for Silent Execution on LLM Agents
H. Hao, D. Min, Z. Zhang, Y. Zhang, M. Xu, Y. Ge, L. Cheng
Submitted to AAAI 2027 | arXiv:2606.07943 (2026)
ICLR 2027Submitted
Beyond Math and Code: Lightweight Corpus-Grounded Process Rewards for Factual Question Answering
S. Fan†, H. Hao†, D. Min†, W. Liu†, H. Zhang, L. Wei, H. P. Zou, C. Guo, J. Yang, H. Bao, Z. Liu, L. Cheng, P. S. Yu
†Equal contribution (co-first authors).
Submitted to ICLR 2027 | arXiv:2605.29648 (2026)
ESWA 2026Published
Progressive Alternating Attribute-Structure Optimization for Multiplex Heterogeneous Graphs
H. Hao, J. Huang, S. Rao
Expert Systems with Applications, 312, 131495 (2026)
NCA 2025Published
Heterogeneous Graph Multi-level Semantics Extraction for Node Classification
H. Hao, J. Huang, S. Rao
Neural Computing & Applications, 37, 11821–11841 (2025)
Research Projects
Trustworthy LLMs, Agents & Post-Training
SafeCRS: Personalized Safety Alignment for LLM-Based Conversational Recommender Systems
Co-first author · ACM SIGKDD 2026
Designed a personalized safety-alignment framework for LLM conversational recommenders that enforces user-specific safety constraints while preserving recommendation quality.
Poise: Position-Aware One-Instruction Skill Injection for Silent Execution on LLM Agents
Co-first author (lead) · Submitted to AAAI 2027
Designed a position-aware, one-instruction skill-poisoning attack. On Skill-Inject with Codex + GPT-5.2, it achieves 89.3% attack success while preserving the legitimate task, 28.0 percentage points above random body placement.
CorVer: Lightweight Corpus-Grounded Process Rewards for Factual QA
Co-first author · Submitted to ICLR 2027
Co-developed sentence-level factual rewards from corpus co-occurrence counts for RL post-training. CorVer achieves the highest accuracy in 17 of 20 baseline-comparison settings and reduces mean complete training time by 5.5–10.4× on Qwen3-4B/8B relative to four factuality-RL baselines. Inference requires no corpus access.
Graph Learning & Graph Foundation Models
CoGraC: Compile Locally, Query Globally
First author · Submitted to ICLR 2027
Developed an end-to-end graph compiler that learns reusable local programs with boundary interactions and recoverable internal-node responses. CoGraC improves compositional link prediction by 7.1–8.2 AP points over full-graph GPR64 and accelerates local structural updates on Cora by 2.4–5.0× with identical predictions.
Graph Representation Learning on Heterogeneous Information Networks
First author · ESWA 2026; NCA 2025
Designed progressive alternating attribute-structure optimization for multiplex heterogeneous graphs, improving robustness under missing attributes and noisy edges (ESWA 2026); developed a multi-level semantics extraction method for node classification (NCA 2025).
Technical Skills
LLM & Agents: LLM evaluation, LLM-as-a-Judge, agent skill, safety alignment, RL post-training, LLM training & inference (HuggingFace, vLLM, Unsloth, TRL), agent systems (Claude Code, Codex, OpenClaw)
Graph & ML: PyTorch, graph neural networks, graph foundation models, uncertainty quantification, conformal prediction, meta-learning, domain adaptation
Programming & Tools: Python, Java, SQL, Git, Docker, Daytona, Linux, LaTeX, TensorBoard, databases (SQL, MongoDB)
Invention Patents
J. Huang & H. Hao. “Node Classification Method, Device, Terminal, and Medium Based on Multi-level Semantic Representation of Heterogeneous Knowledge Graphs.” Approved.
Awards & Honors
Multiple “Three Good Student” awards at the University of Chinese Academy of Sciences and Soochow University for outstanding academic and overall performance.