Currently, I am a first-year Ph.D. student in Computer Science at The University of Hong Kong (HKU), advised by Prof. Zuming Jiang. Before that, I received my B.E. in Computer Science from the University of Science and Technology of China (USTC).
My research focuses on trustworthy and controllable AI systems, particularly the safety and interpretability of language models, LLM post-training, and multi-agent systems. My work has been accepted at leading international AI conferences, such as ICML, ACL, AAAI, KDD, NeurIPS, and ICLR, and has received 1,000+ citations.
Trustworthy AI · LLM Post-training · Autonomous Agents
* Equal contribution
Tracks how refusal circuits emerge and evolve during alignment, then confines safety updates to those circuits to improve robustness, reduce over-refusal, and preserve general capabilities.
Paper
Uses reinforcement learning to identify sparse circuits behind reasoning and medical QA failures, then repairs only those components while preserving the model’s broader abilities.
Paper* Equal contribution
Introduces a universal method to attribute safety behavior to model circuits, enabling consistent diagnosis and targeted analysis across different language models.
Co-first authorCCF-A
Paper
Uncovers hidden backdoor triggers and traces their effects to responsible circuits, providing an interpretable route to diagnose and control backdoor behavior.
Co-first authorCCF-A
Paper
Maps the threats facing trustworthy LLM agents and organizes practical countermeasures across data, training, deployment, interaction, and governance.
Co-first author200+ citations
Paper
Studies how communication topology shapes safety and attack propagation in multi-agent systems, exposing risks that behavior-only evaluations miss.
Co-first author~100 citations
Paper
Aggregates known backdoors into a unified collapse strategy that suppresses unknown threats while retaining the language model’s useful capabilities.
Co-first authorCCF-A
Paper
Discovers latent acoustic triggers that implant and expose backdoors in audio-LLM alignment, connecting hidden signals to measurable safety failures.
Co-first authorCCF-A
Paper
Uses topology-aware diagnosis and treatment to detect security weaknesses in LLM-based multi-agent systems and reduce risks through structural interventions.
ContributorCCF-A100+ citations
Paper
Builds hierarchical memory graphs to trace how queries, insights, and interactions flow through multi-agent systems over time.
ContributorCCF-A~200 citations
Paper
Learns adaptive communication topologies with graph neural networks, helping multi-agent systems coordinate more efficiently under changing tasks and constraints.
ContributorCCF-A200+ citations
Paper
Decomposes competing objectives into orthogonal subspaces so alignment can improve multiple goals while reducing destructive interference between updates.
ContributorCCF-A
Paper
Shows that multi-agent systems can reason more reliably by asking targeted questions before acting, reducing avoidable errors and coordination failures.
ContributorCCF-A
PaperSee the complete and up-to-date publication list on Google Scholar.
