Curriculum Vitae

Education

• M.Phil, The HongKong University of Science and Technology (Guangzhou)
  • AI thrust&INFO Hub
    • Majoring in Artificial Intelligence(AI, CGA:3.323/4)

2024-Present

• B.S, China University of Petroleum Beijing
  • Department of Mathematics, School of Arts and Sciences
    • Majoring in Statistics(STAT,GPA:3.58/4)

2020-2024

Research experience

Postgraduate(2024-Present):

LLM-Based Intelligent Companion AI Digital Humans with Consciousness and Memory

Project summary: In the digital age, addressing challenges such as prompt injection attacks, data leakage, low interpretability, and insufficient emotional engagement faced by LLM-based companion agents in fields like education and healthcare, we constructed a comprehensive framework to enhance security protection, improve reasoning capabilities, quantify uncertainty, and customize emotional support. The goal is to create a safe, reliable, and empathetic AI assistant. Responsible for the post-training and interpretability modules of the LLM.

Outcomes

• Agent Module: Adopted Qwen3-14B as the pre-trained model and performed single-GPU fine-tuning (A800 80G) on the CBT-bench dataset.
• Interpretability Module:
  • Explored potential relationships among conceptual annotations in the dataset using the Apriori algorithm for association rule learning, enhancing the interpretability of causal variables.
  • Designed an innovative framework of Latent Disentanglement-Concept Bottleneck Models (LDCBMs). Compared with previous models on the CUB, AwA2, and CelebA datasets, the concept alignment rate increased by 0.1242 ± 0.0576, and the label accuracy improved by 0.1316 ± 0.0203, achieving SOTA performance. Additionally, the entire input-concept-output process is interpretable.

Backdoor Attack Defense for Concept Bottleneck Models

Project Summary: In explainable AI (XAI), Concept Bottleneck Models (CBMs) enhance interpretability via understandable underlying concepts, but are vulnerable to concept-level backdoor attacks (hidden triggers in concepts causing undetectable misbehavior). First proposed Conceptguard defense: constructed poisoned datasets, divided data subsets, and used majority voting to mitigate data-driven backdoor impacts.

Key Contributions:

• Theoretically proved a minimum trigger size threshold, above which Conceptguard effectively defends against attacks (average backdoor success rate reduced by ~30\%).

• Led CBMs baseline and Conceptguard experiments on CUB dataset, demonstrating improved concept accuracy. Identified cluster mechanism as an unsupervised method to avoid concept-level category conflicts and enhance feature learning for better concept correlation capture.

Publications

Guarding the Gate: ConceptGuard Battles Concept-Level Backdoors in Concept Bottleneck Models

Towards more holistic interpretability: A lightweight disentangled Concept Bottleneck Model

Awards

• Sep. 2022First Prize China Undergraduate Mathematical Contest in Modeling(CUMCM)
• Jun. 2022Grand Prize(1/4983) ”Renzheng Cup” National Student Mathematical Modeling Competition
• Nov. 2021H Prize, “Shuwei Cup” International Student Mathematical Modeling Competition
• Nov. 2021Third Prize “Huajiao Cup”National University Mathematics Competition
• Feb. 2022H Prize Interdisciplinary Contest In Modeling(MCM/ICM)
• Aug. 2022Second Prize ”MathorCup” National Student Mathematical Modeling Competition
• May. 2022Second Prize ”Renzheng Cup” National Student Mathematical Modeling Competition
• Jun. 2022Second Prize China Student Computer Design Competition
• Sep. 2022Second Prize Nation University math network championship