CV

You can find my full CV here: Jiaming Zhang

Education

• B.S. in Statistics, Renmin University of China, 2021 - 2025

Internships

• 7/2025 - Now: Remote Intern
  • PRADA Lab, King Abdullah University of Science and Technology (KAUST)
  • Duties includes: Applied Differential Privacy (DP) techniques to enhance data privacy in machine learning models. Investigated the development of trustworthy Large Language Models (LLMs) with a focus on adversarial robustness theory and adversarial training methods.
  • Supervisor: Di Wang
• 10/2024 - 6/2025: Intern
  • Intelligent Game and Decision Laboratory
  • Duties includes: Developed and implemented anomaly detection algorithms for spacecraft telemetry data using advanced generative models (LSTM-VAE, Transformer-AE) and kernel density estimation.
  • Supervisor: Xiaohu Zheng

Publications

Efficient Forward-Only Data Valuation for Pretrained LLMs and VLMs Submitted to AAAI 2026

• Wenlong Deng, Jiaming Zhang, Qi Zeng, Christos Thrampoulidis, Boying Gong, Xiaoxiao Li
• We propose For-Value, a forward-only data valuation framework for LLMs and VLMs. It estimates per-sample influence with a single forward pass using hidden representation alignment and prediction errors, avoiding gradients or retraining. For-Value achieves scalable influence estimation and outperforms gradient-based baselines in fine-tuning and mislabeled data detection.

Towards User-level Private Reinforcement Learning with Human Feedback CoLM 2025

• Jiaming Zhang*, Mingxi Lei*, Meng Ding, Mengdi Li, Zihang Xiang, Difei Xu, Jinhui Xu, Di Wang
• Developed AUP-RLHF, a novel framework integrating user-level label differential privacy (DP) into Reinforcement Learning with Human Feedback (RLHF) to protect user preference privacy. Demonstrated that traditional Random Response algorithms perform suboptimally in user-level settings. Proposed a lower bound for user-level DP-RLHF and designed the AUP-RLHF algorithm, achieving ($\varepsilon$,$\delta$)-user-level privacy with improved estimation error.

A Case-Based Reasoning Framework for Multi-Hazard Assessment of Earthquake Impacts ICCBR2024

• Yiding Dou*, Jiaming Zhang*, Yuxin Li, Ruyi Qi, Zimeng Yuan, Yanbing Bai, Erick Mas, Shunichi Koshimura
• Developed an enhanced causal graph Bayesian network integrated with a case-based reasoning framework for multi-hazard impact assessment. This method shows robust adaptability to noisy data and significantly improves accuracy compared to prior models (e.g., USGS). Applied to analyze major earthquakes in Turkey (2023) and Japan (2024), it offers a high-precision, scalable, and unsupervised learning approach for earthquake damage estimation.

Teaching

• 09/2024-12/2024, Served as a Teaching Assistant for Regression Analysis, Renmin University of China

Skills

• Programing: C/C++, Python(PyTorch, Tensorflow, Scikit-learn, etc.), SQL, MATLAB, R, GitHub, LeTaX
• Languages: CET6