I am a Master’s student in Data Science at ETH Zürich and a Research Assistant at the ETH AI Center. My research focuses on LLM evaluation, interpretable capability assessment, and privacy in language agents.
Previously, I studied Computer Science and Mathematics at Sorbonne University. I also worked on time-series forecasting at ESIEE Paris, supervised by Dr. You Jiang.
Email / CV / Google Scholar / GitHub / LinkedIn
I work on evaluating language models beyond aggregate benchmark scores, including privacy–utility trade-offs, ranking robustness, and interpretable skill profiles.
NeurIPS 2026 E&D Track.
Project PagearXivGitHubDataset
A diagnostic benchmark with 7,852 samples across 10 domains for evaluating the trade-off between task utility and user-defined privacy constraints in LLM agents.
NeurIPS 2026 TAE Workshop.
We study the robustness of near-tied LLM rankings across five benchmarks using item response theory and differential item functioning, with owner-disjoint evaluation folds and matched-random controls.
Preprint. Under review.
Combining item response theory with LLM-based contrastive hypothesis generation to explain question difficulty, with validation through held-out difficulty prediction and targeted question rewriting across reasoning benchmarks.
Under review.
Combining LLM-based skill discovery with neural cognitive diagnosis to construct interpretable ability profiles for 3,811 LLMs across 9,523 items from five benchmarks.
Under review.
Project PagearXivGitHubDataset
A benchmark of 720 synthetic counterfactual pairs across three domains evaluating privacy leakage and downstream utility across five output-release settings. We develop workflow-constrained privacy quantization (WCPQ) to reduce secret recovery while preserving action branches and costs across 18 frozen workflow/cost combinations.
Preprint.
I contributed translation test cases and reviewed English, Chinese, and French benchmark submissions.
Research Assistant · Zurich, Switzerland
Developing a multilingual benchmark for legal compliance in LLMs, including reproducible scenario-generation, review, and evaluation pipelines. Collaborating with legal experts and investigating applications for LLM post-training.
Research Assistant · Paris, France
Implemented Transformer-based models and PatchTST for time-series forecasting under Dr. You Jiang, adapting training pipelines and comparing model performance and computational efficiency on synthetic datasets. [code]
MSc in Data Science
BSc in Computer Science / Mathematics
Sorbonne University
Implemented Penrose, Wang, and Hat tiling algorithms in Python; optimized generation with duplicate detection and spatial pruning.
ESIEE Paris · GitHub
Compared Q-learning and SARSA across 15 custom environments. Built a map editor and environments adopted for master’s-level teaching at UPEC.
I speak Chinese, English, and French. Outside research, I enjoy playing drums and saxophone, cooking, reading, and calligraphy.
You can reach me at zqiaoyuan@ethz.ch.