cv

Basics

Name Zeyu (Steven) Zhang
Label PhD Student
Email zeyuzhang2028@u.northwestern.edu
Url https://ZeyuZhang1901.github.io
Summary PhD student at Northwestern University working on trustworthy and reliable LLMs: temporal knowledge leakage in LLM backtesting, black-box interpretability, and RL-based post-training.

Work

  • 2025.09 - Present
    Graduate Researcher - Trustworthy LLM Backtesting (TEMPO, TimeSPEC & Leakage-Adjusted Evaluation)
    Northwestern University
    Detecting, attributing, quantifying, and eliminating temporal knowledge leakage in LLM backtesting. Three first-author papers: Findings of EMNLP 2026, plus two under review or in submission (NeurIPS 2026, TMLR).
    • Formalized temporal knowledge leakage: models exploit post-cutoff pre-training knowledge on historical prediction tasks, inflating reported accuracy and invalidating evaluation
    • Proved the leakage inflation is non-identifiable from passive backtest scores (recency alone reproduces the leakage signature: four flagship models fail the naive pre/post check on questions they cannot have memorized) and derived a concentration law: inflation = honest uncertainty x extraction factor, so partial memorization is disproportionately rewarded
    • Showed one external reference restores measurement: a known cutoff (regression discontinuity) or a matched clean control (difference-in-differences) yields a leakage-adjusted score with confidence intervals; validated by planting leakage in twin models (dose recovery, null on the clean twin) and deployed on frontier models, detecting one cutoff-localized signature and clearing five models whose advantages were recency alone
    • Introduced Shapley-DCLR, a claim-level Shapley-weighted metric quantifying what fraction of decision-critical reasoning is contaminated
    • Designed TimeSPEC, an inference-time architecture interleaving temporally filtered retrieval with claim-level supervision, requiring no retraining
    • Developed TEMPO, GRPO-based post-training with a two-mode reward; proved monotonic leakage decrease and convergence to the leak-free optimum
    • Reduced leakage from 2-13% to 0.6-3.7% across three tasks and two models while improving task performance by 6-13% where valid signals exist
  • 2024.06 - 2025.01
    Graduate Researcher - Factual Memorization in LLMs
    Northwestern University
    Analysis of LLM memorization behaviors under SFT and DPO. Qualifying examination paper.
    • Analyzed LLM memorization under SFT and DPO fine-tuning, reproducing and extending Stanford's FineTuneBench
    • Proposed a formal distinction between passive memorization (exposure) and positive memorization (direct QA supervision)
    • Showed injected future-dated facts do not generalize beyond training, and that a temporal-consistency system prompt mitigates the resulting overfitting
  • 2024.01 - 2025.01
    Graduate Researcher - Black-Box LLM Interpretability (LAMP)
    Northwestern University
    Co-developed LAMP for interpreting black-box LLMs. Accepted as Spotlight at AISTATS 2026.
    • Co-developed LAMP, which treats an LLM's self-reported explanations as a coordinate system and fits locally linear surrogate decision surfaces
    • Designed perturbation-based probing that extracts decision surfaces without gradients, logits, or internal activations, enabling audits of proprietary LLMs
    • Ran experiments across sentiment analysis, controversial-topic detection, and safety-prompt auditing; surfaces align with human and expert judgments
  • 2022.05 - 2023.12
    Research Collaborator - Unified Off-Policy Learning to Rank (CUOLR)
    Princeton University (remote)
    Mentored by Prof. Mengdi Wang and Prof. Huazheng Wang. Published at NeurIPS 2023.
    • Unified ranking under general stochastic click models as a Markov Decision Process, enabling principled offline RL for off-policy learning to rank
    • Proposed CUOLR, a click-model-agnostic algorithm requiring no explicit debiasing or prior knowledge of the click process
    • Achieved state-of-the-art performance on large-scale benchmarks with consistent robustness across heterogeneous click models

Teaching

Education

  • 2023.09 - 2028.06

    Evanston, IL, USA

    Doctor of Philosophy
    Northwestern University
    Statistics and Data Science
    • GPA: 3.95/4.00
    • Advisor: Prof. Bradly C. Stadie
    • Expected graduation: June 2028
  • 2020.09 - 2023.06

    Hefei, Anhui, P.R.China

    Certificate (Minor)
    University of Science and Technology of China (USTC)
    Artificial Intelligence
    • Talent Program in Artificial Intelligence
    • Outstanding Undergraduate Honorary Rank (top 5%)
  • 2019.09 - 2023.06

    Hefei, Anhui, P.R.China

    Bachelor of Engineering
    University of Science and Technology of China (USTC)
    Electronic Information Engineering
    • GPA: 3.93/4.30, Rank: 5/213 in School of Information Science and Technology
    • China National Scholarship, Ministry of Education of the PRC (top 1%)
    • Wang Xiaomo Talent Program in Cyber Science and Technology

Awards

Publications

Skills

Programming
Python
PyTorch
Hugging Face Transformers
R
C/C++
Bash
Git
LaTeX
ML / LLM
RL post-training (GRPO, DPO, RLHF)
LLM evaluation & benchmarking
Retrieval-augmented pipelines
Offline reinforcement learning
API-based LLM systems (OpenAI platform)

Languages

Chinese
Native speaker
English
Fluent