CV
Email · Google Scholar · LinkedIn
Research Interests
Reinforcement learning, large language models, statistical machine learning, multi-armed bandits, and mechanism design.
Education
- Purdue University — Ph.D. in Computer Science, August 2022–May 2027 (expected). Advisor: Steve Hanneke.
- Tsinghua University — B.S. in Mathematical Science, August 2017–June 2022.
Experience
Amazon — Applied Scientist Intern
August 2026–present
- Designed and built an end-to-end product embedding and KNN retrieval pipeline to predict multiple cost and profitability (CP) components for low-volume and cold-start products, reducing prediction MAE by 15–30% across different CP components compared to the existing economics-based baseline.
- Post-trained a Transformer-based embedding model using multi-objective contrastive learning to learn task-specific product representations for similarity-based retrieval and downstream CP component prediction.
DoorDash — Machine Learning Engineer Intern
May–August 2026
- Formulated peak-time incentive allocation as an MDP to capture sequential supply and demand dynamics.
- Developed and evaluated offline reinforcement learning algorithms to improve service quality while controlling incentive cost.
- Built an end-to-end RL pipeline covering offline training and evaluation, online experimentation, and online serving.
Publications and Manuscripts
* Equal contribution; alphabetical author order is indicated separately.
PAC Learning with Bandit Feedback: Sharp Sample Complexity in the Realizable Setting
Steve Hanneke, Qinglin Meng, Shay Moran, Amirreza Shaeiri (alphabetical order)
Advances in Neural Information Processing Systems (NeurIPS 2026), 2026 · Paper
Towards Intrinsic Interpretability of Large Language Models: A Survey of Design Principles and Architectures
Yutong Gao*, Qinglin Meng*, Yuan Zhou, Liangming Pan
Annual Meeting of the Association for Computational Linguistics (ACL 2026 Main Conference, Oral), 2026 · Paper
An Optimal Sauer Lemma Over k-ary Alphabets
Steve Hanneke, Qinglin Meng, Shay Moran, Amirreza Shaeiri (alphabetical order)
arXiv preprint arXiv:2604.12952, 2026 · Paper
Discover, Explain, and Repair: Adaptive Testing of LLM Failures in Medical Safety
Y. Zhou, S. Guo, Qinglin Meng, X. Zhang
Manuscript under submission, 2026, 2026
Representation Preserving Multiclass Agnostic to Realizable Reduction
Steve Hanneke, Qinglin Meng, Amirreza Shaeiri (alphabetical order)
International Conference on Machine Learning (ICML 2025), 2025 · Paper
Learning Dynamic Mechanisms in Unknown Environments: A Reinforcement Learning Approach
Shuang Qiu*, Boxiang Lyu*, Qinglin Meng*, Zhaoran Wang, Zhuoran Yang, Michael I. Jordan
Journal of Machine Learning Research (JMLR), 2024 · Paper
Research and Working Projects
Reward-Free RL for Mechanism Design and Dynamic Pricing Optimization
Advisors: Michael I. Jordan, Zhaoran Wang, and Zhuoran Yang
- Studied online mechanism design with linear function approximation.
- Developed reward-free exploration over a rich policy space to estimate prices.
- Designed optimistic/pessimistic Q-function approximation algorithms to recover mechanisms over repeated interactions while balancing seller and buyer profit.
Adaptive Training Data Allocation for LLM Safety via Non-stationary Bandits
Advisor: Xiangyu Zhang
- Built a data pipeline with OpenFDA drug data, synthetic patient profiles, and generated scenarios for LLM safety evaluation and training.
- Formulated adaptive post-training data selection as a non-stationary multi-armed bandit problem.
- Implemented sliding-window Thompson Sampling to prioritize failure-prone drugs, reducing mean and worst-case error relative to uniform sampling and baseline curricula on LLaMA-3 and Qwen-2.5.
Near-Optimal Bandit Learning for Multiclass Classification
Advisors: Steve Hanneke and Shay Moran
- Proved a sharp Sauer lemma based on the list-DS dimension.
- Developed a chain-of-list learning algorithm achieving near-optimal sample complexity for multiclass learning with bandit feedback.
- Improved an existing list-learning sample complexity bound by a factor of K⁵.
A Survey of Architecture-Level Design for Interpretable LLMs
Advisor: Liangming Pan
- Analyzed architecture-level interpretability techniques, including codebook features, sparse mixture-of-experts architectures, and sparse autoencoder based steering.
Reduction for Multiclass Learning
Advisor: Steve Hanneke
- Studied multiclass reductions with potentially infinite label spaces.
- Developed a representation preserving agnostic-to-realizable reduction by applying a realizable learner to realizable subsets of the input sample.
Academic Service
Reviewer: Transactions on Machine Learning Research (TMLR), COLM 2026, NeurIPS 2026, and ACL ARR 2026.
Teaching
Teaching Assistant, Purdue University
- CS 177: Python Programming, Fall 2023–Fall 2025.
- CS 573: Data Mining, Spring 2023.
Skills
- Programming: Python, C++, R, MATLAB, SQL.
- Software and systems: PyTorch, TensorFlow, Git, Databricks, Snowflake.
