Hanqing Wang

Hanqing Wang

Ph.D. Student at Shanghai University of Finance and Economics (SUFE)

I am a 4th-year Ph.D. student at SUFE advised by Prof. Yun Chen, and previously a research intern at THUNLP & QY Lab supervised by Shuo Wang. My work has appeared at ACL, EMNLP, NAACL, and NeurIPS. I was recently selected as SUFE Academic Star (5 graduate students per year university-wide).

Open to industry research positions · Expected to graduate July 2027
Research Interests

My research is unified by one question: how do we transfer capabilities into and across large language models efficiently? I started with zero-shot style transfer via lightweight adapters (StyleBART), then built parameter-efficient fine-tuning methods (LoRA-Flow, MiLoRA) and training-free delta compression (Delta-CoMe), and most recently latent-vulnerability discovery (SAGE) plus Python-pivot SQL generation (Pi-SQL) for Text-to-SQL. Increasingly I'm asking whether the same lens — modular, efficient capability — applies to agent skill design.

News
  • Apr 2026One paper accepted by Findings of ACL 2026!
  • Nov 2025One paper accepted by Findings of EMNLP 2025!
  • Mar 2025One paper accepted by NAACL 2025!
  • Jan 2025Completed research internship at THUNLP & QY Lab.
  • Sep 2024Awarded First-Class Academic Scholarship at SUFE.
  • Sep 2024One paper accepted by NeurIPS 2024!
  • May 2024One paper accepted by ACL 2024!
  • Sep 2023Started research internship at THUNLP & QY Lab.
  • Oct 2023One paper accepted by Findings of EMNLP 2023!
  • Jun 2022Obtained bachelor's degree from SUFE.
Selected Publications
SAGE method overview ACL 2026
Beyond Static Rules: Automated Discovery of Latent Vulnerabilities in Text-to-SQL
Hanqing Wang*, Yongdong Chi*, Jian Yang, Lei Yang, Jiehui Zhao, Yun Chen, Guanhua Chen
ACL 2026 Findings · CCF-A
TL;DR: SAGE (Systematic Automated Guided Exploration) automates discovery of latent failure patterns in LLM Text-to-SQL — generating vulnerability hypotheses with a continuously evolving Vulnerability Codex of perturbations.
Pi-SQL pivot programming language pipeline EMNLP 2025
Pi-SQL: Enhancing Text-to-SQL with Fine-Grained Guidance from Pivot Programming Languages
Yongdong Chi*, Hanqing Wang*, Yun Chen, Yan Yang, Jian Yang, Zonghan Yang, Xiao Yan, Guanhua Chen
Paper · Code · EMNLP 2025 Findings · CCF-B
TL;DR: Uses Python as a pivot intermediate representation to give Text-to-SQL fine-grained structural guidance — leveraging LLMs' stronger Python understanding to bootstrap better SQL generation.
MiLoRA minor singular component initialization NAACL 2025
MiLoRA: Harnessing Minor Singular Components for Parameter-Efficient LLM Finetuning
Hanqing Wang, Yixia Li, Shuo Wang, Guanhua Chen, Yun Chen
Paper · Code · NAACL 2025 Main · CCF-B
TL;DR: Initializes LoRA adapters using the minor singular components of pre-trained weights — reducing interference with the base model's existing knowledge while still enabling effective task adaptation.
Delta-CoMe mixed-precision delta compression NeurIPS 2024
Delta-CoMe: Training-Free Delta-Compression with Mixed-Precision for Large Language Models
Bowen Ping, Shuo Wang, Hanqing Wang, Xu Han, Yuzhuang Xu, Yukun Yan, Yun Chen, Baobao Chang, Zhiyuan Liu, Maosong Sun
Paper · Code · NeurIPS 2024 · CCF-A
TL;DR: Compresses the delta (fine-tuned minus base weights) of an aligned LLM with mixed-precision quantization — recovering near-full performance with a fraction of storage and zero retraining.
LoRA-Flow dynamic LoRA fusion ACL 2024
LoRA-Flow: Dynamic LoRA Fusion for Large Language Models in Generative Tasks
Hanqing Wang*, Bowen Ping*, Shuo Wang, Xu Han, Yun Chen, Zhiyuan Liu, Maosong Sun
Paper · Code · ACL 2024 Main · CCF-A
TL;DR: Replaces fixed LoRA fusion weights with a per-token dynamic gate — letting multiple task-specific LoRA adapters be combined adaptively during generation rather than averaged with a static recipe.
StyleBART style adapter architecture EMNLP 2023
StyleBART: Decorate Pretrained Model with Style Adapters for Unsupervised Stylistic Headline Generation
Hanqing Wang*, Yajing Luo*, Boya Xiong, Guanhua Chen, Yun Chen
Paper · EMNLP 2023 Findings · CCF-B
TL;DR: Adds lightweight style adapters on top of frozen BART to generate stylistic news headlines without paired data — decoupling content faithfulness from style transfer.