INDUSTRY
SAP
Part-time iXp Research Intern
Remote, USA
Xiangqi (Shawn) Wang
I am a PhD student in Computer Science and Engineering at the University of Notre Dame, advised by Prof. Xiangliang Zhang. My research focuses on continual learning for language models.
Continual learning for language models
Language models must adapt to new experience without losing reliable reasoning or aligned behavior. My research studies this problem through test-time training, reinforcement learning, self-distillation, and preference optimization: how to select reasoning actions, enable adaptation, assess reasoning quality, limit behavioral drift, and balance competing objectives in human–model interaction.
A sequential learning perspective
The schematic below relates action selection and learning gains to future learning value, subject to a constraint on protected utility.
Notation and interpretation
In this schematic, θt denotes model parameters; at, a selected action; gt, an immediate learning gain; and Pt, a long-term learning value. The factor γΔt discounts future value. UH denotes protected utility, θ0 the reference model, and C an allowable utility decrease. These are conceptual definitions for the diagram; their operational meaning depends on the individual research setting.
| Component | Research question and approach | Related work |
|---|---|---|
| Action selection at | How should a model select reasoning actions for different problems?Reinforcement learning for an adaptive reasoning controller. | AdaReasonerNeurIPS 2025 · Spotlight |
| Test-time adaptation gt: reachability | How can initialization prepare a model to learn effectively at test time?Meta-learned initialization for test-time training. | PrimeTTTEMNLP 2026 · Main |
| Reasoning coherence gt: validity | How can policy optimization account for reasoning coherence alongside answer correctness?Coherence-aware reinforcement learning. | CE-POEMNLP 2026 · Main |
| Behavioral preservation Pt+1 | How can self-distillation improve reasoning while limiting behavioral drift?Locality-constrained self-distillation. | LGSDICLR 2027 · Submission |
| Preference alignment UH | How can preference optimization balance truthfulness, anti-sycophancy, empathy, and creativity?Dual preference optimization with multiple behavioral objectives. | Dignified PeersFindings of EMNLP 2026 |
First-author publications
Continual learning, reasoning, online learning, and AI systems.
First author on every work below
-
Right Answers for the Right Reasons: Coherence-Enhanced Policy Optimization for Large Language Models
Preprint title: Causally-Enhanced Reinforcement Policy Optimization (arXiv:2509.23095).
-
From Evasive Servants to Dignified Peers via Dual Preference Optimization
Preprint title: Dual Optimal: Make Your LLM Peer-like with Dignity (arXiv:2604.00979).
-
Locality-Guided Self-Distillation
-
SAGE: A Unified Algebra and Self-Adaptive Execution for AI Functions in SQL
No papers match that search. Try a different topic or .
Co-authored publications
Evaluation, agents, reasoning, and trustworthy AI.
Experience and education
View full CVINDUSTRY
IBM Research
Research Intern · Yorktown Heights, NY
Developed SAGE, a unified algebra and adaptive execution framework for AI functions in SQL, optimizing model selection and query execution for quality, cost, and latency.
EDUCATION
University of Notre Dame
PhD Student · Computer Science and Engineering
Advisor: Prof. Xiangliang Zhang
South Bend, Indiana
EDUCATION
University of Science and Technology of China
B.Sc. in Computer Science · First-Class Honors
School of the Gifted Young · Hefei, China
Prior mentorshipQi Liu ↗Huazheng Wang ↗Jian Kang ↗Zhuoran Yang ↗
Research software
Implementations and research systems.
01 / RL INFRASTRUCTURE
CausalRL
VERL extended with Jacobian-based causal scores and a parallel reward backend for policy optimization.
02 / ADAPTIVE INFERENCE
AdaReasoner
A modular reasoning controller with reproducible training and inference pipelines. The official implementation.
03 / ONLINE LEARNING
Fair Influence Maximization
Fairness-aware combinatorial bandits with scalable optimization and online evaluation on public social graphs.
04 / AI FOR SCIENCE
Knowledge Network of Wildlife
A searchable wildlife knowledge interface and graph-RAG chatbot for species, habitat, and policy queries.
Tools
PyTorch · Transformers · VERL · LLaMA-Factory · SQL · Git · Docker
Teaching, service, and research support
Mentorship & teaching
I guide student-led projects from problem selection and method design through experiments and writing.
Teaching assistant for Graduate Machine Learning at Notre Dame, and Computer Programming A and Data Analysis and Practice at USTC.
Academic service
Reviewer for NeurIPS, ICML, ICLR, KDD, and ACL Rolling Review (2025–2026), WWW (2025), and KDD (2024).
Journal reviewing: TMLR (2025) and IEEE Transactions on Computational Social Systems. Additional service: IEEE NASC Competition.
Research support
NSF DISCOVER ACCESS computing allocation and OpenAI Researcher Access API credits (2025), plus Anthropic research computing support.
Contributions to data and knowledge integration through NSF Proto-OKN, with GPU research infrastructure through NSF ACCESS.
Honors & community participation
Honors
- Chinese Mathematics Competition · Provincial First Prize, 2022
- ICPC China Regional · Competition prize, 2021
- USTC Outstanding Student · Annual award
- SGY 87–00 Scholarship · Nominee
Talks & participation
NSF Proto-OKN Annual Meeting; SASC Symposium; ACM WSDM Tutorial; EMNLP 2024; Third Reinforcement Learning Conference (remote).