Self-improving agents
Agents that learn from successful and failed trajectories, refine reusable skills, and improve how they approach repeated tasks.
Guangya (Wayne) Wan
Ph.D. Candidate in Data Science · University of Virginia
I am a Ph.D. candidate at the University of Virginia, advised by Prof. Sheng Li. My research focuses on self-improving AI, long-horizon agents, efficient reasoning, and trustworthy AI.
News
SkillForge will appear at EMNLP.
COMPASS was accepted to ACL.
BEACON will appear in Findings of EMNLP.
Research
Research Interests
My work connects agent design, test-time reasoning, and post-training to make capable AI systems more reliable and efficient over long horizons.
Efficient reasoning
Adaptive test-time compute, verification, and sampling strategies that spend inference resources only when they are useful.
Trustworthy AI
Methods and evaluations for reliable, fair, and empathetic AI across language, healthcare, and real-world decision settings.
Experience
Research & Industry Experience
Research internship
Waymo
- Built an autonomous AI research agent that searches literature, proposes code changes, runs training, analyzes results, and iterates on motion-forecasting models.
- Integrated the agent with an internal model-development pipeline for training and evaluation on autonomous-driving data.
- Helped design GRPO training for driving-action prediction, including reward signals and an image-to-text data-cleaning pipeline.
Student researcher
Google Cloud — Vertex AI
- Developed production-level code for DeepResearch Agent.
- Led SkillForge, a training-free framework for distilling validated procedural skills from successful and failed agent runs.
- Led COMPASS, which improves long-horizon reasoning by maintaining and evolving an agent’s context as work unfolds.
Graduate research assistant
University of Virginia
- Designed BEACON, a Bayesian optimal-stopping approach for efficient reward-guided LLM sampling.
- Proposed Reasoning-Aware Self-Consistency and Derailer-Rerailer for efficient, adaptive reasoning verification.
- Co-developed WiseMind, a knowledge-guided multi-agent framework for psychiatric diagnosis.
Publications
Selected Publications
2026
EMNLP 2026 · To appear
SkillForge: Learning from Success–Failure Boundaries for Consistent Long-Horizon Agents
G. Wan, H. Wang, X. Ling, M. Ling, X. Ren, S. Li
2026
ACL 2026
COMPASS: Enhancing Agent Long-Horizon Reasoning with Evolving Context
G. Wan, M. Ling, X. Ren, R. Han, S. Li, Z. Zhang
Paper ↗2026
Findings of EMNLP 2026 · To appear
BEACON: Bayesian Optimal Stopping for Efficient LLM Sampling
G. Wan, Z.S. Xu, S. Zorc, M. Baucells, M. Hu, H. Wang, S. Li
Paper ↗2025
NAACL 2025
Reasoning-Aware Self-Consistency: Leveraging Reasoning Paths for Efficient LLM Sampling
G. Wan, Y. Wu, J. Chen, S. Li
Paper ↗2025
Findings of ACL 2025
Derailer-Rerailer: Adaptive Verification for Efficient and Reliable Language Model Reasoning
G. Wan, Y. Wu, H. Wang, S. Zhao, J. Chen, S. Li
Paper ↗2026
npj Digital Medicine
WiseMind: A Knowledge-Guided Multi-Agent Framework for Accurate and Empathetic Psychiatric Diagnosis
Y. Wu*, G. Wan*, J. Li, S. Zhao, L. Ma, T. Ye, I. Pop, Y. Zhang, J. Chen
Paper ↗* Equal contribution.
Background
Education, Service & Skills
My background spans data science, biostatistics, computer science, and applied machine learning.
Education
University of VirginiaAug 2023 — Dec 2027 (expected)
Ph.D. in Data Science · Advisor: Prof. Sheng Li
Harvard UniversityAug 2021 — May 2023
M.S. in Biostatistics · GPA: 3.90/4.0
University of Illinois Urbana-ChampaignSep 2017 — Dec 2020
B.S. in Statistics, Magna Cum Laude · GPA: 3.94/4.0
Academic service
Conference reviewer
NeurIPS, ICLR, ICML, ACL, COLM, AISTATS
Journal reviewer
IEEE Computational Intelligence Magazine
Teaching assistant
UVA and Harvard
Technical toolkit
PyTorch, JAX, Hugging Face, TRL, DeepSpeed, FSDP, VeRL, vLLM, SGLang, Google ADK, LangGraph, DSPy, GCP, AWS, Docker, Kubernetes, Slurm, Python, and SQL.
Contact
Get in touch
I welcome conversations about research and potential collaborations.
wxr9et@virginia.edu