Hi! I’m Leonard (Shang’ao) Li, an undergraduate student in the School of Computer Science at Nanjing University.
My academic focus lies in AI reasoning, machine learning, and robust, verifiable intelligent systems. I have completed nearly all AI-related and core CS courses with strong performance, forming a foundation that supports both theoretical and system-level work.
What I Work On
My research centers on how to build learning systems that can reason, adapt, and maintain coherent internal representations. Recently, I have been:
- Reasoning-enhanced reward models for LLMs — designing pipelines that combine reject sampling, SFT, and RL to improve preference alignment under Dr. Zhen Han. This includes engineering large-scale training systems on limited hardware and stabilizing reward-model behavior in challenging reasoning tasks.
- My current research centers on neuro-symbolic reasoning and how to design learning systems that maintain coherent internal structure while adapting to new information. In the KRistal Group at NJU, I work on structured inference and representation learning, focusing on how models can incorporate symbolic constraints while preserving the flexibility of neural methods. This project aims to better understand how reasoning, learning, and knowledge representation interact in scalable architectures—without relying on any proprietary techniques.
- Interactive theorem proving with LLMs — at the ScaleML Lab (UIUC), integrating Lean4 & mathlib with LLMs for mathematical reasoning. We prototyped a bidirectional interface between model outputs and formal verification, highlighting the complementary strengths of symbolic and neural approaches.
My broader research experience spans reasoning-centric WebAgents (LMU Munich), benchmarking intelligent agents at Huawei 2012 Labs, quantum memory architectures (QUEST Lab @ NC State), and adversarial ML robustness at COSEC Research Group @ NJU.
Across these projects, a common thread emerged: intelligent systems need principled reasoning mechanisms and stable representation learning, not only larger scales.
Research Vision
I am broadly interested in the foundations of reasoning for AI.
My goals include:
- Designing hybrid neuro-symbolic architectures that balance logical precision with neural flexibility.
- Building adaptive knowledge-representation systems that can reorganize themselves as they learn.
- Understanding how agents can form hypotheses, refine internal models, and learn through interaction while maintaining coherence.
- Developing scalable, efficient training pipelines for reasoning-heavy models.
Ultimately, I hope to contribute to the scientific foundations of general-purpose reasoning—systems that can generalize reliably, learn continually, and support interpretable, verifiable decision-making.
Beyond Research
I enjoy hands-on systems work: distributed training, HPC cluster environments, Linux tooling, and re-implementing ML/OS components from scratch to understand them deeply. These projects keep me grounded in both engineering constraints and theoretical goals.
Let’s Connect
You can learn more about my experience through the CV / Publications / Projects section. I’m always happy to hear from collaborators, mentors, or anyone who wants to chat about AI reasoning, neuro-symbolic methods, or systems. Feel free to reach out via email!