PhD Student at UCI
Hi 👋! I am Yuangang Li, a PhD student at the University of California, Irvine (UCI), where I am fortunate to be advised by Prof. Cristina (Crista) Lopes. Together with Prof. Alberto Krone-Martins, we are developing new model architectures for reasoning in physics and astronomy.
Before my PhD study at UCI, I spent wonderful years at the University of Southern California, completing my Master’s degree and working with Prof. Yue Zhao. During that time, I also had the pleasure of collaborating with Prof. Xiyang Hu and Prof. Jiechao Gao.
My research interests are centered around AI agents evaluation and optimization, Trustworthy AI, AI auditing (programming-language verification for AI), and new model architectures.
EMNLP ×2 · NeurIPS E&D ×2 · ACL, etc.
More detailsHide detailsAAAI ×2 · ACL, etc.
More detailsHide detailsarXiv
More detailsHide detailsNature AI · ACM MM · IEEE BIBM · ICDM
More detailsHide detailsAgents are usually scored by whether the final output passes a test, which says little about how they got there or how they compare across tasks. I work on evaluation that looks closer. For coding, CodeRQ-Bench and the VERA evaluator score the reasoning behind a solution rather than only its result, and Spec-Harness checks whether a formal specification that passes a verifier actually constrains the code. Across the field, Harbor Adapters brings more than 80 agentic benchmarks behind one interface, and Harbor-Index distills them into 82 hard tasks on which no tested model–harness pair clears 30%; Agents’ Last Exam adds over 1,000 long-horizon tasks written by practitioners. My earlier benchmarks cover anomaly detection in text: NLP-ADBench compares 19 methods on eight datasets, and AD-LLM tests what LLMs add as zero-shot detectors, data augmenters, and model selectors. On the optimization side, I study where capacity should go inside agent pipelines — which stage of a generate–critique–revise loop actually needs a large model — and why automation stalls in operationally complex industries.
All papers in this line of work; the full list is on the Publications page.
A language model can reach the right answer for the wrong reason, and accuracy alone cannot tell the two apart. My work tries to make that difference visible and to shrink it. One part is reasoning itself: fine-tuning on explicit causal structure reduces hallucination. Another is interpretability: transferring what a language model knows into Tsetlin Machines gives classifiers whose decisions can be read clause by clause. The same concern runs through radiology reports anchored to the anatomy and simulations of human decision-making grounded in real population data.
All papers in this line of work; the full list is on the Publications page.
Some properties are easier to build into a model than to train into it. STRAT makes the separation between data and the rules that act on it part of the architecture: it splits the residual stream into orthogonal Data and Type subspaces, and Type-based attention and gating decide how Data is transformed. Trained from ten base examples per dataset, it beats a Transformer baseline on all 11 datasets tested and holds up better under distribution shift. With Prof. Cristina Lopes and Prof. Alberto Krone-Martins, I am now developing new model architectures for reasoning in physics and astronomy.
All papers in this line of work; the full list is on the Publications page.
When data cannot leave the hospital or the device that produced it, the model has to be trained where the data lives, and every round of communication costs bandwidth and energy. My work here makes federated training cheaper without giving up accuracy. H-FedSN trains personalized sparse networks across a client–edge–cloud hierarchy and cuts communication by up to 238×; FedBCGD, aimed at large models, has each client upload a single block of parameters per round. The same constraints show up in applications: personalized medication across distributed healthcare systems, and spatio-temporal graph learning where raw data is never pooled.
All papers in this line of work; the full list is on the Publications page.
A few questions I get often, answered ahead of time. For anything else, just email me.
My first-author paper Beyond Output Correctness has been accepted to the Main Conference of EMNLP 2026 (15.4% acceptance rate)! The paper argues that grading only the final answer misses how a model got there, and introduces CodeRQ-Bench and the two-stage evaluator VERA to score the reasoning itself. See you in Budapest. [arXiv] [GitHub]