Enhancing Temporal Reasoning Capabilities of LLMs

Associate Professor, Dr. Sheng Li (Data Science); Collaboration with VT
Published

November, 2026

Ask a chatbot who invented the iPhone, and the answer is reliable every time. Ask who currently runs a major company, and the story gets shakier. Large language models learn from a fixed snapshot of text collected up to a certain point, and when a fact changes after that snapshot, the model does not always know it changed. It may confidently say an old CEO still runs a company long after that stopped being true. 

To tackle this problem, Dr. Sheng Li, an associate professor in UVA's School of Data Science, partnered with Dr. Dawei Zhou at Virginia Tech through a 4-VA at UVA collaborative research project, "Enhancing Temporal Reasoning Capabilities of LLMs." As Dr. Li put it, models are generally reliable on questions with a fixed answer. But on anything tied to a specific moment in time, they can quietly rely on outdated data and answer with just as much confidence. A model that admits uncertainty is a minor inconvenience. A model that states an outdated fact as if it were current can actively mislead someone. 

The team's approach is to give the model outside help in the form of knowledge graphs, which are structured networks of facts and how they connect to one another, with each fact tagged for when it was true. Their framework lets the model check its answer against this time-stamped structure before responding. The goal is not simply a smarter chatbot; it is a more dependable one for everyday questions whose answers depend on the news, the calendar, or the current state of the world. 

That pairing of structured data and model reliability did not come together by accident. Dr. Li and Dr. Zhou had known each other professionally for years and were familiar with each other's published work but had never found the occasion to build something together. That changed after a single conversation about how their two areas of expertise might combine to tackle a hard problem. The 4-VA at UVA request for proposals appeared shortly after --good timing for an idea that already existed but had not yet found a formal home.  

In this project, the two labs combine genuinely different strengths. Dr. Zhou contributes deep expertise in graph-based learning on how to structure and reason over connected data. The UVA side focuses on trustworthy AI. Making sure models are reliable, robust, and safe is a research thread Dr. Li's lab has pursued for roughly a decade, evolving from traditional machine learning to deep learning and now to large language models. 

Day to day, technical work has been driven primarily by Weili Shi, a Ph.D. student in Dr. Li's lab. He has worked closely with Dr. Li to frame the research problem, study recent papers in the area, and build and test early versions of the model.  Weili turned the idea into a working prototype, and his proof-of-concept experiments have already shown encouraging early results. He then brought the team back together to review those results and refine the method where it fell short. The hardest problems tended to surface inside the training process itself, which took repeated experimentation with different sample-selection strategies. “This taught me that successful LLM alignment depends not only on the training algorithm, but also heavily on careful data construction and experimental tuning,” he said. 

That same care was carried into his work mentoring the undergraduate researcher on the team, Yifan Qiao. With the guidance of Weili, Yifan conducted the literature review, assisted with data collection and analysis, and helped prepare survey materials used in the research. Explaining the project’s ideas and methods to someone newer to research pushed Weili to be more patient and structured, and to revisit his own assumptions along the way. He found walking someone else through the problem helped him see it from a new angle. Weili expects the experience to pay off whether he stays in academia or moves into industry. The research has sharpened his technical and problem-solving skills, and the cross-institutional collaboration and mentoring strengthened his capacity to communicate and lead, strengths he will carry into future teams. He is set to graduate at the end of this year, after one more semester in the lab. 

Closing the distance between confident and correct AI is a long-running problem, and the 4-VA grant gave the idea enough runway to prove it was worth pursuing. The team is already working to extend the work beyond the one-year grant by preparing a proposal for external funding. It is the kind of early-stage work that is often hardest to fund through traditional channels, where a small amount of timely support can make an outsized difference for a research direction that will make a notable impact.