RONGKUN ZHOU / RESEARCHER & ENGINEER
Building AI systemsthat reason withevidence.
Language reasoning · Retrieval · Reliable machine learning
I turn model failures into precise research questions, then build benchmarks and systems that make the answers measurable.
01 / Featured research
ACCEPTED AT COLM 2026
SciTaRC
Scientific Tabular Reasoning and Claims
The question
Can a model reason over scientific evidence—not just describe it?
SciTaRC is an expert-authored benchmark for questions that require language understanding and complex computation over scientific tables.
Instead of treating reasoning as one opaque score, the study separates three stages. That makes it possible to see whether a system misunderstood the evidence, chose the wrong strategy, or failed while carrying out a sound plan.
Understand
Read the table, units, labels, and relationships in context.
Plan
Translate the question into a sequence of valid operations.
Execute
Carry out the plan faithfully and return a supported answer.
What the benchmark reveals
Even leading reasoning models often form the right strategy, then fail during execution. The bottleneck is not always knowing what to do—it is doing it correctly.
WHAT I WORK ON
One question, seen from three sides.
How can intelligent systems make claims that are grounded, inspectable, and dependable in practice?
Structured reasoning
Understanding how models plan and execute multi-step operations over tables, text, and interactive environments.
Evidence & retrieval
Connecting generated claims to supporting sources so answers are grounded, inspectable, and easier to trust.
Applied intelligence
Translating models into measurable systems across recommendation, logistics, forecasting, and computer vision.
EDUCATION
A quantitative foundation for intelligent systems.
M.S. in Computer Science
B.S. in Mathematics · Minors in Computer Science and Statistics