01MATHEMATICAL REASONING
MA-ProofBench — Read paper
How well can language models prove theorems? A two-tier benchmark for formal theorem proving in mathematical analysis.
MATHEMATICS · REASONING · AI
Hi, I’m Xinheng.谢欣桁
From equations
to reasoning.
Mathematician by training. AI researcher by curiosity.
I explore how mathematical ideas can help machines learn, reason, and become a little more reliable.
Explore my researchCurrently exploring mathematical proof reasoning at ModelBest.
+AN OPEN QUESTION01 — ∞
A little change in perspective.
01 /THE PERSON BEHIND THE PAPERS
A path with
a few turns.
✳
PhD viva passed
Expected graduation · November 2026
Electrical engineering. Financial mathematics. Then doctoral research in mathematics. My route into AI has taken a few turns, and each has given me a different way to look at a problem.
At the University of Strathclyde, my research has connected numerical methods, differential equations, and deep learning. That work led me to image denoising, anomaly detection, and an AI safety collaboration with The Alan Turing Institute.
These days, I work on mathematical proof reasoning in large language models at ModelBest. This is a place to share that work, and to leave a little room for whatever I get curious about next.
Electrical engineeringFinancial mathematicsMathematicsMachine reasoning
02 /SELECTED RESEARCH
Different questions.
A mathematical thread.
From understanding a noisy image to checking a mathematical proof: a selection of the problems I’ve worked on.
02ANOMALY DETECTION
2DSig-Detect — Read paper
Finding the unusual in an image. Bringing 2D-signatures and semi-supervised learning together for anomaly detection.
03NEURAL DIFFERENTIAL EQUATIONS
NODE-ImgNet — Read paper
A clearer image, through mathematics. Using differential-equation structure and neural ODEs to learn how to remove noise.
04ADVERSARIAL ROBUSTNESS
NODE-AdvGAN — Read paper
Small changes, different answers. Studying model vulnerabilities through a dynamical approach to adversarial image generation.
THE MATHEMATICAL FOUNDATIONS
Randomness, delays & numerical methods.
Reliable approximations for systems shaped by uncertainty and their past. My work on stochastic and delay differential equations is part of the foundation behind my research in AI.
03 /ALONG THE WAY
The route
so far.
A few places, people, and problems that have shaped how I think.
2026 — NOWNOW
ModelBest
Research Intern · LLM Mathematical Reasoning
Since March 2026, working on proof-related training data, model training, evaluation, and inference for mathematical reasoning.
2023 — 2024
The Alan Turing Institute × Strathclyde
AI Safety Research Collaboration
October 2023–October 2024. Contributed to research on model vulnerabilities and image anomaly detection using 2D-signatures.
2022 — NOW
University of Strathclyde
Doctoral Research · Mathematics
Since October 2022. Numerical analysis, differential equations, and deep learning. Viva passed; expected graduation November 2026.
2021 — 2022
China Textile Information Center
Researcher · Industry Economics
February 2021–July 2022. Industry analysis, technology research, and research writing to support planning and policy.
2019 — 2020
Loughborough University
MSc · Financial Mathematics
September 2019–September 2020. Received the MSc Project Prize.
2014 — 2018
Tianjin Chengjian University
Undergraduate · Electrical Engineering and Automation
September 2014–September 2018. Where the journey began.
WHAT KEEPS ME CURIOUS
Can a model reason through a proof?
Mathematical reasoning, proof-related training data, and evaluating what a model actually understands.
Where do learned models break?
Anomalies, adversarial examples, and the gap between a convincing output and a reliable one.
What can mathematics bring to AI?
Using differential equations and numerical methods to think about the design of learning systems.
04 /A LITTLE SPACE FOR THINKING
Notes in the
making.
Research notes, things I’m learning, and questions that stay with me. A quieter corner of this site, coming soon.
Nothing published just yetLET’S COMPARE NOTES
A good question
is a good start.
Working on mathematical reasoning, reliable AI, or something at the intersection? I’d be glad to hear from you.