Adrian de Wynter

I am a principal scientist at Microsoft and a researcher (PGR) at the University of York. I work in projects related to fundamental problems in AI and science, such as reasoning, measurement, and social impact.

My primary research interest is reasoning as it relates to language in humans and machines. Lately I have focused on LLM reasoning capabilities (e.g. in measurement and strategic reasoning). I also raised here the question of whether understanding (as measured Wittgenstein-style) is required for 'good' dialogue; and demonstrated that in-context learning is a weak form of learning.

I favour pragmatic/intuitionistic approaches, where proofs of complexity and convergence must be constructive. Since I work in industry, these solutions typically apply to production problems. For example, we used category theory to prove that some prompting strategies are objectively better than others; and that they produce more preferrable outcomes by users (it ended up being part of Copilot!). I also recently wrote an algorithm with cryptographic guarantees for determining trust in LLMs-as-judges.

In earlier work I showed that finding a globally optimal solution to model compression is undecidable; but proved that polytime approximation algorithms exist. I applied these results to BERT, reaching a (then) SOTA on model compression. This was later adapted on quantum circuit optimisation in work at ORNL. I also proved (bridging learning theory and topological data analysis) how (and when) LLM-based data augmentation works.

My other research interests include recreational mathematics (games), preserving endangered languages, and computational social science. In the latter I have worked on mitigating toxicity, unfairness, and other harms of LLMs; personalisation and sycophancy in LLMs; research on LLM research; and did one of the first studies of the impact of ChatGPT on loneliness. And I also publish in SIGBOVIK, because this job is actually fun.

In terms of service, I have--like everyone and their grandma--reviewed for AAAI, *ACL, NeurIPS, ICLR and so on. I also review for Nature (Communications, Artificial Intelligence), IEEE Transactions on Games, and ACM TIST.

If you are stalking me, here's my Google Scholar and LinkedIn. Media coverage of my work is below.

Last updated: Feb '26.