We are proud to congratulate Andrey Ustyuzhanin (Constructor University · Constructor Knowledge Labs) and his co-authors Nikita Kazeev and Maxim Borisyak on winning 1st Prize at the Poster Presentation Competition of the AI4X Conference 2026, held 15–19 June at the Raffles City Convention Centre in Singapore.
The conference is co-hosted by NUS Institute for Functional Intelligent Materials, the University of Toronto, and the Acceleration Consortium. The certificate was signed by Conference Co-Chairs Konstantin Novoselov and Alan Aspuru-Guzik, and Poster Session Chair Daria Andreeva.
The award-winning poster — “The Geometry of Reasoning Failure” — asks a simple question: instead of reading what an LLM agent says while it reasons, can we measure how its reasoning moves? Each chunk of a chain-of-thought is collapsed to a single scalar — its Semantic Level of Detail — and the shape of that trajectory turns out to predict whether the agent will fail. The result: a lightweight, interpretable detector stack (0.89 ROC-AUC; +9.8 pp on FrontierScience; 39.5% of the Pass@8 oracle gap recovered) without any per-model training.
Congratulations to the team!
