At the 2026 World Young Scientists "AI+" Dialogue held in Macao on 28 March, a cross-disciplinary team from the University of Macau, Hong Kong Baptist University, The Education University of Hong Kong, Guangdong University of Technology and the Guangdong Digital Industry Research Institute globally launched the AI-safety "Twin Star" architecture — HoloAuditor and HoloBlocker. Built on pure mathematical topology, the architecture targets the "fatal hallucination" problem of large language models entering the physical world, laying a safety floor for embodied intelligence.
Opening the launch, Academician Prof. Ho Kin-Chung of The Education University of Hong Kong presented striking data: under the strictest TCM logic framework (monarch–minister–assistant–envoy), the newest generation of large models showed up to 6.0% extreme semantic-ranking collapse. On highly structured, axiom-driven tasks, reasoning does not improve with scale — it degrades systematically.
Prof. Ho stressed that as AI moves from virtual conversation into the physical world, hallucinations can cause irreversible physical consequences. The four-university team therefore open-sourced the "Twin Star" architecture, building a "hard boundary" for AI safety with the indisputability of mathematics.
Prof. Hu Hao of the University of Macau's State Key Laboratory of Quality Research in Chinese Medicine announced that the team would open-source HoloAuditor, a medical-logic auditor. The bottleneck of TCM internationalisation, he said, is quality control of real-world data: HoloAuditor maps "pattern differentiation and treatment" logic onto continuous mathematical manifolds, auditing AI output geometrically in real time — raising data quality and supporting cross-border regulatory mutual recognition.
Dr. Li Tak-Ho (Alex), Executive Dean of the Guangdong Digital Industry Research Institute and Visiting Associate Professor at HKBU, revealed the underlying technology: "as AI gets smarter at conversation, it may be getting dumber at reasoning." HoloAuditor maps the millennia-old monarch–minister–assistant–envoy logic into mathematical manifolds — geometry sets the rules for semantics. When AI tries to output a "popular but wrong" answer, the system computes a geometric distance and intercepts it within 0.82 milliseconds.
Dr. Li then introduced the twin result — HoloBlocker. The same topological logic transfers to industry: when an embodied agent (a drone, a robotic arm) receives an instruction that is "rule-permitted but logically lethal", HoloBlocker instantly triggers topological energy fusing and locks the hardware, stopping the dangerous action. "Agents gave AI wings; HoloBlocker is its radar and parachute."
Prof. Liu Kaijie, Vice-Dean of the GBA ESG & New Quality Productivity Research Institute at Guangdong University of Technology, added that HoloBlocker is more than a lab result: the team has encoded international industrial-safety standards as mathematical manifolds, making the technology a physical foundation for the industrial internet and embodied intelligence — key infrastructure both against physical accidents and for the "AI+" empowerment of new productive forces under global ESG governance standards.
At the close of the launch, Prof. Ho Kin-Chung, Prof. Hu Hao, Dr. Li Tak-Ho, Prof. Song Menghuan and Dean Liu Kaijie started a global countdown. As the two QR codes for the HoloAuditor and HoloBlocker releases appeared on screen, applause filled the hall — marking the formal opening of an era of AI topological safety led by Chinese scholars.
The team noted that the core algorithms have been filed for national invention patents, with results already on the SSRN preprint platform (DOI: 10.2139/ssrn.6368538; 10.2139/ssrn.6421619; 10.2139/ssrn.6368638) and source code to be opened worldwide, welcoming academia and industry into the safety ecosystem.
Organising committee / joint research team
Contact: Dr. Li Tak-Ho
Email: alexlihk@hotmail.com
Tel: +86 18420600450
The number most worth remembering from this report is "6.0% semantic-ranking collapse" — it shows that a bigger model is not a steadier reasoner: on highly structured, axiom-driven tasks, the newest-generation models degrade systematically. This is one of the core reasons the Committee promotes public AI-safety literacy: AI's capability curve and its reliability curve are not in sync, and users need to ask "what happens when it is wrong?"
The two wings of the "Twin Star" also map onto two of our "three locks": HoloAuditor's geometric audit guards reliability (verifiable data and logic), while HoloBlocker's topological fusing guards safety (a hard boundary in the physical world). Encoding safety standards as mathematical manifolds and intercepting dangerous actions within milliseconds shows AI governance moving from "accountability after the fact" to "computable interception during the event" — a direction industry on both sides of the boundary should watch. Preprints are on our publications page (SSRN 6368538, 6421619, 6368638).