Improving Machine Self-Assessment by Turing Learning

Szerzők és előadók

Hajnal Zsófia · Vásárhelyi Orsolya · Lendák Imre

Szekció és helyszín

MI, Kína és az algoritmikus társadalom
2026. november 26., csütörtök · 11:30–12:45
107-es terem, Szentkirályi utca 28.

Teljes, beküldött absztrakt

The phenomenon of benchmark saturation (forms of artificial intelligence surpassing the performance standards humans can establish through tests) means that it is no longer only up to ourselves to safeguard against deception by artificial intelligence (AI). Beyond watermarking, one promising primary tool against deception through AI (scams, synthetic content, fake bots, etc.) may be the given AI model itself, provided that it is better at self-detection than deception. To fulfil this latter condition ought to be a cybersecurity imperative, but the methods are not straightforward. This study has an exploratory and a translational goal. One is the mapping and testing of distinct challenges to large language models along the threads of computing, logic, and geometry. Our respective starting points are the halting problem (as proposed by Alan Turing), the application of the liar paradox (as proposed by the Ancient Greek philosopher Epimenides), and operations on non-orientable surfaces, such as the Klein bottle and the Möbius strip. These examples contain elements with the potential to perplex large language models (LLMs). By being able to describe the characteristics of the challenges at the frontier of LLM capabilities in dynamic terms, we aspire to generalise the method of proposing such challenges, essentially building a Turing learning (a machine learning method for two competing systems) based algorithm for AI to be able to self-assess and challenge itself.

Forrás