Auto-Relational Reasoning

In a new paper submitted to the Journal of Artificial Intelligence (JAIR), researchers Ioannis Konstantoulas, Dimosthenis Tsimas, Pavlos Peppas, and Kyriakos Sgarbas from the University of Patras introduce Auto-Relational Reasoning, a theoretical framework that marries the scalability of machine learning with the rigor of formal reasoning. The core idea is to automate reasoning through object-relations and integrate it directly into artificial neural networks, addressing the well-known soft limits of large models that show diminishing returns in reasoning tasks.

To demonstrate the framework, the team built a system that solves Intelligence Quotient (IQ) problems with zero prior knowledge of the problem domain. On a standard IQ test, the system achieved a 98.03% solving rate, corresponding to the top 1% percentile of human performance—an IQ score between 132 and 144. The authors note that this result is constrained only by the small model size and the processing capabilities of the machine it ran on, implying that scaling up could yield even higher scores. With the integration of prior knowledge and dataset expansion, the system is designed to generalize to a wide category of problems, favoring few-shot or zero-shot solutions.