There’s an impedance mismatch between symbolic and statistical views of world knowledge: one grounded in logic, the other in probability. Natural language, during the last century programming languages and more recently large language models, offer the strongest bridges we’ve ever seen between these paradigms. Perhaps one day they’ll be fully fused. And perhaps none of us will truly understand how.
It seems like if this really works, then it should do well on ARC.
At a glance it seems that the authors do not think this is viable in the real world. Without looking it up it appears that it is from the field of symbolic AI.