The Aviv Regev paper you link was recently recommended to me as a useful reference for something. It was a nice surprise to see that Regev's thesis advisor was Ehud Shapiro, known to the Prolog community from his co-authorship of one of the good Prolog books (The Art of Prolog, with Leon Sterling - https://mitpress.mit.edu/9780262691635/the-art-of-prolog/). Indeed, Regev's thesis (and the paper above) propose a system based on a Flat Concurrent Prolog.

Shapiro was also the author of one of the two PhD theses that were a major influence to Inductive Logic Programming, a field at the intersection of logic programming and machine learning.

A lot of the kind of "deep work" you mention used to be done in the logic programming and ILP community in times past, before everyone seemingly switched to neural nets and statistical machine learning.