Ask HN: LLM Logic Programming Research?
Is there any research into using something like Transformer but for direct logical inference?
I have found papers which discuss using LLMs to transform natural language into something like Prolog and run logic as a separate and traditional symbolic engine.
I have a hunch that it would be more powerful to skip the generation step and somehow integrate symbolic reasoning at the attention-wordvec level.
Recommend any papers? There is a lot of work on neurosymbolic systems, see this review https://arxiv.org/pdf/2305.00813.pdf I'd note that it would be much more powerful to produce code to run in an SMT solver https://en.wikipedia.org/wiki/Satisfiability_modulo_theories as opposed to Prolog. It's clear to me how you can get the LLM to generate code for such a system but many of the tough problems about decidibility are as tough as they ever were. On top of that there is the problem of efficient logical inference. For instance the Cyc system has more than 1000 specialized inference engines for particular kinds of problems, which doesn't seem like a terribly sustainable route. Ah, seems I was thinking of category 1b. I’ll keep reading. Great find.