Practical Prolog Planner Prompting

Summary

Using Large Language Models to generate Prolog planners Combining the strength of Large Language Models (LLMs) with Prolog combinatorial search is a natural fit actually delivering tangible progresses in mainstream automated planning. State of the art in LLM-generated planners Parts of the public discussion about LLMs revolves around unachievably high goals such as "General AI", but LLMs are just statistical language models after all and excel in language translation and summarization tasks foremost. And while LLMs can be trained to do almost anything including optimization and other math and reasoning tasks, it's exceedingly clear results are determined by the used training data sets, and remain highly volatile and unpredictable for problems outside that set, a fact well known acknowledged in papers such as aiw and posp. Rather than having LLMs perform planning itself, it's thus only natural to complement LLMs with the raw combinatorial power of Prolog, much like a human would be reaching out to a pocket calculator. Or, as tnltopgwl formulates it Our empirical results [...] show that LLMs are much better suited towards translation rather than planning. Putting LLMs to use for translating natural language into Prolog is also a straightforward extension of its original goals considering Prolog was originally devised for natural language processing. Planning being another field of application right from the start for Prolog, it isn't surprising many researchers are experimenting with Prolog, Contraint Logic Programming (CLP), and other logic-based programming (LP) languages as a target for code generation using LLMs (see the bibliography for a small selection of reports). Practicitioners are also turning to solutions targetting Prolog rrbnl after having observed (swipldiscourse1) that targetting domain-specific planning languages and formalisms lacking the inherit logical underpinnings that Prolog has, such as the one discussed in tnltogw, after initial promising resu...