An AI research paper proposes using large language models to solve one of the oldest bottlenecks in symbolic AI: getting logic programs to invent useful new concepts from scratch.
Inductive Logic Programming, or ILP, is a technique that learns human-readable rules from examples - think teaching a system to recognize poker hands by showing it data. The catch is that ILP needs explicit logical predicates (named, structured concepts) to work from, and inventing good ones has always required domain experts. ADVENT, described in a new arXiv preprint, swaps in an LLM to generate candidate predicates, then runs them through a Prolog interpreter for formal verification. The loop repeats - the LLM sees what broke, refines its guesses - until something holds. Invented predicates get saved to a shared knowledge pool so later tasks can reuse them instead of starting over.
The numbers are specific enough to be interesting: across nine poker-hand concepts tested on seven different LLMs, the LLM-only approach succeeded 58% of the time where standard ILP failed entirely. Adding formal verification pushed that to 80%, and reusing pooled predicates added up to 31 percentage points on top. More importantly, the resulting rules were human-readable - which is the entire point of ILP over a black-box neural net.
Symbolic AI has spent years playing understudy to deep learning, but hybrid approaches that pair LLMs with formal reasoning keep showing up in research. ADVENT is one more data point suggesting the combination is more than the sum of its parts - though poker hands are a long way from production knowledge engineering.