ALBI: LLM-Guided Inductive Logic Programming via Iterative Generate-Test-Constrain Refineme

| Source: arXiv AI

Tags: ALBI, inductive logic programming, SWI-Prolog, Popper, neuro-symbolic, Gemini 3.1 Pro

ALBI swaps the symbolic hypothesis generator in inductive logic programming for an LLM checked by SWI-Prolog, and solves all 30 standard ILP benchmarks with three frontier models in 1.4-1.7 iterations on average.

Details

ALBI (Agent for Learning By Induction) keeps the Learning from Failures paradigm but replaces its ASP-based hypothesis generator with an LLM, while SWI-Prolog stays as the formal verifier. The loop is generate-test-constrain: the LLM proposes hypotheses from structured prompts, Prolog tests them, and diagnostic feedback drives refinement. The authors also introduce semantic anonymization, a metamorphic test that renames predicates to check whether the model reasons structurally rather than matching on names. On 30 standard ILP benchmarks ALBI hit 100% success with Gemini 3.1 Pro, GPT 5.4 Pro and Claude Opus 4.6, matching Popper while needing only 1.4-1.7 iterations. Under anonymization all models stayed at 100% with 1.36-1.67 times more iterations.