ALBI pairs AI-generated logic rules with a Prolog checker

A language model can propose a plausible rule, but plausibility is not the same as correctness. ALBI separates those jobs: a model writes candidate logic programs, and a symbolic engine checks them against supplied examples. The reported results illustrate both the promise of that division of labor and the limits of what example-based verification can establish.

Learning a rule from examples means finding something more general than a list of answers. In inductive logic programming, the learner receives background facts, examples that should follow from the rules, and examples that should not. Its output is a logic program: explicit rules that can be inspected and executed.

In an arXiv preprint dated November 3, 2026, researchers at Tallinn University of Technology introduce ALBI, short for Agent for Learning By Induction. Their system combines a large language model with SWI-Prolog, a logic-programming engine. The paper reports encouraging benchmark results, but those findings remain preprint claims rather than established evidence of broadly reliable reasoning.

Propose, check, revise

ALBI adapts the Learning from Failures approach used by the existing system Popper. Popper generates candidate programs using an answer set programming solver, tests them, and uses failures to restrict subsequent search. ALBI replaces that generator with a language model.

The model receives background knowledge, positive and negative examples, and declarations describing the target and permitted supporting predicates—the named relationships used in logic programs. It is asked for three candidate programs. SWI-Prolog then evaluates each candidate against the examples, with timeout protection.

ALBI favors candidates that cover no negative examples, then rewards coverage of positive examples. Failures produce feedback about missed positives, incorrectly covered negatives, or syntax problems. This feedback guides another generation round; it is not equivalent to an exhaustive symbolic search with formally enforced pruning constraints.

For programs needing several rules, ALBI can retain clauses that explain some positives without covering negatives, then work on the remaining examples. A separate optimization stage attempts to simplify successful programs and checks the proposed revisions against the examples again.

Complete benchmark coverage, but slower execution

The authors evaluated 30 noise-free tasks from Popper’s benchmark suite, covering program synthesis, game-related rules, spatial reasoning, robotics, and other problems. They tested ALBI with Gemini 3.1 Pro, GPT 5.4 Pro, and Claude Opus 4.6.

All three configurations and Popper reportedly solved all 30 tasks. Here, success means finding a program that entails every supplied positive example and none of the supplied negative examples. It does not establish correctness on unseen cases.

With meaningful predicate names, ALBI averaged 1.4–1.7 generate-test cycles, compared with 6.8 reported refinement steps for Popper. These counts describe different generation procedures: an ALBI cycle requests three candidates, while Popper uses solver-based search. They are not interchangeable measures of computational work.

Nor did fewer cycles mean faster answers. Popper’s reported mean wall-clock time was 81.8 seconds, compared with 333.6 seconds for ALBI with Gemini, 397.7 with Claude, and 1,631.9 with GPT. The paper attributes ALBI’s longer runtimes to API latency and generation time. These measurements do not establish why a particular model reasoned as it did.

Removing meaningful names

To investigate dependence on linguistic clues, the researchers consistently replaced predicate names with abstract identifiers while preserving the tasks’ logical structure. A relationship name can otherwise suggest a familiar interpretation before the model examines its facts.

All three ALBI configurations retained full benchmark success after renaming. Average iteration counts increased to 2.6 for Gemini, 1.9 for GPT, and 2.5 for Claude.

The authors interpret this as evidence of structural induction rather than reliance on predicate-name meaning. The narrower conclusion is that meaningful names were not necessary for success on these tasks within ALBI’s feedback loop. Renaming alone does not establish general symbolic reasoning ability or exclude familiarity with benchmark structures from pretraining.

A checker cannot repair missing evidence

One example exposes the limits of verification against supplied cases. The paper describes an optimization that simplified a correct rule into one that still passed the examples but was incorrect more generally. Insufficient negative examples left the checker unable to reject it. A shorter program was not necessarily a better explanation.

Comparison with Popper also requires care. ALBI does not use Popper’s additional structural and type restrictions, and it sometimes introduces Prolog features unavailable to Popper’s configured generator. The systems therefore explore different effective program spaces.

ALBI also lacks a complete-search guarantee, depends on changing model APIs, and was evaluated only on a small, noise-free benchmark collection. Its reported achievement is specific: model-generated proposals plus executable feedback solved these selected tasks. Whether that approach remains reliable on broader, noisier problems is still open.