ADVENT: LLM-Driven Automatic Predicate Invention for ILP

What Happened

Researchers have introduced ADVENT, a novel framework that leverages large language models to automate predicate invention within Inductive Logic Programming (ILP). Predicate invention—the process of creating new logical predicates to expand the hypothesis space—has historically been a manual, expertise-heavy bottleneck in ILP systems. ADVENT addresses this by using LLMs to generate meaningful new predicates automatically, rather than relying on human domain experts to predefine them. The system evaluates candidate predicates for both logical utility and semantic clarity, producing outputs that are interpretable and reusable across different learning tasks.

Why It Matters

Predicate invention has long been the Achilles' heel of ILP. Traditional approaches required practitioners to handcraft auxiliary predicates, which not only demanded deep domain knowledge but also often resulted in opaque, task-specific constructs that could not transfer to other problems. This severely limited ILP's scalability and practical adoption.

ADVENT's contribution is significant for three reasons. First, it removes the human bottleneck by automating a process that previously required specialized expertise. Second, by using LLMs to generate predicates with clear semantics, it addresses the "opacity problem"—where invented predicates are so abstract or arbitrary that they become meaningless to human analysts. Third, the approach creates predicates that are potentially reusable across domains, which could dramatically reduce the engineering overhead for ILP applications.

This work sits at an interesting intersection: it uses the statistical, pattern-matching strengths of LLMs to enhance the symbolic, rule-based reasoning of ILP. Rather than pitting neural and symbolic approaches against each other, ADVENT demonstrates a complementary relationship where LLMs serve as creative generators of logical primitives.

Implications for AI Practitioners

For researchers and engineers working on neuro-symbolic AI, ADVENT offers a concrete blueprint for hybrid systems. Practitioners building knowledge discovery tools or automated scientific discovery platforms should pay attention: the ability to automatically generate interpretable predicates could accelerate hypothesis generation in fields like bioinformatics, materials science, and program synthesis.

However, there are practical considerations. LLM-generated predicates still require validation—the framework must filter out nonsensical or logically redundant candidates. This introduces a computational cost that practitioners need to factor into their pipelines. Additionally, the quality of invented predicates will depend heavily on the LLM's training data and prompt design, meaning domain-specific tuning may still be necessary.

For ILP practitioners specifically, ADVENT could reduce the barrier to entry. Instead of spending weeks defining auxiliary predicates for a new domain, they can now focus on higher-level problem formulation and result interpretation. The trade-off is a new dependency on LLM infrastructure and API costs.

Key Takeaways

• ADVENT automates the historically manual process of predicate invention in ILP by using LLMs to generate semantically meaningful new predicates.
• The framework addresses two core ILP bottlenecks: reliance on domain expertise and production of opaque, non-transferable predicates.
• This represents a practical neuro-symbolic integration where LLMs handle creative generation while ILP maintains logical rigor and interpretability.
• Practitioners should weigh the reduced human effort against new computational costs and potential domain-specific tuning requirements for LLM prompts.