LLMs generate SHACL code without syntax errors—but they make mistakes in terms of meaning

The most dangerous error in SHACL generation doesn’t look like an error. Modern LLMs are already quite good at generating syntactically valid shapes, but when it comes to complex logical and structural constraints, they often fail to preserve the original meaning of the requirement.

SHACL is used to verify whether an RDF knowledge graph complies with specified rules. The problem is that writing such rules by hand is a task typically handled by technical specialists, while domain experts formulate requirements in natural language.

The idea behind NL2SHACL sounds practical: describe a condition in words and get a ready-made schema. But we cannot limit ourselves to comparing strings here. Two shapes can look and be structured differently while still specifying the same constraint. That’s why the new NL2SHACL-Bench checks not only the form of the response but also its semantic correspondence to the original requirement.

This is a significant shift in the evaluation of AI tools. If a model returns a document that passes the parser’s validation, it does not necessarily mean that it has understood the business rule. For simple requirements, automation already looks promising, while complex conditions still require independent verification of meaning.

It’s also worth keeping the study’s limitations in mind: the abstract does not name the four models tested, their specific results, or the composition of the complex test cases. Therefore, while the study highlights the problem and provides a basis for comparison, it does not answer which specific LLM performs better with NL2SHACL.

When you verify an LLM’s output for your task, is syntactic validity sufficient for you—or do you already check that the rule preserves the original meaning? Source: cs.AI updates on arXiv.org