RuleChef turns LLM task knowledge into fast, human-editable rules

RuleChef, introduced by Kovács et al., bridges the gap between LLM flexibility and rule‑based reliability. The framework generates executable rules for text classification, NER, and relation extraction by first prompting an LLM with a task description and a few labeled examples. The rules are then iteratively patched based on misclassifications on a held‑out set, and optionally refined with human feedback on existing rule logic. This means LLMs are only invoked during the learning phase — the resulting rule system runs without any model calls, making it orders of magnitude faster and fully deterministic. Importantly, rules are human‑readable and editable, so domain experts can inspect, tweak, or override logic without retraining. RuleChef can also bootstrap rules directly from any existing model’s input‑output pairs, useful for distilling black‑box models into transparent systems.

Preliminary tests on classification and NER tasks show competitive accuracy while maintaining interpretability. The open‑source release (Apache 2.0) allows teams to integrate RuleChef into their own pipelines. For enterprises that need auditable AI decisions — finance, healthcare, legal — RuleChef offers a practical path to combine the reasoning of LLMs with the transparency of hand‑crafted rules. It reduces dependency on expensive API calls at inference time and gives teams full control over their NLP systems.