From Natural Language to Executable Narsese: A Neuro-Symbolic Benchmark and Pipeline for Reasoning with NARS

Researchers Mina Gabriel and Pei Wang have introduced a novel neuro-symbolic framework designed to address the reliability gap in AI reasoning. Their system translates natural-language reasoning problems into executable formal representations using First-Order Logic (FOL) and Narsese, the programming language of the Non-Axiomatic Reasoning System (NARS). This approach moves beyond traditional large language models that generate verbal responses, instead producing structured symbolic representations that can be executed and validated.

To support this direction, the team created NARS-Reasoning-v0.1, a comprehensive benchmark containing natural-language reasoning problems paired with FOL forms, executable Narsese programs, and three gold labels: True, False, and Uncertain. They developed a deterministic compilation pipeline from FOL to executable Narsese and validated examples through runtime execution in OpenNARS for Applications, ensuring symbolic targets are both syntactically correct and behaviorally aligned with intended answers.

The researchers also present Language-Structured Perception (LSP), a formulation where an LLM is trained to produce reasoning-relevant symbolic structure rather than only final verbal responses. As a proof of concept, they trained and released a Phi-2 LoRA adapter on their benchmark for three-label reasoning classification. This work positions executable symbolic generation and execution-based validation as a practical path toward more reliable neuro-symbolic reasoning systems that can handle explicit symbolic structure, multi-step inference, and interpretable uncertainty.