Neuro-Symbolic Framework Translates Natural Language to Executable Narsese for Reliable Reasoning

The inherent unreliability of large language models (LLMs) in tasks requiring explicit symbolic structure, multi-step inference, and interpretable uncertainty presents a significant bottleneck for their deployment in critical applications. A novel neuro-symbolic framework directly confronts this limitation by establishing a pipeline to translate natural language reasoning problems into executable formal representations using First-Order Logic (FOL) and Narsese, the language of the Non-Axiomatic Reasoning System (NARS). This approach represents a crucial step towards integrating the generative power of LLMs with the rigorous, verifiable reasoning capabilities of symbolic AI, promising a new era of more robust and transparent intelligent systems.
Central to this framework is the introduction of NARS-Reasoning-v0.1, a benchmark designed to evaluate this translation and execution process. This benchmark pairs natural language problems with their corresponding FOL forms, executable Narsese programs, and three gold labels: True, False, and Uncertain. A deterministic compilation pipeline ensures that the translation from FOL to executable Narsese is not only syntactically sound but also behaviorally aligned with the intended answers, validated through runtime execution in OpenNARS for Applications (ONA). Furthermore, the concept of Language-Structured Perception (LSP) is introduced, training LLMs to generate reasoning-relevant symbolic structures rather than just final verbal responses, as demonstrated by a Phi-2 LoRA adapter trained on the benchmark for three-label reasoning classification.
The strategic implications are substantial, offering a practical pathway to mitigate the "black box" problem and hallucination tendencies prevalent in purely neural architectures. By enabling execution-based validation and focusing LLMs on producing structured symbolic outputs, this framework lays the groundwork for AI systems capable of not just generating text, but also providing verifiable, step-by-step reasoning. This could unlock new applications in areas demanding high-stakes decision-making, such as legal analysis, scientific discovery, and complex engineering, where the justification of an AI's conclusion is as important as the conclusion itself. The long-term trajectory points towards hybrid AI architectures as the standard for achieving truly intelligent and trustworthy systems.