NeuroNL2LTL translates natural language to logic with 86% verified satisfiability

NeuroNL2LTL, developed by Paapa Kwesi Quansah and Ernest Bonnah, bridges natural language and Linear Temporal Logic (LTL) for formal verification. It uses a neurosymbolic architecture: neural components learn fluent translations, but outputs are routed through an intermediate representation that preserves structure when mapped to LTL. Generated specifications then undergo satisfiability and non-triviality checks, with a minimal-edit repair mechanism correcting near-miss translations before they reach downstream verification tools. The key innovation is verifier-in-the-loop training—verification outcomes act as reward signals for reinforcement learning, making neural components optimize directly for formal correctness rather than just linguistic fluency.

Tested on over 200,000 requirements spanning aerospace, robotics, autonomous vehicles, and ten other domains, NeuroNL2LTL achieves 28% exact semantic equivalence with reference specifications while ensuring 86% of outputs are satisfiable (via a solver). The system also generates contextually grounded explanations from LTL, allowing domain experts to validate specifications without specialized training. This work demonstrates that formal verification can serve as both a training objective and a runtime filter, offering neural-based specification tools with reliability grounded in logical guarantees rather than statistical confidence.