Specifying complex requirements for autonomous and cyber-physical systems traditionally relies on Signal Temporal Logic (STL). However, translating natural language requirements into structured STL formulas is a persistent bottleneck, demanding specialized expertise and incurring significant costs with commercial LLMs, while also raising privacy concerns. The new ReasonSTL framework tackles this challenge head-on.
Decoupling Complexity in Formal Specification
ReasonSTL innovates by decomposing the natural language to STL generation process. It moves beyond end-to-end black-box approaches, instead leveraging explicit reasoning steps, deterministic tool calls, and structured formula construction. This modular design enhances transparency and control over the translation pipeline, a critical factor for industrial adoption where understanding the generation process is paramount.
Performance Gains with Local, Open-Source Models
A key breakthrough is demonstrating that even a compact 4B parameter open-source model, when trained with ReasonSTL, achieves state-of-the-art performance. This is validated through both automatic metrics and human evaluations on their novel extsc{STL-Bench} benchmark. This finding suggests a paradigm shift, enabling organizations to achieve high-fidelity natural language to STL generation without the prohibitive costs or privacy risks associated with proprietary, cloud-based LLM APIs. The focus on process-rewarded training further refines the model's ability to effectively utilize tools and produce accurate final formulas.
