SCP-NL2TL: Selective Conformal Prediction for Reliable NL-to-TL Translation

A recent preprint on arXiv, identified as 2608.05439, presents SCP-NL2TL, a framework designed for selectively translating natural language instructions into temporal logic specifications. This initiative aims to enhance the reliability of safety-critical autonomous systems. Drawing from selective conformal prediction, the framework not only creates formal specifications but also assesses their trustworthiness. It employs two complementary black-box signals: the fidelity of back-translated specifications and the variability of repeated translations that maintain semantic equivalence. These signals identify different types of errors, allowing for a more precise distinction between incorrect translations when combined. Conformal risk control transforms this score into a decision-making rule, enabling the system to refrain from generating unreliable outputs. This framework is particularly suited for robots and autonomous systems that require planning, reasoning, and formal verification of their actions.