<ul><li>Large Language Models have shown impressive performance in mathematical reasoning tasks when guided by Chain-of-Thought prompting.</li><li>A structured framework that models stepwise confidence as a temporal signal and evaluates it using Signal Temporal Logic (STL) has been proposed.</li><li>Formal STL-based constraints are defined to capture desirable temporal properties and compute robustness scores for structured, interpretable confidence estimates.</li><li>Experiments show that this approach consistently improves calibration metrics and provides more reliable uncertainty estimates than conventional methods.</li></ul>

Temporalizing Confidence: Evaluation of Chain-of-Thought Reasoning with Signal Temporal Logic

Discover more