llm-baseline

This paper proposes a cycle-consistent neural architecture that generates faithful natural language explanations of formal verification certificates, achieving 90% soundness and 860x faster inference than LLM baselines.