Analyzing the Narration Gap in LLM-Solver Loops

The paper presents an analysis of the "narration gap" in LLM-solver loops, focusing on the interaction between language models and formal solvers in producing sound answers to logic-based queries. It introduces a verified decision procedure to model the LLM-solver loop and evaluates five open-source models under prompt injection, finding that while certificate gating can ensure soundness, adversaries can still manipulate outputs. This research highlights the need for improved robustness in the narration step, as current mitigations do not fully protect against adversarial attacks, which is critical for practitioners relying on LLMs in safety and security contexts.