Formal safety verification for LLM‑integrated trajectory planners

Establish a formal safety verification framework for trajectory planning systems that integrate large language model–derived commonsense guidance into Monte Carlo Tree Search via a trust‑weighted Dirichlet policy, as instantiated by the C‑TRAIL framework for autonomous driving, providing provable guarantees that planned actions satisfy specified safety properties under the Recall–Plan–Update loop.

Background

C‑TRAIL integrates LLM‑derived commonsense into trajectory planning using a dual‑trust mechanism and a trust‑guided MCTS with a Dirichlet prior. While empirical results show improved safety and robustness, the paper notes that these are not backed by formal guarantees.

Because autonomous driving is safety‑critical, informal or empirical validation is insufficient to assure correctness under all conditions. The authors therefore highlight that devising formal verification methods for planners that rely on LLM guidance remains unresolved, motivating a precise verification framework tailored to LLM‑integrated planning systems like C‑TRAIL.