Large Language Model-Tuned Rule Distillation for Adversarial Ramp Cut-in Testing of Autonomous Vehicles

Highway ramp merging is a safety-critical interaction for automated-vehicle validation because gap acceptance, longitudinal control, and lane-changing responses are coupled within a short time horizon. This paper presents a rule-distillation framework in which a large language model (LLM) serves only as an offline calibrator of a compact, interpretable rule expert: the model receives scalar replay statistics and proposes bounded updates of three rule parameters, every candidate is verified by closed-loop simulation, and the accepted rule expert is re-executed under legal-action masking and safety filtering to produce feasible demonstrations for an adversarial ramp actor, which is then fine-tuned by multi-agent reinforcement learning on a fixed scenario-replay library. On the training-distribution library, the distilled adversary reaches the highest evaluation return and reward-collision rate of the three schemes. This advantage only partly carries over to 150 held-out scenarios, where its reward-collision rate becomes indistinguishable from that of training from scratch (42.7% against 40.7%, overlapping intervals) while its evaluation return and its rate of successfully induced evasive responses remain higher. Against random-search and Bayesian-optimization calibrators evaluated under the same parameter ranges, score and simulation budget, the LLM attains a comparable calibration score with roughly one third of the closed-loop simulations and with a smaller run-to-run spread, but it is not a better global optimizer at the full budget, so it is best characterized as a sample-efficient and low-variance calibrator. The results indicate that verified rule calibration provides effective initialization for adversarial reinforcement learning within the evaluated simulation setting.