Executing Signal Control From Text: LLM Fidelity on a CV Platoon-Based Optimization Algorithm

Connected vehicle (CV) signal control algorithms are often published as prose, yet deployment requires exact numeric execution; this paper tests whether large language models (LLMs) can execute a CV platoon-based signal optimization workflow correctly when the only specification is text. We model a two-approach intersection with through movements, group vehicles into ordered platoons, and define a control decision as a departure sequence that interleaves West and South platoons while preserving within approach order. For each candidate sequence, a deterministic Python reference enumerates feasibility, computes discharge times, selects advisory speeds for moving platoons under speed bounds, and returns average vehicle delay (AVD). We query LLMs under three Levels of Guidance (LoG), parse outputs into a scenario–sequence table, and score them against the reference using AVD error, optimum-sequence match, regret, rank agreement, speed-bound state match, and response time. Under High LoG, Gemini matches the reference optimum in 30/30 scenarios with median absolute error 0.005 s/veh and 90th percentile 0.0177 s/veh on per-sequence AVD, while ChatGPT matches 27/30 optima with low regret (median 0.068 s/veh); under Medium LoG, performance degrades sharply for some models, including large-regret failures tied to incorrect speed-bound clamping that shifts the timing recursion. These results show that procedural structure in the prompt drives execution fidelity in this setting and motivates compliance and style evaluation before making deployment claims.