OracleTSC: Enhancing Traffic Signal Control with LLMs and Uncertainty Regularization
Researchers have introduced OracleTSC, a novel framework designed to stabilize Large Language Model (LLM)-based Traffic Signal Control (TSC) systems. Traditional reinforcement learning methods often lack interpretability, while standard LLM fine-tuning for TSC suffers from instability due to sparse and delayed feedback. OracleTSC addresses these challenges through two key mechanisms: a reward hurdle that filters weak learning signals by subtracting a calibrated threshold, and uncertainty regularization that encourages consistent decision-making. Experiments using the LibSignal benchmark demonstrate that OracleTSC enables a compact LLaMA3-8B model to significantly improve traffic efficiency, achieving a 75% reduction in travel time and a 67% decrease in queue length compared to pretrained baselines. Furthermore, the system exhibits strong cross-intersection generalization, transferring effectively to structurally different intersections without additional fine-tuning. This approach preserves transparency through natural language explanations, aiming to build public trust in automated traffic management systems while enhancing operational effectiveness.
Wire timeline
OracleTSC: Enhancing Traffic Signal Control with LLMs and Uncertainty Regularization
Researchers have introduced OracleTSC, a novel framework designed to stabilize Large Language Model (LLM)-based Traffic Signal Control (TSC) systems. Traditional reinforcement learning methods often lack interpretability, while standard LLM fine-tuning for TSC suffers from instability due to sparse and delayed feedback. OracleTSC addresses these challenges through two key mechanisms: a reward hurdle that filters weak learning signals by subtracting a calibrated threshold, and uncertainty regularization that encourages consistent decision-making. Experiments using the LibSignal benchmark demonstrate that OracleTSC enables a compact LLaMA3-8B model to significantly improve traffic efficiency, achieving a 75% reduction in travel time and a 67% decrease in queue length compared to pretrained baselines. Furthermore, the system exhibits strong cross-intersection generalization, transferring effectively to structurally different intersections without additional fine-tuning. This approach preserves transparency through natural language explanations, aiming to build public trust in automated traffic management systems while enhancing operational effectiveness.
cs.AI updates on arXiv.org