LLM-Assisted Light: Empowering Intelligent Traffic Signal Control with Large Language Models

The LLM-Assisted Light project aims to use large language models to achieve human-like complex urban traffic signal control, constructing a five-stage hybrid decision-making framework by combining reinforcement learning and tool calling. This project addresses the problems of traditional reinforcement learning methods, such as lack of interpretability and unstable performance in extreme scenarios, providing an efficient and interpretable new solution for intelligent traffic signal control.

Urban traffic congestion is one of the core challenges faced by modern cities. Traditional traffic signal control systems are based on fixed timing or simple induction logic, making it difficult to handle complex and changing scenarios; reinforcement learning is widely applied but lacks interpretability and has unstable performance in extreme scenarios. The strong reasoning and knowledge understanding capabilities of large language models bring new opportunities to this field, leading to the birth of the LLM-Assisted Light project, which aims to build a hybrid intelligent system by combining LLM's human-like reasoning with RL methods.

The core innovation of LA-Light is the five-stage hybrid decision-making process:

1. Task Planning: The LLM clarifies the traffic management role and understands high-level semantic goals (e.g., prioritizing main roads, emergency vehicle passages);
2. Tool Selection: Dynamically call tools such as road network structure perception, traffic state monitoring, and sensor diagnosis, reducing information load through on-demand perception;
3. Environment Interaction: Collect real-time traffic data via the TransSimHub simulation platform;
4. Data Analysis: The LLM performs reasoning by combining tool data with pre-trained knowledge (e.g., inferring accidents from abnormal occupancy rates);
5. Execution Feedback: Output signal instructions and natural language explanations (e.g., reasons for extending green light duration).

LA-Light and RL collaborate complementarily: it includes a complete RL training and evaluation process (TSCRL directory, providing training/evaluation scripts), and the LLM intervenes and optimizes in edge scenarios that RL finds difficult to handle (e.g., sensor failures, emergency vehicles). The tool calling mechanism draws on LangChain, breaking down tasks into tool chains; the LLM independently decides the calling sequence, and new tools do not require retraining—only updating descriptions, making it highly scalable.

Three scenarios verify the system's capabilities:

1. Regular Congestion Handling: Call road network query and occupancy detection tools to extend the green light for congested directions and provide explanations;
2. Sensor Failure Tolerance: Simulate E2--s sensor failure (script example: python llm_rl.py --env_name '4way' --phase_num 4 --edge_block 'E1' --detector_break 'E2--s'), and the LLM uses common sense and historical data to compensate for decision-making;
3. Emergency Vehicle Priority: Understand the "life first" instruction, interrupt the phase to open a green channel, and explain the rationality of the decision.

The team later launched the VLMLight framework (accepted by NeurIPS 2025 in 2025), with upgrades including image perception capabilities, a dual-branch architecture for safety-critical decisions, and meta-control capabilities. The open source ecosystem is based on TransSimHub, LangChain, stable-baselines3, etc. The team gives back to the community (TransSimHub has become an important traffic simulation tool) to accelerate technology iteration.

LLM-Assisted Light represents a paradigm shift in traffic signal control: from purely data-driven black-box models to knowledge-enhanced interpretable intelligence. Core contributions:

1. Human-machine collaboration framework (decision assistant rather than replacing experts);
2. Tool-enhanced intelligence (overcoming LLM hallucinations and RL's interpretability defects);
3. Semantic understanding capabilities (understanding high-level instructions and complex scenarios). In the future, it is expected to be truly deployed to support smart cities, providing researchers with an entry point for large model applications in physical world decision-making.