LLM-Assisted Light: Implementing Human-like Traffic Signal Control with Large Language Models

LLM-Assisted Light is an intelligent traffic signal control system developed by the Traffic-Alpha team. Its core is a five-stage hybrid decision-making framework that integrates the reasoning capabilities of large language models (LLMs) and the optimization capabilities of reinforcement learning (RL) to achieve human-like traffic management. The project is open-sourced on GitHub (https://github.com/Traffic-Alpha/LLM-Assisted-Light), and the related paper was published in March 2024 (arXiv:2403.08337). It is continuously updated to address issues such as poor adaptability of traditional control methods and insufficient interpretability of pure RL models.

Urban traffic congestion is a major challenge for modern cities. Traditional signal control relies on fixed timing or rule-based algorithms, which struggle to adapt to complex environments. While pure RL methods have made progress, they lack interpretability and perform unstable in extreme scenarios (e.g., sensor failures, emergency vehicle passage). The reasoning, common sense, and tool-using capabilities of LLMs bring new possibilities to solve these problems, so the LA-Light project explores a hybrid decision-making system combining LLMs and RL.

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

1. Task Planning: The LLM defines role responsibilities, understands scene objectives (such as traffic efficiency and emergency priority) and formulates strategy directions;
2. Tool Selection: Dynamically calls tools like road network perception, traffic status, RL decision support, and fault detection;
3. Environment Interaction: Collects real-time traffic data through the TransSimHub (SUMO-based) simulation platform;
4. Data Analysis: Fuses LLM reasoning, RL outputs, and real-time data to generate control strategies;
5. Execution Feedback: Implements decisions and provides an interpretable basis (e.g., "Extend green light because section E3 has a high occupancy rate and there is an emergency vehicle").

Technical implementation details:

• RL Module: Trained using the stable-baselines3 framework. Command: python train_rl_agent.py. Models are stored in the TSCRL/result/ directory and can be evaluated with eval_rl_agent.py;
• Pure LLM Mode: No tool calls needed. Command: python llm.py --env_name '3way' --phase_num 3 --detector_break 'E0--s'. Suitable for simple scenarios;
• Hybrid Decision Mode: Full functionality. Command: python llm_rl.py --env_name '4way' --phase_num 4 --edge_block 'E1' --detector_break 'E2--s'. Simulates complex scenarios (congestion + sensor failure).

Experimental validation scenarios:

• Normal Scenario: Dynamically adjust signal duration to optimize overall traffic efficiency;
• Emergency Vehicle Scenario: Prioritize emergency vehicle passage while balancing impacts on other directions;
• Sensor Failure Scenario: Use LLM reasoning combined with historical data to infer traffic status, demonstrating robustness advantages.

The team will launch the next-generation framework VLMLight in 2025, which introduces vision-language models (VLMs) to enhance scene perception and real-time decision-making capabilities. The related paper has been accepted by NeurIPS 2025, indicating that the LA-Light research approach has been recognized by the academic community and is evolving toward multimodality.

Practical significance:

1. Interpretability: Provides decision-making basis for easy manual supervision and fault diagnosis;
2. Robustness: LLM's common sense reasoning enhances performance under abnormal conditions;
3. Flexibility: Tool-using mechanism adapts to different scenario needs;
4. Scalability: The framework supports the introduction of new tools and capabilities. Application prospects: The hybrid decision-making method is expected to be applied in more real-world scenarios, promoting traffic management toward a more intelligent and humanized direction.

Summary: LA-Light demonstrates the innovative application of LLMs in the field of traffic signal control. The five-stage framework integrates LLM reasoning and RL optimization to achieve human-like management, providing a reference for LLM decision-making in the physical world. The open-source ecosystem relies on TransSimHub, LangChain, and stable-baselines3, uses the Apache 2.0 license, and encourages community contributions and secondary development.