Article Source

• RL Loses Against Multi-Agent Intelligence

RL Loses Against Multi-Agent Intelligence: CoMAL

Abstract

The CoMAL (Collaborative Multi-Agent LLMs) framework addresses the complexities of mixed-autonomy traffic, where autonomous and human-driven vehicles coexist. By leveraging Large Language Models (LLMs) for high-level reasoning, collaboration, and decision-making, CoMAL allows autonomous vehicles (CAVs) to communicate, assign roles, and coordinate traffic strategies in real-time. Each CAV functions as an agent with modules responsible for perception, memory, collaboration, reasoning, and execution, which work together to ensure safe, efficient traffic flow. LLMs generate human-readable descriptions of traffic scenarios to facilitate natural, interpretable decisions and coordination among vehicles.

A key strength of the CoMAL framework is its role assignment system, which dynamically assigns tasks like leading or following to agents, preventing conflicts and enhancing collaboration. These roles allow for task specialization and reduce the computational complexity of managing traffic, improving traffic flow and safety. By combining LLM-based reasoning with rule-based models like the Intelligent Driver Model (IDM) for low-level control, CoMAL ensures that agents make generalizable high-level decisions while adhering to strict real-time safety and precision requirements in execution.

While the LLM-centric approach offers flexibility, generalization, and collaboration, it introduces inefficiencies in real-time processing. Alternatives such as reinforcement learning (RL) and hybrid models could improve low-level control while maintaining LLMs’ ability to manage complex, dynamic environments. Future research might explore integrating RL and LLMs or adopting neuro-symbolic systems to strike a balance between efficiency, precision, and adaptability in collaborative traffic systems.