Integrating Large Language Models into Traffic Systems: Integration Levels, Capability Boundaries, and an...

Area of Science:

• Intelligent Transportation Systems
• Artificial Intelligence
• Information Theory

Background:

• Large language models (LLMs) offer advanced capabilities like semantic abstraction and multimodal fusion for intelligent traffic systems.
• Existing research on LLM integration spans data representation to autonomous agents.

Purpose of the Study:

• To review LLM integration in intelligent traffic systems through an information-theoretic lens.
• To analyze the limitations of LLMs in continuous, safety-critical traffic environments.
• To propose hybrid intelligence architectures for next-generation traffic systems.

Main Methods:

• Information-theoretic analysis conceptualizing LLMs as entropy-minimizing systems.
• Examination of LLM integration patterns and limitations.
• Delineation of boundaries between LLM capabilities and classical models.

Main Results:

• LLMs excel at managing high semantic entropy (contextual understanding, knowledge integration).
• A fundamental mismatch exists between LLM's discrete reasoning and traffic systems' continuous, causal nature.
• Classical physics-based and optimization models are crucial for low-entropy domains (real-time control, safety verification).

Conclusions:

• Hybrid intelligence architectures are essential for next-generation traffic systems.
• These architectures must bridge LLM-based semantic processing with physical system modeling.
• LLMs are vital for semantic tasks, while classical models ensure physical and causal integrity.