From Code to Collaboration: A Cognitive Agent Framework for Large Language Model (LLM)-Based Human-Vehicle Teaming

Abstract

This study develops a human-centered cognitive-agent framework for understanding how large language models (LLMs) can support human-vehicle teaming in automated driving. Following PRISMA guidelines, we reviewed 1,126 records published between 2021 and 2025 and included 52 studies after screening and full-text assessment. The synthesis identified four recurring capability clusters: perception and awareness, reasoning and decision-making, action and control, and interaction and communication. Across these functions, LLMs show promise for improving semantic scene understanding, explainable decision-making, high-level planning, and bidirectional communication with drivers. However, hallucinations, incomplete physical grounding, non-deterministic reasoning, and latency remain important limitations in safety-critical settings. The findings suggest that LLMs are most effective as high-level cognitive partners integrated with verified task-specific modules rather than as standalone controllers. The proposed framework offers design guidance for safer, more transparent, and collaborative human-vehicle systems.