LLMs for Air Traffic Control: Prompt Engineering and Evaluation.

Air traffic control (ATC) communication remains a human-centric, safety-critical domain, despite increasing automation in other areas of air traffic management. This study investigates the potential of large language models (LLMs) to generate realistic ATC transmissions. Researchers designed a multi-turn pipeline where LLMs acted as ATC, responding to a fixed pilot transcript, and tested various prompt structures and conditioning methods across several LLMs. The experiments revealed that providing an in-context example significantly improved the similarity of LLM-generated responses to ground truth. Counterintuitively, lighter, less constrained prompts performed better than heavily scripted ones, which tended to accumulate errors over the dialogue. Injecting correct historical dialogue could repair these errors. The findings highlight both the promise and current limitations of using LLMs for ATC, suggesting a concrete path for developing LLM-assisted systems.