Researchers combine language models with flight trajectory data to predict airline delays

Scientists have developed LLM4Delay, a framework that harnesses large language models to predict flight delays by unifying textual aeronautical data with aircraft trajectory information. The approach targets air traffic management from the perspective of controllers monitoring planes as they enter the terminal maneuvering area—the critical airspace surrounding an airport where landing procedures begin.

How it works#

The system combines multiple types of information into a single prediction model. It ingests flight records, weather reports, and aerodrome notices alongside multiple trajectory streams that capture airspace conditions. Rather than treating text and trajectory data separately, LLM4Delay converts trajectory representations into language-compatible form through what researchers call instance-level projection—a technique that maps trajectory-based features into the same semantic space as language model embeddings.

This cross-modal approach lets the framework tap into knowledge already embedded in both pretrained trajectory encoders and established language models, allowing them to complement each other rather than work in isolation.

Why it matters#

Flight delays ripple through aviation systems, creating operational inefficiency and cascading problems for passengers and airlines. Accurate prediction in real-world operations requires controllers to integrate diverse information sources—some numerical and sequential (trajectories), others textual and qualitative (weather, notices). By bridging these modalities, LLM4Delay demonstrates better performance than existing air traffic management systems.

The framework also supports continuous prediction updates: as new information emerges during a flight's approach, the model can refine its forecast, offering practical value for real-time decision-making.

Research context#

The work, submitted to arXiv in the machine learning community, has undergone multiple revisions since October 2025, reflecting ongoing refinement. It represents a growing trend of adapting large language models to specialized domains beyond text—in this case, the intersection of aviation data and operational forecasting.