AIBRIEF
M3 Former, a maritime trajectory model, uses a language model to read a ship's attributes and intent, then forecasts its position four hours out, cutting displacement error 4–5% on Danish data: a modest result.
Tens of thousands of cargo ships broadcast position every few seconds through the Automatic Identification System (AIS). Port authorities, shipping operators, and sanctions watchdogs all want to know where each ship will be a few hours from now. A new preprint from the University of Chinese Academy of Sciences uses a large language model to forecast ships' future positions.
The model, M3-Former, uses a language model to encode a vessel's static attributes (size, type, destination) and navigational intent as semantic priors, then layers a dual-granularity Mixture-of-Experts on top. Sequence-level experts model global navigation trends; token-level experts refine fine-grained maneuvering. A steering-weighted loss upweights the rare sharp-turn samples that drive the worst collision and detection failures.
The reported result on a Danish Maritime Authority AIS dataset covering January 1 to March 31, 2023 is modest: 4.4% lower Average Displacement Error and 5.1% lower Final Displacement Error at the four-hour horizon, against the strongest baseline the authors compared against.
The arXiv preprint was submitted on August 2, 2026 and is not peer-reviewed. The code is public on GitHub; a Hugging Face model card shows two downloads last month and an empty README. No independent benchmark, multi-region, or multi-season replication is established, and the paper describes no commercial deployment.
Physical-infrastructure AI is now borrowing language models as semantic encoders for things that don't speak, and the first independent benchmark on a non-Danish dataset will tell whether the 4–5% gain survives.