Arctic shipping IRL study: latent context cuts model performance by 16.5%

A team of researchers from six institutions, including Vaishnav Vaidheeswaran and Gabriel Spadon, ran a controlled evaluation of inverse reinforcement learning (IRL) for AI-assisted Arctic shipping navigation. Using 3,186 AIS-derived voyages from 202 vessels across nine Arctic shipping seasons, they compared a linear shared reward, a nonlinear shared reward, and a latent-context meta-IRL model built on the same nonlinear architecture. The nonlinear reward improved held-out likelihood by 50.9% over the linear baseline, but adding vessel-specific latent context reduced performance by 16.5%, suggesting the latent variables were not capturing genuinely hidden preferences.

Behavioral analysis, context probes, and a pre-registered feature-hiding ablation revealed that apparent vessel-level variation is largely explained by observable route and environmental conditions—like sea-ice density and weather—rather than hidden vessel-specific factors. The study also found that predictive accuracy, route fidelity, and reward transfer yield different model rankings, demonstrating that no single metric can fully evaluate learned rewards. The authors recommend testing whether observed route, environmental, and vessel features already explain behavioral variation before adding per-vessel latent context, which supports more reliable and interpretable AI deployment in safety-critical maritime domains.