Study: Visual Grounding in Long-Horizon VLMs Predicts Out-of-Distribution Generalization

A recent study published on arXiv (2603.06828) indicates that long-horizon vision-language models (VLMs) with temporally grounded beliefs exhibit superior generalization to out-of-distribution (OOD) data. The researchers define 'behavioral faithfulness' as a measurable characteristic that assesses the consistency of a model's intermediate reasoning with the changing visual context. Analyzing eight models across three long-horizon benchmarks, they demonstrate that the Step Grounding Rate (SGR) correlates with OOD retention at r = 0.83 (permutation test p = 0.003). This correlation persists among models of similar capacity and is not attributable to scale or in-distribution accuracy. The authors argue that traditional benchmarks focusing solely on final-answer accuracy may overlook the model's use of visual information, highlighting the importance of temporal grounding as a key metric for enhancing VLMs in long-horizon applications.