Are Video Reasoning Models Ready to Go Outside?

arXiv cs.CV / 3/12/2026

💬 OpinionSignals & Early TrendsIdeas & Deep AnalysisModels & Research

共有:

Key Points

ROVA is a robustness-focused training framework for vision-language models that uses a robustness-aware consistency reward under spatio-temporal corruptions to improve performance in real-world disturbances.
It employs a difficulty-aware online training strategy that re-estimates sample difficulty via self-reflective evaluation to adapt training based on the model’s evolving capability.
The authors introduce PVRBench, a benchmark that injects real-world perturbations into embodied video datasets to evaluate both accuracy and reasoning under disturbances.
Evaluations on PVRBench, UrbanVideo, and VisBench show models suffer up to 35% accuracy drop and 28% reasoning drop under perturbations, while ROVA boosts relative accuracy by at least 24% and reasoning by over 9% compared with strong baselines, with gains transferring to clean benchmarks.

Abstract

In real-world deployment, vision-language models often encounter disturbances such as weather, occlusion, and camera motion. Under such conditions, their understanding and reasoning degrade substantially, revealing a gap between clean, controlled (i.e., unperturbed) evaluation settings and real-world robustness. To address this limitation, we propose ROVA, a novel training framework that improves robustness by modeling a robustness-aware consistency reward under spatio-temporal corruptions. ROVA introduces a difficulty-aware online training strategy that prioritizes informative samples based on the model's evolving capability. Specifically, it continuously re-estimates sample difficulty via self-reflective evaluation, enabling adaptive training with a robustness-aware consistency reward. We also introduce PVRBench, a new benchmark that injects real-world perturbations into embodied video datasets to assess both accuracy and reasoning quality under realistic disturbances. We evaluate ROVA and baselines on PVRBench, UrbanVideo, and VisBench, where open-source and proprietary models suffer up to 35% and 28% drops in accuracy and reasoning under realistic perturbations. ROVA effectively mitigates performance degradation, boosting relative accuracy by at least 24% and reasoning by over 9% compared with baseline models (QWen2.5/3-VL, InternVL2.5, Embodied-R). These gains transfer to clean standard benchmarks, yielding consistent improvements.