Are Video Reasoning Models Ready to Go Outside?

TL;DR

ROVA is a training framework that enhances video reasoning model robustness against real-world disturbances by adaptive difficulty-aware training and a new benchmark, PVRBench.

Contribution

The paper introduces ROVA, a novel robustness-aware training method, and PVRBench, a benchmark for evaluating video reasoning under real-world perturbations.

Findings

ROVA reduces accuracy and reasoning drops by up to 35% and 28% under disturbances.

ROVA improves accuracy by at least 24% and reasoning by over 9% compared to baselines.

Performance gains from ROVA transfer to standard benchmarks, showing broad effectiveness.

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.