VLM-RobustBench: A Comprehensive Benchmark for Robustness of Vision-Language Models

Vision-language models (VLMs) and vision-language-action models (VLAs) are increasingly deployed in real-world applications. There, a small perturbation to the recorded camera image may change a decision significantly. However, existing benchmarks for these models only sample perturbations, which does not guarantee the absence of a failure in the untested region. We present the first robustness validation of six VLMs (drawn from the Gemma, InternVL, LLaVA, and Qwen families) and five VLAs (drawn from the GR00T, OpenVLA, and $π$ families) over entire continuous regions of photometric and geometric image perturbation: brightness shifts, camera rotations, and their composition. To this end, we build on the validation framework H$^2$V and introduce H$^2$V-M, a margin-aware convergence rule that makes validation affordable at the 32B parameter scale. We demonstrate that H$^2$V-M outperforms H$^2$V by an order of magnitude in model queries and that it finds counterexamples faster than random sampling while providing soundness guarantees. Our VLM and VLA robustness validation shows that robustness is mostly dependent on the perturbation type, rather than the model, and that VLMs are more robust to large camera rotations than VLAs. For VLAs, even perturbations as small as $\pm1^\circ$ can change the commanded action in many cases. We also show that robustness depends more on model family than on model size.

Vision-language models (VLMs) are increasingly used in settings where sensitivity to low-level image degradations matters, including content moderation, image restoration, and quality monitoring. Yet their ability to recognize distortion type and severity remains poorly understood. We present DistortBench, a diagnostic benchmark for no-reference distortion perception in VLMs. DistortBench contains 13,500 four-choice questions covering 27 distortion types, six perceptual categories, and five severity levels: 25 distortions inherit KADID-10k calibrations, while two added rotation distortions use monotonic angle-based levels. We evaluate 18 VLMs, including 17 open-weight models from five families and one proprietary model. Despite strong performance on high-level vision-language tasks, the best model reaches only 61.9% accuracy, just below the human majority-vote baseline of 65.7% (average individual: 60.2%), indicating that low-level perceptual understanding remains a major weakness of current VLMs. Our analysis further reveals weak and non-monotonic scaling with model size, performance drops in most base--thinking pairs, and distinct severity-response patterns across model families. We hope DistortBench will serve as a useful benchmark for measuring and improving low-level visual perception in VLMs.

The robustness of Vision Language Models (VLMs) is commonly assessed through output-level invariance, implicitly assuming that stable predictions reflect stable multimodal processing. In this work, we argue that this assumption is insufficient. We introduce a representation-aware and frequency-aware evaluation framework that measures internal embedding drift, spectral sensitivity, and structural smoothness (spatial consistency of vision tokens), alongside standard label-based metrics. Applying this framework to modern VLMs across the SEEDBench, MMMU, and POPE datasets reveals three distinct failure modes. First, models frequently preserve predicted answers while undergoing substantial internal representation drift; for perturbations such as text overlays, this drift approaches the magnitude of inter-image variability, indicating that representations move to regions typically occupied by unrelated inputs despite unchanged outputs. Second, robustness does not improve with scale; larger models achieve higher accuracy but exhibit equal or greater sensitivity, consistent with sharper yet more fragile decision boundaries. Third, we find that perturbations affect tasks differently: they harm reasoning when they disrupt how models combine coarse and fine visual cues, but on the hallucination benchmarks, they can reduce false positives by making models generate more conservative answers.