CDH-Bench: A Commonsense-Driven Hallucination Benchmark for Evaluating Visual Fidelity in Vision-Language Models

Vision-language models (VLMs) achieve strong performance on many benchmarks, yet a basic reliability question remains underexplored: when visual evidence conflicts with commonsense, do models follow what is shown or what commonsense suggests? A characteristic failure in this setting is that the model overrides visual evidence and outputs the commonsense alternative. We term this phenomenon $\mathbf{\text{commonsense-driven}}\mathbf{\text{hallucination}}$ (CDH). To evaluate it, we introduce $\mathbf{\text{CDH-Bench}}$, a benchmark designed to create explicit $\mathbf{\text{visual}}\mathbf{\text{evidence–commonsense}}\mathbf{\text{conflicts}}$. CDH-Bench covers three dimensions: $\mathit{\text{counting}}\mathit{\text{anomalies}}$, $\mathit{\text{relational}}\mathit{\text{anomalies}}$, and $\mathit{\text{attribute}}\mathit{\text{anomalies}}$. We evaluate frontier VLMs under $\mathit{\text{binary}}\mathit{\text{Question}}\mathit{\text{Answering}}\mathit{\text{(QA)}}$ and $\mathit{\text{multiple-choice}}\mathit{\text{QA}}$, and report metrics including $\mathit{\text{Counterfactual}}\mathit{\text{Accuracy}}$ (CF-Acc), $\mathit{\text{Commonsense}}\mathit{\text{Accuracy}}$ (CS-Acc), $\mathit{\text{Counterfactual}}\mathit{\text{Accuracy}}\mathit{\text{Drop}}$ (CFAD), $\mathit{\text{Commonsense}}\mathit{\text{Collapse}}\mathit{\text{Rate}}$ (CCR), and $\mathit{\text{Relative}}\mathit{\text{Prior}}\mathit{\text{Dependency}}$ (RPD). Results show that even strong models remain vulnerable to prior-driven normalization under visual evidence–commonsense conflict. CDH-Bench provides a controlled diagnostic of visual fidelity under visual evidence–commonsense conflict.