Researchers have developed a new evaluation method called TimeCatch to assess the ability of vision-language models (VLMs) to capture temporal structure. They tested several VLMs on synthetic and real-world datasets by introducing anomalies in consecutive frames or individual frames, and found that while these models can detect frame-level anomalies accurately, they struggle to identify temporal anomalies. This suggests that current VLMs have a limitation in integrating information across frames to reason about temporal consistency. --- Why it matters: This matters because it highlights the limitations of current vision-language models in understanding temporal relationships between frames, which is crucial for applications like video analysis and surveillance. Source: https://arxiv.org/abs/2608.23474