This work examines whether vision‑language models possess commonsense reasoning about duration. We built a corpus of geotagged tweets that were co‑annotated by humans and large language models, providing a testbed for inferring stay length from noisy textual, visual, and contextual cues. Experiments reveal that even with multi‑turn chain‑of‑thought prompting, state‑of‑the‑art models struggle with duration inference. Error analysis shows that models often fail to integrate essential contextual information or rely too heavily on salient visual details while overlooking subtle hashtags and semantic nuances. These results highlight the substantial challenges that remain for context‑aware multimodal understanding.
Review