TempCloze: Can Video-LLMs Identify the Missing Middle?
Temporal reasoning benchmarks for Video-LLMs are often mediated by language, leaving room for linguistic shortcuts from option wording, answer correlations, or language priors. To reduce such shortcuts, we introduce TempCloze, a video cloze benchmark for evaluating visual temporal reasoning in Video-LLMs. Given the beginning and ending clips of a video, models must identify the true missing…
We haven't written up this one. arXiv cs.AI has the full story — the link below goes straight to it.
