Vision-language models (VLMs) can increasingly handle extended contexts and answer questions about long videos—but can they understand the narrative structure of movies as humans do? To investigate this question, we introduce MF² (Movie Facts and Fibs), a fully manually curated benchmark designed to evaluate whether models can identify, recall, and comprehend information central to the movie’s storyline, using both the visual and textual modalities. MF² includes over 50 open-licensed movies, each paired with manually constructed sets of claim pairs: one true claim (fact) and one plausible but false (fib), totaling over 850 pairs. Our analysis across 20 open-source and 2 proprietary models reveals a substantial gap relative to human performance. While subtitles provide a strong signal for narrative understanding, other textual sources (e.g., synopses) are less informative. In contrast, when visual information is distributed across the entire movie and jointly embedded with subtitles, models struggle to identify and effectively use the relevant content. However, when the visual input is restricted to claim-relevant visual content, models improve significantly, indicating substantial room for improvement in processing large multimodal contexts and distinguishing relevant from irrelevant information. Overall, our findings highlight persistent challenges in multimodal narrative understanding for current VLMs. We release our data and code to support further research in this direction.

This paper introduces a novel approach to learn visually grounded meaning representations of words as low-dimensional node embeddings on an underlying graph hierarchy. The lower level of the hierarchy models modality-specific word representations, conditioned to another modality, through dedicated but communicating graphs, while the higher level puts these representations together on a single graph to learn a representation jointly from both modalities. The topology of each graph models similarity relations among words, and is estimated jointly with the graph embedding. The assumption underlying this model is that words sharing similar meaning correspond to communities in an underlying graph in a low-dimensional space. We named this model Hierarchical Multi-Modal Similarity Graph Embedding (HM-SGE). Experimental results validate the ability of HM-SGE to simulate human similarity judgments and concept categorization, outperforming the state of the art.