TV-TREES is a novel multimodal entailment tree generator designed to enhance the interpretability and reasoning capabilities of video-language models. The paper introduces TV-TREES as the first system that generates trees of entailment relationships between simple premises directly entailed by videos and higher-level conclusions. This approach enables interpretable joint-modality reasoning by producing trees that provide human-interpretable evidence and natural language explanations for each logical operation. The system is evaluated on the challenging TVQA dataset, demonstrating state-of-the-art zero-shot performance on full video clips, offering a contrast to black-box methods. TV-TREES focuses on the manipulation of atomic "facts" retrieved from video clips to answer VideoQA questions. It jointly reasons over both modalities and is compatible with long video inputs. The system's evaluation method is based on informal logic and textual entailment tree generation, adapting these ideas to the multimodal domain with an emphasis on reliable evaluation. The system's results show that it performs competitively on zero-shot VideoQA for the difficult TVQA dataset while providing interpretable reasoning traces. The paper also introduces the task of multimodal entailment tree generation to assess the reasoning ability of such systems. The system is evaluated using three main qualia: acceptability, relevance, and sufficiency. These are measured through human annotations and GPT-4 evaluations. The results show that TV-TREES outperforms existing zero-shot methods on full clips, demonstrating the effectiveness of its approach in video understanding. The system's performance is evaluated on the TVQA dataset, comparing its performance against a text-only version of the architecture and competing zero-shot VideoQA approaches. The results show that TV-TREES achieves state-of-the-art performance on the TVQA benchmark, highlighting the effectiveness of its multimodal entailment tree generation approach. The system's results suggest that interpretable, neuro-symbolic approaches to video understanding are a strong alternative to existing methods and present exciting directions for future research.TV-TREES is a novel multimodal entailment tree generator designed to enhance the interpretability and reasoning capabilities of video-language models. The paper introduces TV-TREES as the first system that generates trees of entailment relationships between simple premises directly entailed by videos and higher-level conclusions. This approach enables interpretable joint-modality reasoning by producing trees that provide human-interpretable evidence and natural language explanations for each logical operation. The system is evaluated on the challenging TVQA dataset, demonstrating state-of-the-art zero-shot performance on full video clips, offering a contrast to black-box methods. TV-TREES focuses on the manipulation of atomic "facts" retrieved from video clips to answer VideoQA questions. It jointly reasons over both modalities and is compatible with long video inputs. The system's evaluation method is based on informal logic and textual entailment tree generation, adapting these ideas to the multimodal domain with an emphasis on reliable evaluation. The system's results show that it performs competitively on zero-shot VideoQA for the difficult TVQA dataset while providing interpretable reasoning traces. The paper also introduces the task of multimodal entailment tree generation to assess the reasoning ability of such systems. The system is evaluated using three main qualia: acceptability, relevance, and sufficiency. These are measured through human annotations and GPT-4 evaluations. The results show that TV-TREES outperforms existing zero-shot methods on full clips, demonstrating the effectiveness of its approach in video understanding. The system's performance is evaluated on the TVQA dataset, comparing its performance against a text-only version of the architecture and competing zero-shot VideoQA approaches. The results show that TV-TREES achieves state-of-the-art performance on the TVQA benchmark, highlighting the effectiveness of its multimodal entailment tree generation approach. The system's results suggest that interpretable, neuro-symbolic approaches to video understanding are a strong alternative to existing methods and present exciting directions for future research.