**Abstract:** Performing question-answering over complex, multimodal content like television clips is challenging due to the reliance on single-modality reasoning, limited performance on long inputs, and lack of interpretability in current video-language models. To address these issues, the authors propose TV-TREES, the first multimodal entailment tree generator. TV-TREES enhances video understanding by producing trees of entailment relationships between simple premises directly entailed by the videos and higher-level conclusions. The authors introduce the task of multimodal entailment tree generation to evaluate the reasoning quality of such methods. Experimental results on the TVQA dataset demonstrate state-of-the-art zero-shot performance on full video clips, showcasing a best-of-both-worlds contrast to black-box methods. **Introduction:** The paper highlights the challenges in automated reasoning over semantically complex video-language data, particularly in narrative-centric video question-answering (VideoQA). While large, joint-modality transformer models often outperform smaller, domain-specific architectures, they lack interpretability and robustness. The authors propose TV-TREES, a multimodal entailment tree generator that jointly reasons over both modalities and provides human-interpretable evidence and explanations for each logical operation. **Contributions:** 1. The first multimodal entailment tree generator, a fully explainable video understanding system. 2. The task of multimodal entailment tree generation and a corresponding metric for evaluating step-by-step video-text reasoning quality. 3. State-of-the-art performance on zero-shot VideoQA using full-length video clips. **Related Work:** The paper reviews existing work on VideoQA, explainable multimodal understanding, and entailment tree generation, highlighting the limitations of current approaches and the need for more robust and interpretable methods. **TV-TREES:** The paper details the TV-TREES system, which involves three primary procedures: retrieval, filtering, and decomposition. The system recursively searches for evidence from the transcript and video frames to construct entailment trees. The authors also introduce an evaluation methodology based on informal logic theory, including metrics for acceptability, relevance, and sufficiency. **Experiments:** The TV-TREES system is evaluated on the TVQA dataset, comparing its performance against zero-shot VideoQA approaches. The results show that TV-TREES outperforms existing methods, particularly in terms of interpretability and robustness. The paper also discusses limitations and future directions for improving the system's performance and interpretability.**Abstract:** Performing question-answering over complex, multimodal content like television clips is challenging due to the reliance on single-modality reasoning, limited performance on long inputs, and lack of interpretability in current video-language models. To address these issues, the authors propose TV-TREES, the first multimodal entailment tree generator. TV-TREES enhances video understanding by producing trees of entailment relationships between simple premises directly entailed by the videos and higher-level conclusions. The authors introduce the task of multimodal entailment tree generation to evaluate the reasoning quality of such methods. Experimental results on the TVQA dataset demonstrate state-of-the-art zero-shot performance on full video clips, showcasing a best-of-both-worlds contrast to black-box methods. **Introduction:** The paper highlights the challenges in automated reasoning over semantically complex video-language data, particularly in narrative-centric video question-answering (VideoQA). While large, joint-modality transformer models often outperform smaller, domain-specific architectures, they lack interpretability and robustness. The authors propose TV-TREES, a multimodal entailment tree generator that jointly reasons over both modalities and provides human-interpretable evidence and explanations for each logical operation. **Contributions:** 1. The first multimodal entailment tree generator, a fully explainable video understanding system. 2. The task of multimodal entailment tree generation and a corresponding metric for evaluating step-by-step video-text reasoning quality. 3. State-of-the-art performance on zero-shot VideoQA using full-length video clips. **Related Work:** The paper reviews existing work on VideoQA, explainable multimodal understanding, and entailment tree generation, highlighting the limitations of current approaches and the need for more robust and interpretable methods. **TV-TREES:** The paper details the TV-TREES system, which involves three primary procedures: retrieval, filtering, and decomposition. The system recursively searches for evidence from the transcript and video frames to construct entailment trees. The authors also introduce an evaluation methodology based on informal logic theory, including metrics for acceptability, relevance, and sufficiency. **Experiments:** The TV-TREES system is evaluated on the TVQA dataset, comparing its performance against zero-shot VideoQA approaches. The results show that TV-TREES outperforms existing methods, particularly in terms of interpretability and robustness. The paper also discusses limitations and future directions for improving the system's performance and interpretability.