**Goldfish: Vision-Language Understanding of Arbitrarily Long Videos** This paper addresses the challenges of processing long videos, which are common in movies and TV series, by introducing *Goldfish*, a methodology designed to understand videos of arbitrary lengths. Goldfish employs an efficient retrieval mechanism that initially gathers the top-k relevant video clips before providing the desired response, enabling it to handle long video sequences effectively. The authors also introduce the TVQA-long benchmark, specifically designed to evaluate models' capabilities in understanding long videos with questions that cover both visual and textual content. To facilitate the retrieval process, the authors developed *MiniGPT4-Video*, a model that generates detailed descriptions for video clips. This model extends the MiniGPT-v2 architecture to process multiple frames with aligned subtitles, enhancing the model's ability to interpret and respond to video content. The evaluation results show that Goldfish achieves a 41.78% accuracy rate on the TVQA-long benchmark, surpassing previous methods by 14.94%. Additionally, MiniGPT4-Video outperforms existing state-of-the-art methods by 3.23%, 2.03%, 16.5%, and 23.59% on short video benchmarks such as MSVD, MSRVT, TGF, and TVQA. The contributions of this work include: - Developing Goldfish, a framework for long video understanding that leverages a retrieval mechanism to select the most relevant video clips. - Creating the TVQA-long benchmark, which requires models to understand both visual and textual content in long videos. - Developing MiniGPT4-Video, a model that extends VLMs to process multiple frames and improve content understanding. The authors conclude that their models and code have been made publicly available, and they hope that their work will benefit future research in long video understanding.**Goldfish: Vision-Language Understanding of Arbitrarily Long Videos** This paper addresses the challenges of processing long videos, which are common in movies and TV series, by introducing *Goldfish*, a methodology designed to understand videos of arbitrary lengths. Goldfish employs an efficient retrieval mechanism that initially gathers the top-k relevant video clips before providing the desired response, enabling it to handle long video sequences effectively. The authors also introduce the TVQA-long benchmark, specifically designed to evaluate models' capabilities in understanding long videos with questions that cover both visual and textual content. To facilitate the retrieval process, the authors developed *MiniGPT4-Video*, a model that generates detailed descriptions for video clips. This model extends the MiniGPT-v2 architecture to process multiple frames with aligned subtitles, enhancing the model's ability to interpret and respond to video content. The evaluation results show that Goldfish achieves a 41.78% accuracy rate on the TVQA-long benchmark, surpassing previous methods by 14.94%. Additionally, MiniGPT4-Video outperforms existing state-of-the-art methods by 3.23%, 2.03%, 16.5%, and 23.59% on short video benchmarks such as MSVD, MSRVT, TGF, and TVQA. The contributions of this work include: - Developing Goldfish, a framework for long video understanding that leverages a retrieval mechanism to select the most relevant video clips. - Creating the TVQA-long benchmark, which requires models to understand both visual and textual content in long videos. - Developing MiniGPT4-Video, a model that extends VLMs to process multiple frames and improve content understanding. The authors conclude that their models and code have been made publicly available, and they hope that their work will benefit future research in long video understanding.