Recent leaps in Vision-Language Models (VLMs) have unlocked impressive short-clip understanding, yet their capacity to reason across extended, continuous video—where meaning emerges slowly through evolving scenes, dialogue arcs, and subtle visual cues—remains largely unexplored. This project asks: what does “long-form multimodal understanding” truly entail, and how far do current VLMs really go? We will probe state-of-the-art models’ temporal memory, cross-shot cohesion, and narrative comprehension by systematically analyzing their behavior on hours-long documentaries, multi-episode series, and unedited real-world footage. We mainly focus on uncovering conceptual strengths and brittleness—identifying where today’s architectures succeed, where they falter, and which inductive biases or prompting strategies might extend their reach. The work is open-ended and research-intensive; it demands curiosity about video cognition, solid coding skills, and the persistence to dissect complex, ambiguous results.