Tell Its Own Story: Personality-Guided Human-Centric In-depth Video Reasoning Based on Iceberg Theory

Despite advancements in Multimodal Large Language Models (MLLMs) for video understanding, human-centric video reasoning remains challenging due to complex emotion-behavior interactions and limited depth at cognitive level. Inspired by the Iceberg Theory, we argue that truly in-depth reasoning is fundamentally driven by internal personality traits and external interaction patterns, which guide motivations behind behaviors. Based on this insight, we introduce Event-Twin, a collaborative reasoning framework comprising two levels of agents: (1) Witness Agent at the perceptual level, which analyzes spatiotemporal context and inter-character interactions from a global and outside perspective; (2) Role-Play Agents at the cognitive level, which construct corresponding character twins by integrating personality, interaction, and memory, interpreting events from a fine-grained and inside perspective. During inference, the Witness Agent retrieves relevant memories according to the question, and activates Role-Play Agents, who answer the question from a first-person perspective, establishing a new tell its own story paradigm. Additionally, we propose the BMPBench, which stratifies human-centric Video QA pairs into the Behavior, Motivation, and Personality-guided layers. Extensive experiments demonstrate that EventTwin significantly outperforms existing baselines on BMP-Bench. To our knowledge, this is the first work to incorporate personality into video reasoning, thereby paving the way for human-centric in-depth video reasoning.