Recent Multimodal Large Language Models (MLLMs) have made rapid progress in video understanding, demonstrating strong performance across a wide range of perceptual and reasoning tasks. However, their Social Intelligence remains underexplored and difficult to evaluate. Existing benchmarks for social understanding suffer from two major limitations: they primarily evaluate short-horizon interactions and focus largely on explicit perceptual cues, such as facial expressions and speech sentiment. However, real-world social understanding often requires inferring implicit emotions, intentions, beliefs, and interpersonal dynamics from extended contextual interactions rather than isolated expressions or utterances. Consequently, robust evaluation of social intelligence requires long-horizon reasoning over complex social contexts. To address this gap, we introduce SMMU, a Massive Multimodal Social Understanding benchmark. SMMU comprises 264 long-form audiovisual scenes and 1519 question-answer pairs, to systematically evaluate each model across five social dimensions (Emotions, Intent, Knowledge State, Perspective, Relationships) and three cognitive tasks (Comprehension, Reasoning, Prediction). SMMU is designed to require models to integrate contextual evidence across extended interactions in order to interpret, reason about, and anticipate complex human behavior. Evaluating a suite of state-of-the-art MLLMs, we find that even the strongest models achieve only 0.43 Social Understanding Score compared to the average human baseline of 0.85, and 0.33 Social Agency Score compared to the human baseline of 0.78, revealing a substantial gap between current MLLMs and human social intelligence.

Recent video generation models can produce smooth and visually appealing clips, but they often struggle to synthesize complex dynamics with a coherent chain of consequences. Accurately modeling visual outcomes and state transitions over time remains a core challenge. In contrast, large language and multimodal models (e.g., GPT-4o) exhibit strong visual state reasoning and future prediction capabilities. To bridge these strengths, we introduce VChain, a novel inference-time chain-of-visual-thought framework that injects visual reasoning signals from multimodal models into video generation. Specifically, VChain contains a dedicated pipeline that leverages large multimodal models to generate a sparse set of critical keyframes as snapshots, which are then used to guide the sparse inference-time visual-state adaptation of a pre-trained video generator only at these key moments. Our approach is tuning-efficient, introduces minimal overhead and avoids dense supervision. Extensive experiments on complex, multi-step scenarios show that VChain significantly enhances the quality of generated videos.