VideoLatent: Video-Language Learning via Latent Self-Forcing

VideoLatent is a novel multimodal large language model (MLLM) designed for video understanding and reasoning, introducing a latent injection module that employs a latent self-forcing training paradigm. This model achieves significant computational efficiency, reducing training and inference overhead by approximately 6x and 68x, respectively, while outperforming existing MLLMs across 14 benchmarks in both general video understanding and complex reasoning tasks. Its reliance on standard video-question-answer triplets enhances scalability and transferability, making it a valuable tool for practitioners in the field of AI who require efficient video processing capabilities.