LingBot-Video is a 13B parameter sparse-MoE video diffusion transformer (1.4B active) that has been post-trained with reinforcement learning using six rewards, including a physical-plausibility reward, and supports action-conditioned world model prediction for robot rollouts. This work advances video generation and world modeling by combining sparse MoE for efficiency, RL post-training for improved plausibility, and open release of weights and code, making it a strong candidate for robotics and video synthesis research. The model uses a DeepSeek-V3-style sparse MoE with 128 experts and top-8 routing, achieving top average performance on RBench but ranking second on general text-to-video. The physical-plausibility reward is judged by a VLM, which raises concerns about reward hacking despite adding real-video negatives.
Background
Diffusion transformers (DiTs) replace convolutional U-Net backbones with transformer architectures for denoising in generative models. Sparse mixture of experts (MoE) activates only a subset of parameters per forward pass, enabling larger model capacity with lower computational cost. Reinforcement learning from reward models can fine-tune generative models for specific objectives, such as physical plausibility.