#world models

Researchers from General Intuition, Kyutai, and Epic Games released MIRA, a 5B parameter interactive world model trained on 10k hours of synthetic Rocket League data, enabling real-time 4-player gameplay at 20 FPS on a single B200 GPU. MIRA demonstrates the feasibility of large-scale world models for multiplayer, physically complex environments, potentially accelerating research in game AI, simulation, and reinforcement learning by providing a playable demo and open-source tools. The model uses a latent diffusion architecture to generate video frames conditioned on all four players' actions, and the team released a 1,000-hour dataset of 4-player gameplay, a technical report, and a playable online demo.

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.

Researchers at LingBot World have released an open-weights interactive world model that uses a MoBA attention mask (mixing bidirectional and autoregressive attention) and post-training distillation over long self-rollout trajectories to significantly reduce drift, demonstrated by a continuous 60-minute stable rollout. This work addresses a critical challenge in interactive world models—drift accumulation during long rollouts—which has limited their practical usability. The combination of MoBA attention and long-rollout distillation could enable more stable and reliable interactive simulations for applications like video games, robotics, and virtual environments. The MoBA attention mask dynamically schedules KV-cache to keep long rollouts tractable, and the model uses Plücker embeddings plus AdaLN for camera control. A limitation is that persistence is only in appearance, not identity, meaning regions leaving the context window are regenerated on revisit, not recalled. The weights are released under CC-BY-NC-SA license.