MolmoMotion: Language-guided 3D motion forecasting
Get smart on it
MolmoMotion is a model that predicts how objects will move in the future by taking a video frame, marked points on an object, and written instructions about an action, then forecasting where those points will move in 3D space over the next few seconds. Predicting motion is harder than simply observing motion that has already happened, but it is more useful for applications that need to plan ahead, such as robots reaching for objects or video generators creating physically plausible frames. The model uses a representation based on 3D points attached to objects in world space, which works across different object types and camera viewpoints without requiring full video rendering. Alongside the model, researchers are releasing a large dataset of 3D point trajectories paired with action descriptions from over 1 million videos, plus a human-validated benchmark for measuring forecasting accuracy.
Related news
Vibe coding platform Base44 launches own model as AI startups seek defensibility
Wix-owned vibe coding platform Base44 has started rolling out its own AI model — with hopes that it will eventually outperform frontier models.
DiScoFormer: One transformer for density and score, across distributions
DiScoFormer is a transformer model that estimates both the density and score of a data distribution in a single forward pass without requiring retraining for new distributions. The density describes where data points cluster, while the score, the gradient of log-density, points toward more probable regions and is used in diffusion-based generative models, Bayesian sampling, and scientific simulations. Existing methods force a trade-off: classical kernel density estimation works on any distribution but loses accuracy in high dimensions, while neural score-matching models stay accurate in high dimensions but must be retrained for each new distribution. DiScoFormer significantly outperforms kernel density estimation, cutting score error by 6.5 times and density error by more than 37 times in 100 dimensions, while generalizing to distributions with shapes and complexities not seen during training.
