Found first: a primary source the press has not covered yet.
Researchers at KAIST have published DreamFree, a world action model that eliminates iterative video generation at inference time. On the DROID real-world benchmark, it reaches 64% average success rate against π₀.₅'s 16%, running at 11Hz closed-loop control with a 14B backbone.
What the source says
The paper, submitted November 3, 2026 by Jaehyeong Jo, Sangwon Jang, Sihyun Yu, and colleagues at KAIST, trains a visual encoder to predict the internal representations of a frozen video diffusion model from current observations alone, removing the denoising rollout at inference time entirely. On the real-world DROID benchmark, DreamFree scores 64% average success rate versus 56% for DreamZero, the full-imagining variant, and 16% for π₀.₅. On RoboCasa-Kitchen simulation, the 14B backbone reaches 71.7%, compared to 44.7% for Fast-WAM with a 14B backbone and 67.1% for Cosmos Policy. LIBERO average success across four task suites is 98.6%. The inference speedup over DreamZero is 17.8×.
Why it matters
World action models have delivered strong simulation results by predicting future frames before choosing actions, but that generative step has kept inference too slow for real-time closed-loop control. DreamFree shows the predictive signal can be distilled into a fast encoder, reaching 11Hz on a 14B backbone while beating the imagining variant on actual hardware. On the real-world DROID benchmark, the gap over π₀.₅ is 64% to 16%. The paper also demonstrates learning from unannotated video across different robot embodiments, which bears on how far these models can scale without expensive action labeling.