https://mllog.dev/en/tags/next-state-prediction/Recent content in Next-State-Prediction on MLLog.devhttps://mllog.dev/images/default_mllog.pnghttps://mllog.dev/images/default_mllog.pngHugo -- 0.147.9enSun, 05 Jul 2026 08:00:00 +0100https://mllog.dev/en/posts/orca-world-foundation-model-next-state-prediction/Sun, 05 Jul 2026 08:00:00 +0100https://mllog.dev/en/posts/orca-world-foundation-model-next-state-prediction/Orca (BAAI) replaces next-token, next-frame, and next-action prediction with a single Next-State-Prediction objective. A frozen 4B backbone pre-trained on 12.5K hours of video — with zero action labels — feeds three lightweight readouts, and the action readout, trained on just 200 trajectories per task, beats π0.5 on OOD robot manipulation (32.4 vs 29.4).