MBZUAI releases open world model PAN 2: LLM planning plus 29B diffusion hits 75.1% step accuracy
rohanpaul_ai · x · 2026-11-11
MBZUAI's Institute of Foundation Models (IFM) has released PAN 2, an open foundation world model for physical intelligence built around "simulative reasoning" — letting an agent play out what happens before it acts.
The architecture splits the job in two:
- Conceptual reasoning: an LLM works out what happens, when, and how the camera moves, writing the plan as text.
- Perceptual reasoning: a 29B diffusion model trained from scratch (semantic/VAE encoders plus a latent diffusion backbone) renders that text into video observations.
Ablations back the split: removing the LLM collapses multi-step tasks to 8.7% step accuracy with zero completed plans, since ordering actions correctly is the LLM's job. With both components, PAN 2 achieves 75.1% step accuracy and fully completes 34.0% of plans — the best among models IFM tested. Details in the technical blog.
Related event: MBZUAI open-sources PAN 2, a 29B world model for simulation-based reasoning(2 posts)→
More from Embodied
- Fully DIY in-wheel-motor autonomous RC car runs SLAM on a ¥2000 LiDAR — ___Mufasaa · 2026-11-11
- AHA 2026: AI reads smartwatch ECGs for heart failure and speeds echo interpretation — AmiBhattMD · 2026-11-11
- SpaceX moves Super Heavy booster to the launch pad at Starbase — DimaZeniuk · 2026-11-11
- SkeLAM Uses Skeleton-Grounded Latent Actions to Decouple Motion from Appearance in Robot World Models — jiqizhixin · 2026-11-11
- Hub to open-source most diverse egocentric robot dataset on Hugging Face Nov 12 — alex_verem · 2026-11-11
- GTA V on Specs glasses: light trails and rain leap off the big screen into your room — stspanho · 2026-11-11