Programmable World Model cracks AI game state persistence with 94% accuracy
mhdfaran · x · 2026-09-10
A paper from Alaya Lab, "Programmable World Model", tackles the biggest weakness of AI-generated games: remembering what actually happened. Most video world models generate plausible next frames but lose persistent facts — NPC health, inventory, objectives, what's behind the camera.
The approach splits the system into two parts:
- Engine: maintains true world state via programmable rules like onhit → hp -= damage, staying true even off-screen;
- Video model: renders what that state should look like.
State-augmented 3D bounding boxes track each entity's identity, position, orientation and state, while the video model handles visuals and motion.
On their CombatStateBench, the method hits 94% character-count accuracy and 98% state accuracy, versus 32–41% and 8–58% for tested interactive world models.
Key takeaway: the future of AI games may be programmable state underneath plus generative video on top, rather than one giant model doing everything.
Related event: Alaya Lab's Programmable World Model Gives AI Games Persistent State(2 posts)→
More from Research
- DeepSeek V4.1 Flash is actually 748B params, safetensors analysis shows — DistanceSolar1449 · 2026-09-10
- Neural Harmonic Textures at ECCV beats all 3DGS and NeRF variants, code open-sourced — ZGojcic · 2026-09-10
- NeurIPS Learn2Design competition offers €10,000 for new optimization algorithms for physical systems — MarioKrenn6240 · 2026-09-10
- Salesforce paper: fine-tuning on evolved harnesses makes weak models worse on all 7 tasks — omarsar0 · 2026-09-10
- Gen2Balance: text-to-video synthesis fills long-tailed data for SOTA action recognition — dimadamen · 2026-09-10
- Hierarchical SfM paper and interactive VGGT point-cloud viewer rethink SLAM visualization — AjdDavison · 2026-09-10