Westlake AGI Lab's Code World Model splits world rules (code) from video rendering
jiqizhixin · x · 2026-09-13
Westlake University's AGI Lab introduces Code World Model, addressing a key flaw in video-based training: video records the visible result of world evolution but not the rules and mechanisms producing it. Models must reverse-engineer collision, attack ranges, faction relations, and quest state from sparse visual consequences, while long-timescale changes happen off-camera—scaling video training adds seen results, not the white-box mechanics.
The method splits "how the world evolves" from "how the world is seen": a coding agent decides world evolution, code keeps rules executing persistently, and a video model converts the evolved state into high-fidelity visual observations.
More from Models
- Uncensored 'DeepSeek-V4.1-Flash' FP8 model trends on Hugging Face, unverified — dealignai · 2026-09-13
- Community picks LFM2.5-2.6B and MiniCPM5-2B as best sub-3B open-weight models — reach_vb · 2026-09-13
- User claims DeepSeek kept V4 Pro online under pressure, slowing next-gen training — teortaxesTex · 2026-09-13
- Forced Model Depreciations May Push Users Toward Open-Weight Alternatives — marlene_zw · 2026-09-13
- Abacus AI releases open-weight Smaug Mini 27B, prices Smaug Flash at $0.10/M input — bindureddy · 2026-09-13
- AkbasCore 2.0: damped multi-axis activation steering kernel validated on Qwen2.5-7B — Nearby_Indication474 · 2026-09-13