PhiZero: A Reason-then-Render Paradigm Using Physical Language for World Models
PhiZero is a preprint released on July 30 by a team from the Institute of Automation, Chinese Academy of Sciences (NLPR/CASIA) (arXiv:2607.28624, project page: phi-zero.github.io). It introduces a new paradigm for physical AI, world models, and embodied intelligence: instead of directly predicting future video pixels, the model first reasons about the future inside a compact discrete "Physical Language" and then renders the symbolic prediction into video.
Benchmark Results
PhiZero reaches SOTA on four physics benchmarks:
- Physics-IQ Verified: IQ-Score 41.2 (vs. Cosmos3-Super 39.5, Grok-Video 34.8, Sora 2 26.5)
- PhyGround: Physics Score 3.01 (vs. Veo 3.1 2.85, Wan2.2-14B 2.90)
- WorldModelBench: Physics Adherence 4.88 (vs. Wan2.2-5B 4.51, Runway 4.27)
- IntPhys2: Overall 56.34 (vs. Gemini-2.5 Flash 55.63, GPT-4o 53.75)
- Visual fidelity keeps improving, but physical consistency plateaus
- Strong 1–2 second rollouts collapse beyond 5 seconds
- Hidden-state transitions (post-collision rebounds, gravity-driven deformation, fluid motion) are nearly impossible to learn in pixel space
- Visual detail consumes most of the representation capacity, drowning out the underlying state-transition signal
- Input: pairs of adjacent video frames
- Encoding: a Wan2.2 VAE maps frames to latents; a transition-level Q-Former extracts transition features between adjacent latent states
- Discretization: Finite Scalar Quantization (FSQ) compresses transition features into discrete symbols drawn from a ~25K-token atomic vocabulary
- Decoding: a Wan2.2-5B diffusion decoder takes the discrete symbols plus the first frame and renders the full video
- Initialization: Qwen3-VL-4B (a vision-language model)
- Input: first frame + text action intent
- Processing: autoregressive prediction of a short sequence of Physical Language tokens describing future dynamics
- Output: the predicted token sequence is handed to the diffusion decoder for rendering
- Initial pool: 50K hours of in-the-wild video (YouTube and public datasets)
- Progressive filtering → 10K hours for tokenizer pretraining
- Real + simulated video → 5 million 4-second clips for tokenizer SFT and reasoner pretraining
- Further filtering → 1 million motion-rich, physics-informative clips for reasoner SFT
- Pre-LLVM compilers emitted assembly directly — non-portable and hard to debug
- LLVM introduced an Intermediate Representation (IR), decoupling language frontends from hardware backends and collapsing migration costs
- Pre-PhiZero world models were end-to-end video-to-video black boxes
- PhiZero introduces an Intermediate Representation (Physical Language), decoupling the reasoner (intent → Physical Language) from the renderer (Physical Language → video)
- New Physical Language tokenizers (alternative discretization algorithms, alternative vision encoders)
- New Physical Language reasoners (different VLM initializations, different instruction formats)
- New Physical Language renderers (alternative video backends, 4D Gaussian Splatting, NeRF, 3D pipelines)
- Shared Physical Language vocabularies (standard libraries of physical symbols)
- World-model teams: evaluate whether the PhiZero tokenizer can plug into your existing video backbone. The paper already validates Wan2.2 + Qwen3-VL.
- VLA / embodied-decision teams: Physical Language is an extremely lightweight intermediate representation that strips visual redundancy and retains only state transitions. It may be a better investment than simply scaling VLA parameters.
- Simulation / Sim2Real teams: reason-then-render separation lets you independently optimize scene controllability and visual realism, which may matter more than visual polish alone.
- AI for Science teams: the discrete symbolic structure of Physical Language is naturally compatible with atoms, molecules, and reaction notations, opening a potential world-model × scientific-discovery direction.
- Anyone evaluating video generators: PhiZero shows that physical consistency and visual fidelity are independent optimization axes. Add Physics-IQ, PhyGround, and WorldModelBench to your evaluation suite.
- PhiZero project page (demos and videos): https://phi-zero.github.io/
- Paper HTML: https://arxiv.org/html/2607.28624
- arXiv TLDR: https://arxivtldr.org/abs/2607.28624
- AI Base (Chinese coverage): https://www.aibase.com/news/30037
- The NextGen Tech Insider (English coverage): https://www.thenextgentechinsider.com/pulse/phizero-introduces-physical-language-for-advanced-world-modeling
- Full 7-30 AI briefing: https://www.cnblogs.com/vibe234/p/22114799
It is the first world model to simultaneously beat Sora 2, Veo 3.1, Wan2.2, and Cosmos-3 on physics-grounded metrics, and it does so with a substantially smaller backbone.
Why Pixel Prediction Has Reached Its Ceiling
The dominant 2024–2026 paradigm — VideoPoet, Genie, Sora, Wan2.2, Cosmos, Veo — predicts pixels end-to-end. The limits of this approach are now visible:
PhiZero's insight, echoing Sherlock Holmes — "You see, but you do not observe" — is that the human brain first abstracts predictive structure from visual experience, then organizes it into a symbolic space for explicit reasoning. Large language models have already shown that symbolic reasoning spaces scale. Natural language, however, is too coarse for physical state transitions. PhiZero therefore proposes a dedicated intermediate representation: Physical Language, a compact, discrete vocabulary of world-state transitions learned self-supervised from unlabeled in-the-wild video.
Architecture: Reason-then-Render
The system splits into two modules.
Module 1 — Physical Language Tokenizer
The decoder's pretrained visual prior handles texture, lighting, and other appearance details; the first frame anchors scene appearance. This lets the discrete bottleneck focus exclusively on state transitions — motion, collisions, deformation.
Module 2 — Physical Language Reasoner
The authors call this overall design reason-then-render: future world evolution is first reasoned out as discrete symbols in Physical Language, then rendered into pixels. Dynamics and rendering are fully decoupled.
Training Data
This broad-pretrain + narrow-finetune recipe mirrors mainstream NLP practice, signaling methodological maturity in the world-model field.
Why It Matters: The "Compiler Era" of World Models
PhiZero reframes world modeling the same way LLVM reframed compilers:
Over the next ~12 months, this suggests an engineering ecosystem will form around Physical Language:
Once this ecosystem exists, world models will look like NLP after BERT/GPT — modular, composable, and productizable from tokenizers, reasoner blocks, and renderers rather than a single monolithic model.
Practical Recommendations
One-Line Summary
PhiZero moves world modeling from "predict pixels directly" to "reason in Physical Language, then render" — the same compiler-style shift as programming languages moving from raw assembly to IR-then-assembly.
---