On July 30, a team from the Institute of Automation, Chinese Academy of Sciences (NLPR/CASIA) released the preprint PhiZero (arXiv:2607.28624, project page phi-zero.github.io), proposing a new paradigm for physical AI, world models, and embodied intelligence: reason about the future in a "Physical Language" first, then render the reasoning result into video — instead of predicting pixels directly.
PhiZero achieves state-of-the-art results on four physics benchmarks:
- Physics-IQ Verified: IQ-Score 41.2 (ahead of Cosmos3-Super 39.5 / Grok-Video 34.8 / Sora 2 26.5)
- PhyGround: Physics Score 3.01 (ahead of Veo3.1 2.85 / Wan2.2-14B 2.90)
- WorldModelBench: Physics Adherence 4.88 (ahead of Wan2.2-5B 4.51 / Runway 4.27)
- IntPhys2: Overall 56.34 (ahead of Gemini-2.5 Flash 55.63 / GPT-4o 53.75)
- Visual fidelity keeps improving, but physical consistency has stagnated
- Excellent at short horizons (1–2 s), collapsing beyond ~5 seconds
- Predictions of "hidden states" (post-collision bounce, gravity-driven deformation, fluid motion) are barely learned in pixel space
- Visual details in training data saturate representational capacity, drowning out the "world state transitions" the model should actually learn
- Input: two adjacent video frames
- Encoding: Wan2.2 VAE to latent; a transition-level Q-Former extracts transition features between adjacent latent states
- Discretization: FSQ (Finite Scalar Quantization) compresses transition features into discrete symbols — a 25K-size atomic vocabulary
- Decoding: a Wan2.2-5B diffusion decoder renders the discrete symbols + first frame back to video
- Key trick: the diffusion decoder's pretrained prior recovers visual detail (texture, lighting) while the first frame anchors scene appearance, so the discrete bottleneck can focus on learning state transitions (motion, collision, deformation)
- Initialized from Qwen3-VL-4B (a vision-language model)
- Input: first frame + textual action intent
- Autoregressively predicts future physical-language token sequences
- Sequences are passed to the diffusion decoder for video rendering
- Initial pool: 50K hours of in-the-wild real video (YouTube / public datasets)
- Progressive filtering → 10K hours, used 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" high-quality clips, for reasoner SFT
- Before LLVM, compilers emitted assembly directly — unportable, hard to debug
- After LLVM, an intermediate representation (IR) decoupled frontends (language → IR) and backends (IR → hardware), collapsing porting costs
- Before PhiZero, world models were end-to-end video→video black boxes
- After PhiZero, there is an intermediate representation (Physical Language) that decouples the reasoner (language → physical language) from the renderer (physical language → video)
- New physical-language tokenizers (different quantization algorithms / visual encoders)
- New physical-language reasoners (different VLM initializations, different instruction formats)
- New physical-language renderers (different video backends, 4D Gaussian Splatting, NeRF, 3D)
- Shared physical-language vocabularies (a standard library of physical symbols)
- World model teams: evaluate whether PhiZero's tokenizer plugs into your video-generation backbone. The paper validated Wan2.2 + Qwen3-VL; components are likely reusable.
- VLA / embodied decision teams: physical language is a very lightweight intermediate representation that strips visual redundancy, keeping only state transitions — possibly more cost-effective than simply scaling up VLAs.
- Simulation / Sim2Real teams: reason-then-render lets "controllability of simulated scenes" and "realism of visual rendering" be optimized independently for the first time.
- AI for Science teams: discrete symbolic physical-language representations naturally align with atomic, molecular, and reaction symbol systems — a potential world-model + scientific-discovery intersection.
- Video model evaluators: PhiZero is a reminder that physical consistency and visual fidelity are independent optimization dimensions. A visually beautiful model doesn't understand physics. Consider dedicated benchmarks like Physics-IQ, PhyGround, and WorldModelBench.
- PhiZero project page (demos / videos): https://phi-zero.github.io/
- arXiv:2607.28624 HTML version: https://arxiv.org/html/2607.28624
- arXiv TLDR summary: https://arxivtldr.org/abs/2607.28624
- AI Base coverage (Chinese): https://www.aibase.com/news/30037
- The NextGen Tech Insider coverage (English): https://www.thenextgentechinsider.com/pulse/phizero-introduces-physical-language-for-advanced-world-modeling
- July 30 AI briefing collection: https://www.cnblogs.com/vibe234/p/22114799
This is the first time a world model beats this entire lineup of baselines on the "physical plausibility" dimension simultaneously — with a far smaller backbone resource budget.
The Problem: Direct Pixel Prediction Has Hit Its Ceiling
Over the past two years, the mainstream approach to world models has been direct pixel prediction of future video — VideoPoet, Genie, Sora, Wan2.2, Cosmos, and Veo are all variants. The ceiling of this paradigm became obvious in the first half of 2026:
PhiZero's angle: turn the human brain's habit of abstracting how the world works into model architecture. The paper opens with Sherlock Holmes's line — "You see, but you do not observe." The brain doesn't learn visual details first; it abstracts "predictive structure" from visual experience and uses language to organize that structure for explicit reasoning. LLMs proved symbolic reasoning spaces are scalable in the digital domain, but natural language is too coarse for physical state transitions.
So PhiZero introduces an intermediate layer — Physical Language: a compact, discrete representation of "world state transitions", learned self-supervised from unlabeled in-the-wild video.
Architecture: The Two-Stage reason-then-render Design
Module 1: Physical Language Tokenizer
Module 2: Physical Language Reasoner
The paradigm is named reason-then-render: future world evolution is reasoned out as discrete symbols in physical-language space, then handed to the diffusion decoder for rendering. Unlike mainstream pixel-space prediction, dynamics and rendering are decoupled.
Training Data: From 50K Hours of Raw Video to 5M 4-Second Clips
This "broad-data pretraining + curated-data fine-tuning" two-stage recipe mirrors mainstream NLP practice — a maturity signal for world-model methodology itself.
Why It Matters: A "Compiler Mindset" for World Models
PhiZero shifts world models from the "direct pixel prediction" paradigm to a "compiler" paradigm, echoing the last decade of programming-language evolution:
This suggests that within 12 months, an engineering ecosystem around "physical language" could emerge:
Once this ecosystem forms, world models become modular products — just as NLP split into composable modules (tokenizers, transformer blocks, task heads, prompt formats) after BERT/GPT.
Practical Advice for Embodied AI / World Model Teams
One-Sentence Summary
PhiZero moves world models from "directly predicting pixels" to "reasoning in physical language first, then rendering pixels" — just as programming languages moved from direct assembly to IR-then-assembly. The real significance is not a specific SOTA score, but that it makes the world model's intermediate representation concrete, standardized, and engineerable for the first time.
References