Continuing the series of posts on world models - PhiZero from CASIA is a research model but it introduces a novel idea of physical language.
The problem it starts from is how world models represent change.
Models such as Cosmos 3 generate future video through continuous video latents. Cosmos 3 already separates an autoregressive Reasoner from a diffusion Generator, but the detailed physical evolution of the scene is still represented inside the high dimensional generative path.
PhiZero introduces another representation between reasoning and rendering.
This is how it does it:
1. A tokenizer turns video dynamics into discrete symbols.
2. A four second, 33 frame video is encoded into nine latent states. For every pair of adjacent states, a Q-Former with 32 learned queries extracts features describing what changed ( transitions)
3. Eight transitions with 32 features each gives 256 transition features.
Finite Scalar Quantization then converts them into a vocabulary of 25K discrete symbols.
The individual symbols do not have human readable meanings such as "move left" or "object falls." Their meaning is learned and distributed across the sequence. But the sequence is explicitly trained to represent how the scene changes rather than its appearance.
A reasoner predicts this physical language.
The reasoner starts from Qwen3-VL-4B. PhiZero adds the 25K physical symbols to its vocabulary and trains it to predict the 256-token sequence from the first frame and a text description of the initiating action.
This part uses the standard autoregressive LLM methodolgy
next physical token → next physical token → next physical token.
Once the physical tokenizer has been trained, it is frozen and used to generate the target sequences for reasoner training.
A renderer turns the physical language back into video.
The video decoder is initialized from Wan2.2-5B. The first frame provides scene information. The physical tokens describe the transition. The decoder is LoRA fine-tuned to reconstruct the future video using flow matching.
At a high level this means
current state → physical-language reasoning → video rendering
instead of asking the video generator itself to carry all of the detailed dynamics representation.
Some thoughts
1. The vocabulary is lossy representation of transitions.
2. If this type of architecture is successful then the reasoner can be trained using ordinary autoregressive method rather than requiring diffusion training.
3. There is no symbol/vocabulary for laws of physics like gravity etc.