In the current wave of large models, video generation models are often regarded as the future simulator for embodied intelligence. However, mainstream models mostly use direct prediction in pixel space, which can easily lead to deviations in the stability and consistency of physical laws. To address this industry pain point, the research team has recently officially launched a new physical world model PhiZero. This model abandons the traditional blind prediction approach and innovatively introduces the concept of "physical language," enabling AI to perform explicit physical reasoning before generating videos.
As a core innovation, PhiZero proposes a discrete representation called "physical language." This language extracts state transition patterns from massive outdoor videos through self-supervised learning, capturing dynamic evolution rules in different visual experiences. Based on this, the model builds an innovative architecture of "reasoning first, then rendering": the system first autonomously infers the trajectory of future world evolution within the physical language space, then hands it over to a dedicated diffusion decoder to render it into high-definition video. This design not only separates the underlying dynamics simulation from pixel-level synthesis but also significantly improves the model's performance in physical consistency.
In terms of technical implementation, the system consists of two core modules: a tokenizer and a reasoner. The physical language tokenizer uses a transition-level Q-Former to capture changes between adjacent states and generates compact discrete symbols by combining a limited scalar quantization mechanism. The diffusion decoder restores detailed visual details based on the first frame and discrete symbols. Meanwhile, the physical language reasoner is initialized using a pre-trained vision-language model, accurately predicting the future sequence of physical language based on the first frame and text action intentions.
Multiple benchmark test results show that PhiZero has achieved leading performance in multiple physical reasoning and video generation evaluations, including Physics-IQ Verified, PhyGround, and WorldModelBench. Whether it is object displacement caused by collisions, deformation driven by gravity, or complex chain reactions, the model demonstrates highly realistic physical plausibility.
With the accelerated deployment of embodied intelligence and robot technology, the release of PhiZero opens up a new path for controllable and interactive world modeling, and is expected to play a key role in scenarios such as long-term simulation and motion transfer.
