PhiZero pioneers physical language, enabling AI world models to think like humans
In today’s landscape of large models, video generation is increasingly viewed as the foundational simulator for embodied intelligence. Yet, most mainstream approaches rely on direct pixel-space prediction, often resulting in unstable or inconsistent physical laws. To solve this critical industry challenge, our research team has officially introduced PhiZero , a new physical world model. Moving beyond traditional blind prediction, PhiZero innovatively incorporates "physical language," empowering AI to conduct explicit physical reasoning prior to video generation.
At its core, PhiZero introduces a discrete representation known as "physical language." Through self-supervised learning on vast outdoor video datasets, this language extracts state transition patterns, capturing dynamic evolution rules across diverse visual contexts. Leveraging this, the model adopts a "reasoning-first, then-rendering" architecture: it first autonomously infers future world evolution trajectories within the physical language space, then passes these to a specialized diffusion decoder for high-definition video synthesis. This separation of underlying dynamics simulation from pixel-level synthesis significantly enhances physical consistency.

Technically, the system comprises two key modules: a tokenizer and a reasoner. The physical language tokenizer employs a transition-level Q-Former to capture changes between adjacent states, generating compact discrete symbols via limited scalar quantization. The diffusion decoder then reconstructs detailed visuals based on the initial frame and these discrete symbols. Meanwhile, the physical language reasoner, initialized with a pre-trained vision-language model, accurately predicts future physical language sequences given the first frame and text-based action intents.

Benchmark evaluations confirm that PhiZero achieves state-of-the-art performance across multiple physical reasoning and video generation tests, including Physics-IQ Verified, PhyGround, and WorldModelBench. Whether handling object displacement from collisions, gravity-driven deformation, or complex chain reactions, the model exhibits highly realistic physical plausibility.
As embodied intelligence and robotics accelerate, the release of PhiZero paves a new path for controllable, interactive world modeling. It is poised to play a pivotal role in long-term simulation and motion transfer scenarios.
Related article
Nvidia’s Vera chip: The US$200 billion gamble Jensen Huang wants you to notice
The Nvidia Vera chip rarely grabs headlines when earnings exceed expectations, yet it deserves the spotlight. With Q1 revenue hitting US$81.62 billion—surpassing analyst estimates of US$78.86 billion—and a Q2 guidance of US$91 billion, well above Wal
MingTang Intelligent Unveils MiniCPM-V 4.7 for Edge Devices, Delivering 13+ Token/s on 3GB Memory
Mingshi Intelligence unveiled its latest edge-side multimodal large model, MiniCPM-V 4.7, available in two configurations: 35B-A3B and 1B. The 35B-A3B variant marks the first Mixture of Experts (MoE) model in the MiniCPM-V series, leveraging a propri
Gamma buys out Accel-backed Lica to boost design capabilities
Gamma, a leading presentation platform, has acquired Lica, an Accel-backed design startup, to establish a dedicated design research lab, according to exclusive reports from TechCrunch. Lica’s co-founders will spearhead this new initiative.Originally
Related Special Topic Recommendations
Comments (0)
0/500
In today’s landscape of large models, video generation is increasingly viewed as the foundational simulator for embodied intelligence. Yet, most mainstream approaches rely on direct pixel-space prediction, often resulting in unstable or inconsistent physical laws. To solve this critical industry challenge, our research team has officially introduced
At its core,

Technically, the system comprises two key modules: a tokenizer and a reasoner. The physical language tokenizer employs a transition-level Q-Former to capture changes between adjacent states, generating compact discrete symbols via limited scalar quantization. The diffusion decoder then reconstructs detailed visuals based on the initial frame and these discrete symbols. Meanwhile, the physical language reasoner, initialized with a pre-trained vision-language model, accurately predicts future physical language sequences given the first frame and text-based action intents.

Benchmark evaluations confirm that
As embodied intelligence and robotics accelerate, the release of
Nvidia’s Vera chip: The US$200 billion gamble Jensen Huang wants you to notice
The Nvidia Vera chip rarely grabs headlines when earnings exceed expectations, yet it deserves the spotlight. With Q1 revenue hitting US$81.62 billion—surpassing analyst estimates of US$78.86 billion—and a Q2 guidance of US$91 billion, well above Wal
MingTang Intelligent Unveils MiniCPM-V 4.7 for Edge Devices, Delivering 13+ Token/s on 3GB Memory
Mingshi Intelligence unveiled its latest edge-side multimodal large model, MiniCPM-V 4.7, available in two configurations: 35B-A3B and 1B. The 35B-A3B variant marks the first Mixture of Experts (MoE) model in the MiniCPM-V series, leveraging a propri
Gamma buys out Accel-backed Lica to boost design capabilities
Gamma, a leading presentation platform, has acquired Lica, an Accel-backed design startup, to establish a dedicated design research lab, according to exclusive reports from TechCrunch. Lica’s co-founders will spearhead this new initiative.Originally





Home






