AIAny
Icon for item

PhiZero: A World Model Built Around Physical Language

Learns a discrete “physical language” from unlabeled videos and uses a reason-then-render pipeline: predict compact state-transition tokens, then decode them into future video. Separates dynamics inference from pixel synthesis to improve physical fidelity, controllable simulation, and zero-shot motion transfer.

Introduction

Predicting how the physical world will evolve from video often forces models to bury dynamics inside high-dimensional pixel predictors. PhiZero flips that script: it learns a compact, discrete "physical language" that encodes state transitions, uses a pretrained visual-language reasoner to autoregressively predict those tokens given a current state and action intent, and then renders the predicted sequence into video with a diffusion decoder. The key insight is to make world evolution an explicit, appearance-disentangled reasoning target instead of a direct pixel regression.

Key Findings
  • Compact discrete representation: a tokenizer compresses state transitions into a small set of discrete symbols (256 tokens reported), so what: dynamics are represented sparsely and focus on change rather than appearance, reducing the burden on the visual decoder.
  • Reason-then-render decomposition: the model first infers future evolution in token space then renders frames with a pretrained diffusion decoder, so what: separates temporal/dynamic reasoning from high-fidelity appearance synthesis and enables more physically coherent outcomes.
  • Versatile use cases demonstrated: competitive results on video-generation and intuitive-physics benchmarks, plus action-conditioned rollouts and zero-shot motion transfer, so what: the learned interface supports controllable simulation and transferring motion across embodiments and visual domains.
  • Leverages pretrained visual-language priors: initializing the reasoner from a VLM injects commonsense and visual-semantic knowledge, so what: improves autoregressive prediction of plausible state sequences without heavy supervised labels.
Who it's for and trade-offs

Great fit if you need a video/world-model that emphasizes physical plausibility, controllability, and domain-transfer of motion (e.g., robotics simulation, action-conditioned forecasting, interactive rollouts). Look elsewhere if you require provably exact physics simulations, very long-horizon symbolic planning, or ultra-low-latency on-device inference: PhiZero relies on learned discrete tokens, pretrained diffusion decoders, and VLM priors, which trade strict physical exactness and extreme runtime efficiency for flexible, data-driven physical reasoning.

Where it fits

PhiZero sits between pixel-based future-video predictors and classical physics simulators: it offers more explicit, transferable dynamics representations than end-to-end pixel models while remaining data-driven and more visually realistic than analytic simulators. It is especially relevant to researchers working on video generation, intuitive physics, and embodied AI where appearance-disentangled control over state transitions is valuable.

Information

  • Websitearxiv.org
  • AuthorsShuyao Shang, Yuqi Wang, Ruopeng Gao, Xu Chen, Tieniu Tan, Lue Fan, Zhaoxiang Zhang
  • Published date2026/07/30

More Items

Converts text prompts into physically consistent videos by synthesizing executable Blender programs as a process-level chain-of-thought and using a dual-engine pipeline (deterministic simulation draft + draft-conditioned video editor). Ships with a VideoCoCo-3K draft–instruction–target dataset and shows substantial gains in physical-consistency benchmarks.

Evaluates multimodal context learning across grounding, new information application, and knowledge acquisition using a 3,443-instance benchmark spanning science, finance, long documents, spatial reasoning, and web VQA; finds current multimodal models perform poorly (best score 0.2847) and analyzes failure modes.

Evaluates whether vision-language models can make actionable decisions for a physical body by decoupling decision-making from low-level motor execution. Introduces HumanCLAW-Bench with 1,218 long-horizon egocentric episodes across 41 indoor scenes and diagnoses a lack of embodied self-awareness in current VLMs.