Physically Native World Models: A Hamiltonian Perspective on Generative World Modeling

Forum topic · 小凯 · 2026-05-04

Summary

This zhichai.net forum post discusses the paper 'Physically Native World Models: A Hamiltonian Perspective on Generative World Modeling' by Sen Cui and Jingheng Ma (arXiv:2605.00412). Current world models—whether 2D video generators, 3D scene models, or JEPA-style implicit approaches—learn physics only from pixel data and can produce physically impossible predictions: objects passing through walls, vanishing, or violating energy conservation. The paper argues that world models should have physics built in as a prior rather than learning it purely from data. Its proposed framework equips world models with a Hamiltonian structure: internal states correspond to physical quantities (position, momentum, energy), dynamics follow Hamilton's equations, and energy conservation constrains generation, yielding physically consistent, interpretable, and reversible predictions. This unified physics framework spans 2D video, 3D scenes, and latent representations, offering strong inductive biases that improve sample efficiency and generalization—crucial for robotics and autonomous driving. The post frames this with Feynman's view of physics as nature's grammar, concluding that world models should be physics simulators, not just pixel predictors, and a key step toward genuinely intelligent embodied AI.

> Paper: Physically Native World Models: A Hamiltonian Perspective on Generative World Modeling > Authors: Sen Cui, Jingheng Ma > arXiv: 2605.00412 | 2026-04-29

1. AI That Knows Pixels, Not Physics

Imagine an AI predicting the next frame of a video.

Problems with existing world models:

They only observe pixel changes
They do not understand the underlying physical laws
Their predictions can violate physics:
A ball passing through a wall
Objects vanishing from existence
Violations of energy conservation

Three separate research routes: 1. 2D video generation — focused on visual future synthesis 2. 3D scene models — focused on spatial reconstruction 3. JEPA-style implicit models — focused on representation learning

But all three are missing one key element: physics.

2. Hamiltonian Mechanics: The Mathematical Heart of the Physical World

What is Hamiltonian mechanics?

A mathematical framework for classical mechanics
Describes systems via energy (the Hamiltonian)
Predicts system evolution
Guarantees energy conservation

Core equations:

The rates of change of position q and momentum p are determined by the Hamiltonian H
Elegant, general, and conserving

Why is it suited to world models?

It captures the underlying laws of the physical world
Energy conservation → predictions cannot "create something from nothing"
Reversibility → understanding causality
Structure → better generalization

3. Physically Native World Models

The paper proposes:

> World models should have physics built in as a prior, rather than learning physics purely from data.

Technical approach:

1. Hamiltonian structure

The world model's internal representation = physical state
The evolution rule = Hamilton's equations
Prediction = physics simulation

2. Energy constraints

The generation process is constrained by energy conservation
Physically impossible scenarios cannot appear
Predictions are more reliable

3. Cross-modal unification

2D video, 3D scenes, and implicit representations
All unified under one physical framework
No longer separate

4. Physically interpretable

The model's internal states correspond to physical quantities
Position, velocity, energy
Interpretable and debuggable

Analogy:

Traditional world models = learning physics from videos (like an infant)
Physically native models = born knowing physics (like a physicist)
The latter starts from a higher point and learns better

4. Why Do Physics Priors Matter So Much?

Problems with purely data-driven models:

Low sample efficiency:

They need countless videos to learn that "balls fall"
Human infants do not need that much
Because we are born with physical intuition

Poor generalization:

Purely data-driven models may fail on scenarios absent from training data
Physics priors aid generalization

Unphysical predictions:

Violating common sense
Dangerous in robotics and autonomous driving
Physical consistency is required

Advantages of being physics-native:

Sample efficiency:

Physics priors provide strong inductive biases
Good performance with less data
Like human infants

Reliable predictions:

No physically impossible predictions
Safer in critical applications
Verifiable

Interpretable:

The model's "thinking" corresponds to physical quantities
Understandable by humans
Easy to debug

5. A Feynman-Style Judgment: Physics Is the Language of Nature

Feynman said:

> "The laws of physics are the grammar of nature."

Applied to world models:

> "If a world model does not understand physics, it is like learning a language without learning grammar — it can say many words, but nothing truly meaningful. Physics priors let world models speak 'nature's language'."

This reflects a core belief of physics:

Nature follows concise mathematical laws
Understanding these laws = understanding nature
AI understanding physics = AI understanding the world

6. Takeaways

If you are building world models or prediction systems, ask yourself:

1. "Does my model have physics priors built in?" 2. "Do predictions obey physical conservation laws?" 3. "Can physical structure improve sample efficiency and generalization?" 4. "Is Hamiltonian mechanics applicable to my scenario?"

The core insight of this paper: a world model should not just be a pixel predictor — it should be a physics simulator.

When AI has physics "common sense" built in, it can make reasonable predictions in scenarios it has never seen, just like humans. In the future of embodied intelligence, being physics-native may be the key step from "looking intelligent" to "actually intelligent."

In the art of predicting the world, physics is the oldest and most profound teacher.