NewtPhys: Do Foundation Models Understand Newtonian Physics?

Paper: NewtPhys: Do Foundation Models Understand Newtonian Physics? arXiv: arXiv:2606.03986 | Authors: Sebastian Cavada, Soumava Paul, Tuan-Hung Vu, Andrei Bursuc, Raoul de Charette

🏀 Introduction: An Embarrassing Truth

Imagine standing before "the world's smartest AI" with a tennis ball. It can recite Newton's second law (F=ma), momentum conservation, and energy conservation flawlessly. But ask it how high the ball will bounce when dropped, and it hesitates—unable to intuitively answer that it will bounce slightly lower due to energy lost to heat and sound at each collision.

This is the paper's core question: Can AI that memorizes every physics formula actually understand Newtonian mechanics?

🎓 Background: The Gap Between Two Kinds of "Understanding"

• Symbolic understanding: knowing formulas, definitions, and theorems—solving textbook problems.
• Physical understanding: an intuitive grasp of how objects move through space, like knowing a ball rolling downhill accelerates without computing anything.

Human physicists typically have both. Past AI benchmarks mostly used simple, clean, synthetic images—testing symbolic knowledge rather than physical intuition, and leaving models unprepared for real-world messiness.

📊 The New Benchmark: NewtPhys

Key features:

1. Real-world scenarios — multi-view footage from real physical environments, with noisy backgrounds, imperfect objects, and uneven lighting. 2. 4D physical annotations — not just *what* objects are, but *how they move* (3D forces, trajectories, velocities) and *why* (force analysis). 3. Fine-grained time steps — capturing the entire motion process rather than start/end snapshots. 4. Cross-modal coverage — physics, tracking, semantics, and geometry, requiring models to answer what an object is, where it's going, why, and what happens on contact.

🔬 A Brutal Reality: 56 Models Collectively Fail

The authors evaluated 56 vision-language models (54 open-source plus 2 closed-source frontier models) and 10 visual foundation models (VFMs).

The result was harsh: nearly all models performed poorly on low-level physical reasoning—questions almost instinctive for humans:

• Does a ball rolling down a slope speed up?
• Is momentum conserved after two objects collide?
• What is a projectile's velocity at its apex?
• How does an object on a table move if the table is pulled away?

These aren't calculus problems—they're intuition problems a middle-schooler can answer. Yet state-of-the-art AI models frequently fail.

🧠 Why AI Became a "Nerd"

The paper doesn't deeply explore causes, but plausible inferences include:

1. Training data bias — the internet is full of physics formulas but contains little continuous video of real physical processes. AI learned "cat" as a concept but not the intuition of a bouncing ball. 2. Visual understanding ≠ physical understanding — models can describe static scenes but not dynamics, like predicting a ball rebounding off a table leg with energy loss. 3. Lack of causal reasoning — Newtonian mechanics is a causal chain (force → acceleration → velocity → position). Attention mechanisms capture correlation, not necessarily strict causality. 4. Missing scale — models may lack a sense of how numbers in formulas map to real-world magnitudes.

🏗️ Why Building NewtPhys Was So Hard

• Real scenes, not renders or physics-engine simulations, to capture subtleties like micro-textures, air turbulence, and nonlinear lighting.
• Precise physical annotations requiring expert knowledge and measurement equipment (gravity, friction, elastic forces per frame).
• Multi-view capture and 3D reconstruction of scenes and trajectories.
• Temporal continuity with precise time annotations, or momentum-conservation validation fails.

Building NewtPhys is itself an engineering feat: testing physical understanding is a hundred times harder than testing physics knowledge.

🌌 Meaning: Beyond Benchmarks

NewtPhys raises a philosophical question: what does "understanding" mean? If you can solve problems with F=ma but lack intuition about forces changing motion, do you truly understand mechanics? Today's AI resembles a Socratic knowledge repository—it "remembers" everything written on the internet—but lacks Aristotelian experiential learning: no body stumbling through the real world, learning how balls bounce or how heavy a table is.

🚀 Future Directions

1. More real-world physics video training — a large-scale dataset of real physical processes rather than synthetic scenes. 2. Embodied AI — giving models virtual or physical bodies that interact with the world, push objects, and catch balls. 3. Explicit causal modeling — learning causal chains (collision → force → acceleration → velocity change) rather than mere correlations. 4. Multi-scale learning — linking microscopic (molecular collisions) and macroscopic (object motion) physics.

📝 Conclusion: Newton's Apple and AI's Dilemma

The key to Newton's apple legend is that Newton observed the real world. Today's AI is like someone who has read every book about apples but never seen an apple tree—it can recite colors, weights, and photosynthesis, but cannot catch an apple thrown at it.

NewtPhys is a mirror exposing AI's blind spots: accumulated knowledge is not understanding; memorized formulas are not intuition. For AI to truly grasp physics, it may need to enter the real world and feel gravity's inescapability.

> "Nature does nothing in vain." — Newton

Reference: Cavada, S., Paul, S., Vu, T.-H., Bursuc, A., & de Charette, R. (2026). NewtPhys: Do Foundation Models Understand Newtonian Physics? *arXiv preprint arXiv:2606.03986*.