ARYA Labs says its physics-based AI avoids a ceiling in world models

ARYA Labs says new mathematical proofs show the dominant world-model architecture has a structural limit that no amount of data, compute or capital can fix. The company is pairing that claim with a physics-grounded alternative aimed at robotics, industrial simulation and other safety-critical systems. Why it matters: - Venture funding has poured billions into world-model startups built on the same core architectural bet. - ARYA Labs says that bet has a mathematical ceiling, which could reshape how investors and operators judge AI systems for robotics, autonomous driving, industrial simulation and defense. - The question is not just model quality. It is whether a system can stay reliable over long horizons in the physical world. What happened: - ARYA Labs published two preprints that challenge the dominant world-model approach and introduce its own physics-based alternative. - One paper, Identifiability Without Gaussianity: Symbolic World Models (arXiv:2606.12471), establishes a structural limit on JEPA-style world models. - The second, ARYA: A Physics-Constrained Composable and Deterministic World Model Architecture (arXiv:2603.21340), presents the Physics-Grounded Symbolic Architecture, or PGSA. - The company says the work is available immediately and freely on arXiv. - The Lean 4 formalization of the four central theorems accompanies the identifiability paper. - Commercial licensing of PGSA is handled separately at ARYA Labs . The details: - The proof by Klindt, LeCun and Balestriero says JEPA-style models can recover true structure only when reality behaves like a bell curve drifting around a mean. - ARYA Labs says that assumption breaks for gravity, turbulence, markets, sensor noise and fluid dynamics. - The company argues the architecture introduces a built-in bias that data and compute cannot remove. - ARYA Labs calls the failure mode “right once, drift wrong.” - The company cites the May 2026 stable-worldmodel benchmark, which found planning success on a leading neural world model fell from about 50% in clean conditions to 12% after an agent color change and 6% after a background shift. - ARYA Labs says that performance declined quadratically with distractors. - PGSA represents the world through symbols and physical laws rather than learned statistical patterns. - The company says PGSA does not depend on a bell-curve world and does not drift as it runs. - ARYA Labs says PGSA has a per-step error floor of 10⁻¹⁶, at the limit of double-precision arithmetic. - The company says PGSA holds bounded error over a horizon of 10¹³ on non-chaotic systems with known laws. - ARYA Labs says statistical world models typically last hundreds to a few thousand steps. - The Lean 4 formalization has zero unfinished placeholders in the four central theorems, according to the company. - ARYA Labs says its target market includes the pharmaceutical and biotechnology development loop and the engineering and manufacturing invention loop. - The company says the product is a next-generation CAD/CAE and design-space-exploration platform powered by PGSA. - ARYA Labs says the platform reasons over physics, geometric and material constraints, and manufacturing steps. - The target applications include jet engines, spacecraft, medical devices, electronics and automobiles. Between the lines: - The company is trying to move the debate from scaling laws to architecture choice. - If the proofs hold up, more training data and larger models would not solve the core reliability problem for certain physical-world tasks. - That would matter most in settings where failure is expensive, hard to detect or dangerous. - ARYA Labs is also positioning itself against a wave of large funding rounds in the category, including World Labs, AMI Labs, Wayve, Figure, Skild AI and Physical Intelligence. What’s next: - ARYA Labs says technical briefings are available to qualified press, researchers and enterprise prospects on request. - The company is selling PGSA through separate commercial licensing. - The broader test will be whether outside researchers and enterprise buyers accept the proof, the benchmarks and the product claims in real deployments. - For now, the category faces a new question: whether world models need more scale or a different mathematical foundation. The bottom line: - ARYA Labs is arguing that the future of physical AI depends less on bigger models and more on a provably different architecture.