[论文解读] ARYA: A Physics-Constrained Composable & Deterministic World Model Architecture

Seth Dobrin, Lukasz Chmiel|arXiv (Cornell University)|Mar 22, 2026

Adversarial Robustness in Machine Learning被引用 0

一句话总结

tldr: ARYA 提出一个可组合、受物理约束、确定性世界模型，该模型由纳米模型和始终在线的代理构成，在若干基准测试上实现了零神经网络参数的最先进结果。

ABSTRACT

This paper presents ARYA, a composable, physics-constrained, deterministic world model architecture built on five foundational principles: nano models, composability, causal reasoning, determinism, and architectural AI safety. We demonstrate that ARYA satisfies all canonical world model requirements, including state representation, dynamic prediction, causal and physical awareness, temporal consistency, generalization, learnability, and planning and control. Unlike monolithic foundation models, the ARYA foundation model implements these capabilities through a hierarchical system-of-system-of-systems of specialized nano models, orchestrated by AARA (ARYA Autonomous Research Agent), an always-on cognitive daemon that executes a continuous sense-decide-act-learn loop. The nano model architecture provides linear scaling, sparse activation, selective untraining, and sub-20-second training cycles, resolving the traditional tension between capability and computational efficiency. A central contribution is the Unfireable Safety Kernel: an architecturally immutable safety boundary that cannot be disabled or circumvented by any system component, including its own self-improvement engine. This is not a social or ethical alignment statement; it is a technical framework ensuring human control persists as autonomy increases. Safety is an architectural constraint governing every operation, not a policy layer applied after the fact. We present formal alignment between ARYA's architecture and canonical world model requirements, and report summarizing its state-of-the-art performance across 6 of 9 competitive benchmarks head-to-head with GPT-5.2, Opus 4.6, and V-JEPA-2. All with zero neural network parameters, across seven active industry domain nodes spanning aerospace, pharma manufacturing, oil and gas, smart cities, biotech, defense, and medical devices.

研究动机与目标

• Demonstrate that a physics-constrained, deterministic world model can satisfy canonical world model requirements (state representation, dynamic prediction, causal/physical awareness, temporal consistency, generalization, learnability, planning and control).
• Propose a hierarchical system-of-systems of specialized nano models orchestrated by an autonomous cognitive daemon (AARA) to achieve scalable, efficient learning and planning.
• Introduce architectural safety through an Unfireable Safety Kernel to ensure persistent human control as autonomy scales.
• Show that the architecture offers linear scaling, sparse activation, selective untraining, and sub-20-second training cycles compared to monolithic models.

提出的方法

• Define five foundational principles: nano models, composability, causal reasoning, determinism, and architectural AI safety.
• Implement a hierarchy of nano models forming a system-of-systems orchestrated by AARA for sense-decide-act-learn loops.
• Institute the Unfireable Safety Kernel as an architecturally immutable boundary that governs every operation.
• Provide formal alignment between ARYA’s architecture and canonical world model requirements.
• Evaluate ARYA on seven active industry domain nodes and compare head-to-head with GPT-5.2, Opus 4.6, and V-JEPA-2 without using neural network parameters.

实验结果

研究问题

• RQ1 Can a physics-constrained, deterministic world model satisfy canonical world model requirements across diverse domains?
• RQ2 Does a hierarchical nano-model architecture with an autonomous cognitive agent enable scalable, efficient learning and planning?
• RQ3 Can an architectural safety kernel ensure persistent human control in increasingly autonomous systems?
• RQ4 How does ARYA perform against established baselines on real-world industry tasks without neural parameters?

主要发现

• ARYA satisfies canonical world model requirements including state representation, dynamic prediction, causal/physical awareness, temporal consistency, generalization, learnability, and planning/control.
• The architecture achieves sub-20-second training cycles and linear scaling through sparse activation and selective untraining.
• AARA enables continuous sense-decide-act-learn operation across a system-of-systems of nano models.
• The Unfireable Safety Kernel provides an architecturally immutable safety boundary that cannot be disabled by any component.
• ARYA reports state-of-the-art performance across 6 of 9 benchmarks relative to GPT-5.2, Opus 4.6, and V-JEPA-2, using zero neural network parameters.
• The approach spans seven active industry domain nodes: aerospace, pharma manufacturing, oil and gas, smart cities, biotech, defense, and medical devices.