What does this research mean for the field?

Neural world models can achieve zero-shot compositional transfer across physical environments by using a framework that combines residual state prediction, shuffled training, and symbolic law extraction, outperforming traditional monolithic models. Novelty: ClaimNovelty.NOVEL_FINDING. Consensus alignment: ConsensusAlignment.CHALLENGES_CONSENSUS.

What question did this study set out to answer?

The research aims to explore zero-shot compositional transfer in neural world models across different physical environments using innovative prediction methods.

March 12, 2026Open Access

Compositional Transfer in Neural World Models via Symbolic Law Discovery

Puntos clave

The research aims to explore zero-shot compositional transfer in neural world models across different physical environments using innovative prediction methods.
Developed a three-component framework including residual state prediction, shuffled training, and symbolic law extraction.
Applied ablation studies to compare residual state framing with absolute state prediction and evaluate training methods.
Conducted experiments across five classical mechanics environments to assess prediction accuracy and physical consistency.
Residual framing achieved a 47.3× improvement over absolute state predictions.
Shuffled training resulted in an 8.8× increase in physical invariance compared to sequential training.
In a 2D orbital mechanics environment, the model reached a mean squared error of 0.11 × 10⁻⁴, outperforming monolithic baselines.

Resumen

This paper investigates whether neural world models can achieve zero-shot compositional transfer across physical environments without any training data from the target combined system the gaols of physics training is that it can learn physics with observation. We demonstrate that by framing network predictions as residual state changes rather than absolute next states, isolated neural dynamics models trained on individual physical forces can be algebraically superposed to predict combined environments never encountered during training. We introduce a three-component framework combining residual state prediction, shuffled training for invariance learning, and symbolic law extraction via sparse regression (SINDy). Ablation studies show that residual framing provides a 47.3× improvement over absolute state prediction, while shuffled training achieves 8.8× better physical invariance compared to sequential training and recurrent architectures. Experiments across five classical mechanics environments demonstrate that the zero-shot ensemble composition achieves accuracy competitive with Neural ODE baselines trained on 100,000 environment-specific transitions. In a highly non-linear 2D orbital mechanics system governed by inverse-square gravity and quadratic atmospheric drag, the zero-shot composition achieves a mean squared error of 0.11 × 10⁻⁴, statistically outperforming monolithic baselines trained on 100,000 target transitions while exhibiting physically consistent manifold persistence where monolithic models suffer catastrophic escape velocity hallucination. Furthermore, we demonstrate that the compositional framework applies hierarchically, enabling the unsupervised discovery of the coefficient of restitution as a hidden physical constant through sub-environment decomposition. These results establish a mathematically grounded paradigm for modular world modeling where physical invariants are recovered rather than memorized, providing a scalable path toward structured and interpretable artificial general intelligence.

Leer artículo completoexternamente

Me gusta

Guardar

Ver artículo completo