Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces the Meta-Principle Physics Architecture (MPPA), a neural network architecture that incorporates physical meta-principles (Connectivity, Conservation, Periodicity) to move beyond pure phenomenological fitting. MPPA uses specialized components like the Gravitator, Energy Encoder, and Periodicity Encoder to embody these principles, enabling a more structured and interpretable approach to learning. Experiments demonstrate that MPPA achieves significant improvements in physical reasoning, mathematical tasks, logical tasks, and perplexity compared to standard architectures, while also exhibiting strong out-of-distribution generalization.
Ditch the black box: embedding fundamental physics principles directly into neural architectures unlocks surprising gains in reasoning, math, and logic, suggesting a new path towards AI with genuine understanding.
The essence of current neural network architectures is phenomenological fitting: they learn input-output statistical correlations via massive parameters and data, yet lack intrinsic understanding of the fundamental principles governing physical reality. This paper proposes a paradigm leap from pure phenomenological fitting to the fusion of phenomenological fitting and endogenous deduction. By embedding physical meta-principles into neural network architecture, we construct the Meta-Principle Physics Architecture (MPPA). Specifically, MPPA embeds three core meta-principles - Connectivity, Conservation, Periodicity - into its architecture, implemented via three core components: the Gravitator realizes Connectivity via standard causal attention; the Energy Encoder implements Conservation via log-domain energy tracking and delayed compensation; the Periodicity Encoder fulfills Periodicity via FFT-based spectral analysis and delayed modulation. These components collaborate via a learnable independent gating fusion mechanism, forming a complete physical cognition framework of'local relational connectivity - global conservation constraint - evolutionary periodic law'. Experiments show MPPA achieves significant improvements: physical reasoning (from near zero to 0.436, 0.436 vs 0.000), 2.18x mathematical task improvement (0.330 vs 0.151), 52% logical task gain (0.456 vs 0.300), and 3.69% lower validation perplexity (259.45 vs 269.40), with only 11.8% more parameters (242.40M vs 216.91M). Notably, MPPA shows strong generalization on out-of-distribution physical scenarios, proving the robustness and interpretability of this principle-embedded design. This work establishes a new theoretical foundation and technical path for next-generation AI with physical common sense, causal reasoning, and mathematical rigor.