AXIOMATIC MANIFOLD SYNTHESIS: A NEURO-SYMBOLIC FRAMEWORK FOR STRUCTURAL REASONING IN DESIGN-TO-CODE GENERATION

Zihao Li, Ke Li, Wenliang Zhong, Rui Zhang

Abstract:

Addressing the persistent structural dissonance between fluid
visual intent and rigid syntactic constraints in automated
front-end engineering, we introduce a novel neuro-symbolic
framework termed Axiomatic Manifold Synthesis (AMS),
which moves beyond the limitations of standard autoregressive transformers. Unlike conventional sequence-to-sequence
paradigms that frequently suffer from hierarchical collapse
in deep nesting scenarios, AMS leverages a non-Euclidean
hyperbolic embedding space to map UI visual primitives onto
a latent topological manifold, effectively preserving the inherent recursive cardinality and parent-child dependencies
of Document Object Models (DOM). We formulate the code
generation process as a constrained path-finding problem
over a differentiable graph-symbolic bridge, incorporating
a unique ”Structural Entropy Regularization” objective that
penalizes redundant nesting and promotes semantic-to-syntax
isomorphism. By integrating a self-evolving feedback loop
that evaluates the perceptual fidelity of rendered artifacts
against the original design intent via a multi-modal contrastive loss, our approach transcends mere pattern imitation
to achieve authentic structural reasoning. Empirical evaluations across heterogeneous design-to-code benchmarks
demonstrate that AMS significantly mitigates the ”semantic
gap” and outperforms state-of-the-art large language models in generating production-ready, refactorable code with a
42% improvement in long-range structural consistency and a
substantial reduction in syntactic hallucination.