Answers are grounded in this paper and cite their sources.
Neural Architecture Generation (NAG) seeks to generate high-quality neural architectures more directly than conventional Neural Architecture Search (NAS), but existing diffusion-based approaches commonly rely on pretrained denoisers and do not naturally accommodate constrained discrete architecture spaces. This study proposes Discrete Graph Diffusion Evolution with Tabu Search (DGDETS), a framework with a training-free reverse-generation mechanism that represents architecture decisions as heterogeneous categorical variables and constructs reverse diffusion distributions directly from the current population. Fitness-weighted population members are combined with categorical diffusion compatibility to estimate clean-state distributions, from which discrete reverse transitions generate new architectures without a learned score network or graph denoiser. Constraint-aware projection repairs invalid samples while preserving posterior-supported structural decisions, and late-stage multi-start Tabu Search intensifies promising graph neighborhoods using short-term memory and aspiration. DGDETS is evaluated on NAS-Bench-101, NAS-Bench-201, NAS-Bench-301, and TransNAS-Bench-101, covering heterogeneous architecture representations and downstream tasks. The results show competitive or best-performing architecture quality across the evaluated benchmarks together with low generation time in a controlled same-hardware comparison with directly related diffusion-based NAG baselines. Ablation and sensitivity analyses further demonstrate the importance of fitness-conditioned reverse diffusion, feasibility projection, diversity preservation, and late-stage Tabu refinement. The implementation is available at https://github.com/DK-cpy/DGDETS .
No takes yet. Share an insight, caveat, or question.
Cai et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: