PARDiff: Bridging Autoregressive and Diffusion Models for Order-Agnostic Graph Generation

Arunava Roy, Usman Ahmad Usmani, Dipankar Dasgupta, Junzo Watada

Keywords: Autoregressive Graph Generation, Diffusion Process, Transformer

Abstract: Graph generation has long struggled with the trade-off between structural fidelity and permutation robustness: autoregressive models excel in expressivity but break under node-order sensitivity, while diffusion models offer invariance at the cost of directional coherence. We introduce PARDiff, a Progressive AutoRegressive Diffusion framework that unifies these strengths through block-wise, order-agnostic generation guided by learned structural decomposition. Unlike prior heuristics, PARDiff jointly predicts block sizes, ranks nodes, and applies an equivariant diffusion process to each block, aligning AR directionality with diffusion robustness. This reframes graph synthesis as probabilistic reasoning over learned topological partitions, enabling scalable, semantically faithful, and order-agnostic generation across molecular and non-molecular domains without auxiliary features. Experiments show state-of-the-art results on diverse benchmarks, while its modular, latency-aware design supports real-time applications like drug–drug interaction analysis, positioning PARDiff as a paradigm shift in structured generative modeling.

Primary Area: generative models

Submission Number: 8369