Graph generation stops leaning on task-specific heuristics: hazard-guided sampling cuts steps on molecular and synthetic graphs
Synopsis
Discrete flow matching (DFM) decouples graph-generation training from sampling, but existing graph DFM samplers often rely on task-specific heuristics and need sampling budgets comparable to diffusion, so the speedup potential is only partly realized. H-GLow decomposes the continuous-time Markov chain generator into hazards and conditional jump distributions, then uses both to plan graph edits and adjust sampling time to graph activity. Adaptable to pretrained graph DFM models, it reports better generation quality with substantially fewer sampling steps on molecular and synthetic graph benchmarks; a companion graph DFM model predicts hazards and conditional jumps directly, with a training objective that decomposes exactly into hazard matching and hazard-weighted jump prediction.
Interpretation
Graph generation can now decide when to advance sampling from how active the graph currently is, instead of following a fixed sampling budget. Earlier graph DFM samplers often relied on task-specific heuristics, substantially expanding the hyperparameter space, and their sampling budgets were comparable to diffusion-based approaches. On molecular and synthetic graph benchmarks, H-GLow reports improved generation quality with substantially fewer sampling steps.
The method splits the continuous-time Markov chain generator into hazards and conditional jump distributions, then uses these quantities to plan graph edits and dynamically adjust sampling time. That split lets sampling be scheduled by hazard and jump rather than by a separate heuristic rule for each task. The framework can be adapted to pretrained graph DFM models, changing the sampling strategy without retraining.
A companion graph DFM model built on conditional generator matching predicts hazards and conditional jumps directly. Its training objective decomposes exactly into hazard matching and hazard-weighted jump prediction. This model is evaluated alongside H-GLow on molecular and synthetic graph benchmarks.
Perspective
The demonstrated scope is generation quality and sampling-step comparisons on molecular and synthetic graph benchmarks, plus adaptation of the framework to pretrained graph DFM models.
The step reduction and quality gain are given as relative statements such as substantially fewer and better, with no concrete step counts or metric values; conclusions are limited to molecular and synthetic graph benchmarks.