FLAGG: Flexible Autoregressive Graph Generation

2025-12-31 · Source: JMLR · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

The FLAGG (Flexible Autoregressive Graph Generation) framework addresses limitations in existing deep graph generation methods by combining one-shot and sequential models. While one-shot methods struggle with large graphs and sequential methods underperform on smaller ones, FLAGG sequentially generates graph portions using one-shot models. This approach allows flexibility in choosing the sequential policy, which is defined by a stochastic node removal process that an Insertion Model learns to reverse. Evaluated with the DiGress one-shot model across various datasets and graph sizes, FLAGG demonstrates superior sampling quality compared to both one-shot and autoregressive baselines. This system, published in 2026, offers a unified solution for diverse graph categories.

Key takeaway

For machine learning engineers developing graph generation models, if you encounter performance issues with either pure one-shot or sequential methods across varying graph sizes, you should investigate the FLAGG framework. Its ability to flexibly integrate and autoregressively apply one-shot models offers a robust solution for improved sampling quality. Consider experimenting with FLAGG to enhance your models' applicability across diverse graph domains and topologies.

Key insights

FLAGG flexibly combines one-shot and sequential models to overcome limitations in deep graph generation across diverse topologies.

Principles

• Neither one-shot nor sequential methods universally apply to all graph categories.
• Combining generation methods can overcome individual performance limitations.

Method

FLAGG sequentially generates graph portions using any one-shot model. A stochastic node removal process defines the sequential policy, which an Insertion Model learns to reverse.

In practice

• Integrate existing one-shot models into an autoregressive framework.
• Tailor sequential generation policies via node removal.

Topics

• Graph Generation
• Autoregressive Models
• One-shot Models
• FLAGG Framework
• DiGress
• Deep Learning

Code references

• CognacS/flagg-graphgen

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• VPG: Visual Prefix Guidance for Autoregressive Image and Video Generation — VPG is a training-free inference method that improves autoregressive generation by strengthening prefix posterior support.
• Reward Weighted Classifier-Free Guidance as Policy Improvement in Autoregressive Models — RCFG enables autoregressive models to optimize new reward functions at test time without re-training.
• (1D) Ordered Tokens Enable Efficient Test-Time Search — 1D ordered tokens with coarse-to-fine structure significantly enhance test-time search in autoregressive image generation.
• Can We Build Scene Graphs, Not Classify Them? FlowSG: Progressive Image-Conditioned Scene Graph Generation with Flow Matching — FlowSG progressively generates scene graphs by jointly denoising discrete semantics and continuous geometry via hybrid flow matching.
• Autoregressive Boltzmann Generators — Autoregressive Boltzmann Generators (ArBG) enhance molecular system sampling efficiency by replacing normalizing flows with autoregressive models.
• Safe Autoregressive Image Generation with Iterative Self-Improving Codebooks — Unified multimodal models can self-identify and eliminate unsafe image generations through iterative codebook refinement.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by JMLR.