Graph Energy Matching: Transport-Aligned Energy-Based Modeling for Graph Generation

Researchers have introduced new frameworks for graph generation, including Graph Energy Matching (GEM) and FLAGG, to address limitations in scalability and novelty.

Current graph generative models are limited by scalability and novelty, with diffusion-based methods requiring costly full-adjacency operations and long denoising chains^[2]. Autoregressive and hybrid models often have at least quadratic complexity^[2]. To address these issues, a new lightweight autoregressive framework was proposed, using a structure-guided topological ordering to serialize graphs into regular edge sequences^[2]. Additionally, the framework combines exploration-oriented augmentation with iterative refinement to reduce overfitting and promote controlled novelty^[2]. GEM, introduced on March 24, 2026, and revised on May 28, 2026^[1], is a discrete generative framework inspired by the Jordan-Kinderlehrer-Otto (JKO) transport-map optimization perspective. It learns a permutation-invariant potential energy that guides discrete transport from noise toward high-likelihood graph regions^[1]. GEM matches or surpasses strong discrete diffusion baselines on most reported metrics on molecular graph benchmarks^[1]. FLAGG, another framework, combines one-shot and sequential models to overcome limitations of each method, using a stochastic node removal process to specify the sequential policy^[3]. FLAGG outperforms one-shot and autoregressive baselines in terms of sampling quality^[3].