![[2602.19332] Training-Free Cross-Architecture Merging for Graph Neural Networks](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.19332] Training-Free Cross-Architecture Merging for Graph Neural Networks
Summary
The paper presents H-GRAMA, a training-free framework for merging heterogeneous Graph Neural Networks (GNNs), allowing efficient model integration without retraining.
Why It Matters
As GNNs become increasingly prevalent in various applications, the ability to merge models across different architectures without retraining is crucial for enhancing performance and reducing computational costs. H-GRAMA addresses the limitations of current methods by enabling effective cross-architecture merging, which can significantly improve the efficiency of GNN applications.
Key Takeaways
- H-GRAMA allows for merging GNNs of different architectures without retraining.
- The framework improves inference speed by 1.2x to 1.9x compared to traditional ensembles.
- It formalizes a Universal Message Passing Mixture (UMPM) for heterogeneous GNN layers.
- The approach retains high accuracy while enabling cross-architecture compatibility.
- This innovation can lead to more efficient deployment of GNNs in real-world applications.
Computer Science > Machine Learning arXiv:2602.19332 (cs) [Submitted on 22 Feb 2026] Title:Training-Free Cross-Architecture Merging for Graph Neural Networks Authors:Rishabh Bhattacharya, Vikaskumar Kalsariya, Naresh Manwani View a PDF of the paper titled Training-Free Cross-Architecture Merging for Graph Neural Networks, by Rishabh Bhattacharya and 2 other authors View PDF HTML (experimental) Abstract:Model merging has emerged as a powerful paradigm for combining the capabilities of distinct expert models without the high computational cost of retraining, yet current methods are fundamentally constrained to homogeneous architectures. For GNNs, however, message passing is topology-dependent and sensitive to misalignment, making direct parameter-space merging unreliable. To bridge this gap, we introduce H-GRAMA (Heterogeneous Graph Routing and Message Alignment), a training-free framework that lifts merging from parameter space to operator space. We formalize Universal Message Passing Mixture (UMPM), a shared operator family that expresses heterogeneous GNN layers in a common functional language. H-GRAMA enables cross-architecture GNN merging (e.g., GCN to GAT) without retraining, retaining high specialist accuracy in most cases in compatible depth settings and achieving inference speedups of 1.2x to 1.9x over ensembles. Subjects: Machine Learning (cs.LG) Cite as: arXiv:2602.19332 [cs.LG] (or arXiv:2602.19332v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.26...
