Research map / Graph learning / GNN architectures
G-Merging: Graph Models Merging for Parameter-Efficient Multi-Task Knowledge Consolidation
Jun Chen, Ziyue Qiao, Qin Zhang, Kaize Ding, Xiao Luo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add parameter merging to build the unified model instead of using the pretrained model (ablation) | G-Merging · 8 molecular tasks (GIN contextpred) | average ROC-AUC (%): 72.2 -> 74.2 | Improved |
| add node-level MoE adapters (ablation) | G-Merging · 8 molecular tasks (GIN contextpred) | average ROC-AUC (%): 72.4 -> 74.2 | Improved |
| add graph-level MoE adapters (ablation) | G-Merging · 8 molecular tasks (GIN contextpred) | average ROC-AUC (%): 70.2 -> 74.2 | Improved |
| add the topology-aware Wasserstein distance loss (ablation) | G-Merging · 8 molecular tasks (GIN contextpred) | average ROC-AUC (%): 73.2 -> 74.2 | Improved |
| add the L1 distance (L_MD) loss (ablation) | G-Merging · 8 molecular tasks (GIN contextpred) | average ROC-AUC (%): 73.5 -> 74.2 | Improved |
| increase adapter rank from 10 to 30 (ablation) | G-Merging · 8 molecular tasks (GIN contextpred) | average ROC-AUC (%): 72.9 -> 73.9 | Improved |
| increase adapter rank from 30 to 50 (ablation) | G-Merging · 8 molecular tasks (GIN contextpred) | average ROC-AUC (%): 73.9 -> 73.8 | No clear effect |
| select more experts in top-k routing (k from 1 to 8) (ablation) | G-Merging · 8 molecular tasks (GIN contextpred) | average ROC-AUC (%) | Mixed |
| replace adapters-only inference (G-Merging-s) by MoE adapters | G-Merging · 8 molecular tasks (GIN edgepred) | average ROC-AUC (%): 73.1 -> 73.1 | No clear effect |
| replace adapters-only inference (G-Merging-s) by MoE adapters | G-Merging · 8 molecular tasks (GCN contextpred) | average ROC-AUC (%): 68.8 -> 68.9 | No clear effect |
| G-Merging versus best baseline merging method (EMR-Merging) | G-Merging · 8 molecular tasks (GIN contextpred) | average ROC-AUC (%): 71.5 -> 74.0 | Improved |
| G-Merging versus best baseline merging method (EMR-Merging) | G-Merging · 8 molecular tasks (GCN contextpred) | average ROC-AUC (%): 66.4 -> 68.9 | Improved |
Closest papers by meaning
- Towards Graph Foundation Models: Learning Generalities Across Graphs via Task-Trees (ICML 2025 · similarity 0.88)
- A Unified Graph Language Model for Multi-Domain Multi-Task Graph Alignment Instruction Tuning (NeurIPS 2026 · similarity 0.88)
- Towards Pre-trained Graph Condensation via Optimal Transport (NeurIPS 2025 · similarity 0.88)
- Measuring Task Similarity and Its Implication in Fine-Tuning Graph Neural Networks (AAAI 2024 · similarity 0.87)
- GraphBridge: Towards Arbitrary Transfer Learning in GNNs (ICLR 2025 · similarity 0.87)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
