Research map / Graph learning / GNN architectures

Vision Graph Prompting via Semantic Low-Rank Decomposition

ICML 2025 · Graph learning › GNN architectures · approach: Graph structure learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Vision Graph Prompting (VGP), a parameter-efficient prompting method for Vision GNN (ViG) backbones that uses semantic low-rank decomposition in three prompt types (graph, edge, node), matching full fine-tuning on vision and graph benchmarks while training about 5% of parameters.

Problem

Existing visual prompting methods are designed for Transformer backbones and ignore the graph topology of Vision GNNs, so adapting pre-trained ViG models to downstream tasks lags behind full fine-tuning.

Approach

Observes that semantically connected nodes in ViG graphs share principal PCA components (low-rank semantics), and builds three low-rank prompts: virtual prompt nodes linked by KNN to the graph (SeLo-Graph), a low-rank edge-message prompt (SeLo-Edge), and a low-rank node prompt (SeLo-Node), all trained with the backbone frozen.

Key results

• VGP reaches 89.6% average accuracy on ten vision datasets with ViG-M, versus 89.0% for full fine-tuning and 91.3% for the best ViT-B InsVP, and about 5.0% average gain over prior visual prompting methods on ViG.
• On nine chemistry and biology graph datasets, VGP averages 76.39 versus 72.72 for full fine-tuning (a 3.67% gain) and surpasses prior methods on seven of nine datasets.
• Trainable parameters are reduced by 94.6% on average (2.61M vs 48.68M) with about 3.1% extra FLOPs.

Measured effects reported in the paper

Surprising findings

• On ViG, several visual prompting methods designed for ViT (e.g., VPT, VP) perform worse than adapter and underperform full fine-tuning, which the authors attribute to their reliance on attention.
• Linear probing on ViG-M reaches only 74.6% average, far below prompting variants, showing the graph-level prompt component contributes most of the gain (CUB +5.7%, GTSRB +9.5% from SeLo-Graph alone).

Limitations

• Rank threshold and low-rank analysis rely on PCA visualizations and a single error threshold (epsilon=0.25); the theoretical justification is informal.
• Graph-task results rely on the assumption that low-rank patterns exist in chemistry and biology graphs, which the authors only hypothesize.
• Evaluation is limited to a single ViG-M backbone pre-trained on ImageNet-21k and a fixed 100-epoch schedule.

Experimental setup

Datasets: ImageNet-21k DTD CUB-200 NABirds Stanford Dogs Oxford Flowers Food-101 CIFAR-100 CIFAR-10 GTSRB SVHN MoleculeNet

Benchmarks: DTD CUB NABirds Stanford Dogs Oxford Flowers Food-101 CIFAR-100 CIFAR-10 GTSRB SVHN BBBP Tox21

Metrics: classification accuracy ROC-AUC trainable parameters (M) FLOPs (G)

Baselines: Full Fine-Tune VP VPT VFPT DAM-VP InsVP GraphPrompt GPF-Plus GPPT GPF Adapter Linear probing

Builds on: Vision GNN VPT DAM-VP InsVP GPF-Plus

Closest papers by meaning

• GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks (NeurIPS 2025 · similarity 0.88)
• HeterGP: Bridging Heterogeneity in Graph Neural Networks with Multi-View Prompting (AAAI 2025 · similarity 0.87)
• The Underappreciated Power of Vision Models for Graph Structural Understanding (NeurIPS 2025 · similarity 0.87)
• Open Your Eyes: Vision Enhances Message Passing Neural Networks in Link Prediction (ICML 2025 · similarity 0.87)
• Hypergraph Vision Transformers: Images are More than Nodes, More than Edges (CVPR 2025 · similarity 0.87)
• Edge Prompt Tuning for Graph Neural Networks (ICLR 2025 · similarity 0.87)