MAGE: Meta-path Adaptive Graph Evidence for Transferable Multimodal Graph Learning

Haiyang Jiang, Tong Chen, Xinyi Gao, Hao Wang, Hongzhi YinarXiv 2026 · arXiv

TL;DR

MAGE is a new framework for multimodal graph learning that enables models to generalize to entirely unseen domains (like moving from "Electronics" to "Books") without any fine-tuning. By using adaptive meta-path propagation to fuse text and images and selecting specific "evidence" neighbors to prompt a frozen Large Language Model (LLM), MAGE achieves state-of-the-art zero-shot performance on node classification and link prediction.

What Problem Does This Paper Address?

Multimodal graphs—where nodes contain both text and images (think Amazon products or Reddit posts)—are everywhere. However, training a model on one domain and applying it to another is notoriously difficult.

The authors identify two main hurdles:

  1. Modality Volatility: The "usefulness" of text versus images changes across datasets. On Reddit, text might be the best predictor of a post's category; on a book-review site, the cover art (image) might be more informative. Existing models often use fixed fusion strategies that fail when these priorities shift.
  2. Context Overload for LLMs: While LLMs are great at reasoning, they can't ingest an entire graph. Simply feeding an LLM every neighbor's text and image creates "noise" that obscures the useful signal. We need a way to select the right evidence and compress it into a format the LLM can actually use.

Methods

MAGE (Meta-path Adaptive Graph Evidence) introduces a two-stage approach to bridge the gap between graph structures and multimodal LLMs:

  • Heterogeneous Graph Splitting: MAGE treats the text and image of a single node as two separate, connected nodes. This allows the model to learn specific "meta-paths"—for example, a path that goes from a target's text to a neighbor's image and back to a text representation.
  • MAMP (Meta-path Adaptive Multimodal Propagation): This component uses eight different 2-hop meta-paths to aggregate information. Crucially, it uses a Semantic Prompt Router that looks at local graph statistics (like node degree and feature variance) to decide which paths are most important for a specific node in a specific domain.
  • Q-Former Compression: To keep things efficient, MAGE uses Q-Formers to "squish" long sequences of neighbor features into a small set of "soft tokens" that represent the graph context.
  • EGIT (Evidence-Guided Instruction Tuning): Instead of just asking the LLM about a target node, MAGE finds "evidence" nodes (similar neighbors with known labels) and includes their graph-aware soft tokens in the prompt. This allows the LLM to perform "in-context learning" by comparing the target to labeled examples.

Datasets

The researchers tested MAGE in a multi-source transfer setting.

  • Training (Source) Datasets: Toys, Reddit-S, and Amazon-Sports.
  • Evaluation (Unseen Target) Datasets: Movies, Grocery, Ele-fashion, and Goodreads.

These datasets span diverse categories from e-commerce to social media and books, ensuring the model's "zero-shot" capabilities were truly put to the test.

Key Insights and Ideas

  • Combining modalities isn't always better: The authors found that on some datasets (like Reddit-S), adding images actually decreased performance compared to text-only. This justifies MAGE's adaptive routing—the model learns when to ignore a noisy modality.
  • Meta-paths as "Cross-Modal Bridges": By splitting nodes into text and image components, MAGE can capture interactions that standard Graph Neural Networks miss, such as how a neighbor's visual style might clarify the intent of a target's ambiguous text.
  • Analogical Reasoning via Evidence: The LLM performs much better when it can see "evidence." By providing the LLM with a few labeled neighbors that are structurally and semantically similar to the target, the task shifts from "blind prediction" to "pattern matching."

Key Results

MAGE consistently outperformed competitive baselines (like MLaGA and Mario) across all unseen domains:

  • Node Classification: MAGE improved accuracy over the strongest baseline by 9.33% on the Goodreads dataset and 6.12% on Ele-fashion.
  • Link Prediction: In terms of Macro-F1 (a measure of balance between precision and recall), MAGE saw a massive 16.17% jump on the Movies dataset.
  • Zero-Shot Efficiency: The model achieved these results without updating a single parameter on the target domains, proving it can be deployed to new graph types "out of the box."

Limitations and Future Work

  • Scaling Evidence Retrieval: As graphs grow to millions of nodes, the cost of searching for the "top-K" most similar evidence nodes could become a bottleneck. The authors note that while preparation time scales linearly, it is a factor to watch.
  • Fixed Meta-path Length: Currently, MAGE relies on 2-hop meta-paths. Some complex relational reasoning might require deeper (3+ hop) traversals, which would increase computational complexity.
  • LLM Dependency: The performance is ultimately capped by the underlying multimodal LLM (in this case, Qwen2.5-VL). As base models improve, MAGE will likely get stronger, but it also inherits any biases or hallucinations present in the LLM.

Summary updated

Abstract

Multimodal graphs integrate textual and visual attributes with relational structures, providing a rich representation of complex real-world data. To enable the broad applicability and reuse of graph models, it is crucial for models trained on source graphs to generalize to unseen graph domains. However, this is challenging because cross-modal interactions among neighbors and the neighboring information needed for prediction can vary across target nodes and graph domains. This gives rise to two key challenges: adaptive cross-modal neighborhood fusion, which captures and integrates cross-modal interactions among neighbors to accommodate variations across nodes and domains; and target-specific evidence selection, which identifies the neighboring textual and visual information needed for each target node to support LLM-based prediction. In light of these challenges, we propose MAGE (Meta-path Adaptive Graph Evidence Prompting), a framework for transferable multimodal graph learning. MAGE represents textual and visual node views separately in a heterogeneous graph and introduces two complementary components. Meta-path Adaptive Multimodal Propagation (MAMP) adaptively aggregates neighborhood information along heterogeneous meta-paths and compresses it into graph-aware soft tokens. Evidence-Guided Instruction Tuning (EGIT) selects relevant neighboring evidence for each query and incorporates its graph-aware representations into instructions for a frozen multimodal LLM. Together, these components support transfer from multiple source graphs to unseen target graphs without target-domain parameter updates. Experiments with three source graphs and four unseen target graphs demonstrate strong cross-domain transfer performance on node classification and link prediction, without requiring parameter updates on the target domains.
BibTeX
@misc{jiang2026mage,
  author = {Haiyang Jiang and Tong Chen and Xinyi Gao and Hao Wang and Hongzhi Yin},
  title = {{MAGE: Meta-path Adaptive Graph Evidence for Transferable Multimodal Graph Learning}},
  year = {2026},
  eprint = {2611.20449},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.20449},
}
Figure