OpenNovelty.org

Modality-free Graph In-context Alignment

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
Graph neural networksIn-context learningPretraining
Abstract:

In-context learning (ICL) converts static encoders into task-conditioned reasoners, enabling adaptation to new data from just a few examples without updating pretrained parameters. This capability is essential for graph foundation models (GFMs) to approach LLM-level generality. Yet current GFMs struggle with cross-domain alignment, typically relying on modality-specific encoders that fail when graphs are pre-vectorized or raw data is inaccessible. In this paper, we introduce Modality-Free Graph In-context Alignment (MF-GIA), a framework that makes a pretrained graph encoder promptable for few-shot prediction across heterogeneous domains without modality assumptions. MF-GIA captures domain characteristics through gradient fingerprints, which parameterize lightweight transformations that align pre-encoded features and indexed labels into unified semantic spaces. During pretraining, a dual prompt-aware attention mechanism with episodic objective learns to match queries against aligned support examples to establish prompt-based reasoning capabilities. At inference, MF-GIA performs parameter-update-free adaptation using only a few-shot support set to trigger cross-domain alignment and enable immediate prediction on unseen domains. Experiments demonstrate that MF-GIA achieves superior few-shot performance across diverse graph domains and strong generalization to unseen domains. The code is anonymously available here.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces MF-GIA, a framework enabling pretrained graph encoders to perform few-shot cross-domain prediction without modality assumptions or parameter updates. It resides in the 'Graph In-Context Learning Frameworks' leaf, which contains only three papers total (including this one). This leaf sits within the broader 'In-Context Learning and Prompt Engineering for Graphs' branch, indicating a relatively sparse but emerging research direction. The small sibling count suggests this specific combination of modality-free alignment and in-context reasoning for graphs is not yet crowded.

The taxonomy reveals neighboring leaves focused on 'Prompt Learning and Design for Graph Tasks' (four papers) and 'LLM-Based Graph Learning and Prompt Engineering' (four papers), both emphasizing structured prompts and text-attributed graphs. MF-GIA diverges by avoiding modality-specific encoders and textual attributes, instead using gradient fingerprints to capture domain characteristics. The broader 'Graph Foundation Models and Cross-Domain Transfer' branch (nine papers across three leaves) addresses unified architectures and alignment-based transfer, but typically assumes accessible raw data or domain-specific components—constraints MF-GIA explicitly relaxes through its modality-free design.

Among fourteen candidates examined, none clearly refute the three core contributions. The MF-GIA framework itself was assessed against ten candidates with zero refutable overlaps. The gradient fingerprint-based domain embedder and dual prompt-aware attention mechanism each faced two candidates, again with no refutations. This limited search scope (top-K semantic matches plus citation expansion) suggests that within the examined literature, the specific combination of gradient-based domain parameterization and dual attention for in-context graph alignment appears novel. However, the small candidate pool means unexplored prior work could exist beyond these fourteen papers.

Given the sparse taxonomy leaf and absence of refutations among examined candidates, the work appears to occupy a relatively unexplored niche at the intersection of modality-free graph encoding and in-context learning. The analysis covers a focused subset of the field (fourteen papers from semantic search), not an exhaustive survey. Broader literature or domain-specific venues may contain relevant techniques not captured here, so this assessment reflects novelty within the examined scope rather than definitive field-wide originality.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
14
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: cross-domain in-context learning for graph-structured data. The field addresses how graph neural networks and foundation models can generalize across diverse graph domains by leveraging in-context examples, prompts, or transfer mechanisms. The taxonomy reveals several major branches: Graph Foundation Models and Cross-Domain Transfer focuses on building unified architectures that handle multiple graph types and domains, often through pre-training and adaptation strategies (e.g., Unigraph Natural Language[3], Kumorfm Relational Data[5]). In-Context Learning and Prompt Engineering for Graphs explores how to design structured prompts and few-shot demonstrations that guide graph models without extensive retraining, drawing inspiration from language model paradigms (e.g., GraphICL Structured Prompts[6], Topological Prompts[12]). Few-Shot Learning on Graphs and Cross-Domain Recommendation with Graph Learning tackle scenarios with limited labeled data or cold-start problems, while Domain-Specific Cross-Domain Graph Applications and Transfer Learning and Semi-Supervised Graph Methods address practical deployment in areas like fault diagnosis, molecular property prediction, and spatio-temporal forecasting. Surveys and Security Analysis provide overviews and examine vulnerabilities such as backdoor attacks in cross-context settings. Recent work has intensified around unifying graph representations and enabling rapid task adaptation. A central tension lies between domain-agnostic foundation models that aim for broad generalization versus specialized prompt-based or few-shot techniques that preserve task-specific structure. Modality-free Graph Alignment[0] sits within the In-Context Learning and Prompt Engineering for Graphs branch, specifically among Graph In-Context Learning Frameworks. It shares conceptual ground with Vector-ICL[8] and Prodigy[10], which also explore how to construct effective in-context demonstrations for graph tasks. Compared to these neighbors, Modality-free Graph Alignment[0] emphasizes aligning graph structures across modalities without relying on fixed feature encodings, offering a flexible approach to cross-domain adaptation. This contrasts with works like Unigraph Text-Attributed[4] that integrate textual attributes more explicitly, or Few-Shot Graph Networks[1] that focus on meta-learning strategies. The interplay between prompt design, structural alignment, and domain transfer remains an active area, with open questions about scalability, interpretability, and robustness across heterogeneous graph distributions.

Claimed Contributions

Modality-Free Graph In-context Alignment (MF-GIA) framework

The authors propose MF-GIA, a framework that enables graph foundation models to perform few-shot prediction across diverse graph domains without requiring modality-specific conversions or parameter updates. The framework achieves true in-context learning by satisfying three criteria: post-training-free inference, cross-domain alignment, and modality-free operation.

10 retrieved papers
Gradient fingerprint-based domain embedder

The authors introduce a method to capture domain characteristics using gradient fingerprints—single-step parameter updates that encode how a graph's features, labels, and structure influence a shared encoder. These fingerprints are used to generate domain embeddings that parameterize domain-conditioned transformations for aligning features and labels across heterogeneous domains.

2 retrieved papers
Dual Prompt-Aware Attention (DPAA) mechanism with episodic objective

The authors develop a dual prompt-aware attention mechanism that operates on both feature and label spaces, trained with an episodic objective that simulates few-shot scenarios. This mechanism enables the model to match query items against support examples for in-context reasoning without parameter updates during inference.

2 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Modality-Free Graph In-context Alignment (MF-GIA) framework

The authors propose MF-GIA, a framework that enables graph foundation models to perform few-shot prediction across diverse graph domains without requiring modality-specific conversions or parameter updates. The framework achieves true in-context learning by satisfying three criteria: post-training-free inference, cross-domain alignment, and modality-free operation.

Contribution

Gradient fingerprint-based domain embedder

The authors introduce a method to capture domain characteristics using gradient fingerprints—single-step parameter updates that encode how a graph's features, labels, and structure influence a shared encoder. These fingerprints are used to generate domain embeddings that parameterize domain-conditioned transformations for aligning features and labels across heterogeneous domains.

Contribution

Dual Prompt-Aware Attention (DPAA) mechanism with episodic objective

The authors develop a dual prompt-aware attention mechanism that operates on both feature and label spaces, trained with an episodic objective that simulates few-shot scenarios. This mechanism enables the model to match query items against support examples for in-context reasoning without parameter updates during inference.

Table of Contents