Showing 33 papers for 2026-06-03
This paper introduces GFFMERGE, a principled framework for closed-form merging of Graph Neural Force Field models. It leverages the linear structure of message-passing to cast merging as a convex embedding-alignment problem with an analytical solution, enabling adaptation to new chemical systems without retraining.
We propose a Riemannian graph foundation model equipped with neural vector bundles to learn transferable common substructures across graphs. The work investigates the structural transferability of graph motifs and provides theoretical and empirical insights into when and how substructures can be reused.
This study analyzes adaptation strategies for Graph Foundation Models and finds that message-tuning outperforms graph prompt tuning. It introduces a prismatic-space framework to measure adaptation capacity and examines the limits of graph prompts.
The work proposes a multi-scale hypergraph approach to capture higher-order brain connectivity beyond pairwise edges. By moving beyond predefined hypergraphs, it models relations at multiple scales to improve brain-disease classification.
We present a Graph Foundation Model that combines spectral parsing with prototype-guided spatial propagation to handle diverse graph structures. By aligning graph spectra and using prototypes to guide propagation, the model improves cross-graph transferability.
This work introduces a Multi-Modal Graph Neural Network with Transformer-guided Adaptive Diffusion for preclinical Alzheimer classification. The diffusion process aggregates information from multi-modal brain regions, with a Transformer guiding long-range dependencies for improved accuracy and interpretability.
The paper conducts a systematic stability analysis of knowledge graph embedding models across random seeds. It reveals that high-performing KGEMs can produce divergent predictions and unstable embedding spaces, highlighting seed sensitivity beyond standard rank-based metrics.
We propose Topology-Aware Gaussian Graph Repair to improve robustness of GNNs to imperfect topology. The method repairs graph topology with a Gaussian-based mechanism to denoise and recover missing connections, leading to more reliable performance under noisy graphs.
HiSE is a lightweight hierarchical semantic explainer for heterogeneous GNNs. It aims to reveal semantic hierarchies in explanations with low computational cost, improving fidelity without expensive search or perturbation procedures.
This paper provides a mechanistic analysis of Graph Language Models, examining how internal LLMs interpret graph tokens and whether graph tokens carry meaningful structural information.
The authors present a limit-analysis of GNNs on wireless conflict graphs, establishing theoretical transferability results for graphs in sparse wireless networks and clarifying the conditions under which GNNs generalize from small to large scales.
TAGSAM is a text-attributed graph condensation method that compresses TAGs while preserving training accuracy. It introduces subgraph text selection and attribute similarity matching to reduce text and topological complexity.
The paper proposes Contrastive Neural Algorithmic Reasoning for Graph Coloring, enabling a neural approach to learn general graph-coloring strategies via contrastive learning rather than instance-specific optimization.
Visual Graph Scaffolds for Structural Reasoning in LLMs investigates using graphs as internal scaffolds to organize multi-step reasoning, akin to mind maps, improving performance on multi-hop question answering.
RelGT-AC is a Relational Graph Transformer designed for autocomplete tasks in relational databases, handling multi-table, heterogeneous, and temporal data to predict existing column values from relational context.
ReciNet leverages reciprocal-space representations and Fourier series to model long-range interactions in crystalline materials for property prediction, addressing periodicity and scalability.
TIDFormer introduces a dynamic graph transformer that explicitly encodes temporal and interactive dynamics, aiming to improve the effectiveness and efficiency of dynamic graph modeling.
The paper studies whether explanations amplify the risk of decision-logic leakage and model stealing in graph models, showing that explanation mechanisms can expose critical logic to attackers and discussing mitigations.
Learning the Neighborhood proposes C-FREE, a contrast-free, multimodal self-supervised pretraining framework for molecular graphs that integrates 2D topology and 3D structure via ego-nets, avoiding hand-crafted augmentations.
Towards Fair Graph Prompting introduces a dual-prompt mechanism to mitigate attribute and structural biases in graph prompting, aiming to improve fairness and reduce bias mismatch with downstream tasks.
This paper analyzes how graph structure affects membership inference risk in graph neural networks. It formalizes membership inference over node-neighborhood tuples and studies how structural factors influence attack success, highlighting graph-specific privacy concerns.
We propose Random Add-Drop Edge (RADE), a stochastic graph augmentation that both drops and adds edges during training. By regularizing both overfitting and long-range information flow (mitigating over-squashing) and improving train-inference alignment, RADE enhances GNN generalization.
SAHG introduces a sector-anisotropic hyperbolic graph model to better capture social graphs for bot detection. By leveraging hyperbolic geometry to represent hierarchy and scale-free structure, the approach reveals relational patterns and coordinated activity that Euclidean graphs miss, improving detection of social bots in large networks.
SkillDAG models inter-skill relationships as a typed directed graph and exposes a retrieval interface to an LLM agent, enabling structural skill selection at execution time. The framework evolves the skill graph automatically during use, addressing dependencies, conflicts, and duplication that are hard to capture with flat embeddings.
Code-on-Graph presents iterative programmatic reasoning over knowledge graphs via LLMs, addressing bottlenecks of predefined operators with limited expressiveness. The approach enables dynamic, programmable interactions with KGs, enabling more flexible and compositional reasoning for complex KG questions.
Generalizing Graph Foundation Models via Hyperbolic Retrieval-Augmented Generation argues that GFMs struggle under distribution shifts and proposes a hyperbolic RAG framework to inject external knowledge in a geometry-consistent way. The hyperbolic retrieval space better matches graph-structured data and improves robustness and generalization.
The paper proposes a training-free mixture-of-agents system for multi-document summarization that combines LLMs with knowledge-graph-based agents. It leverages the complementary strengths of LLMs and KG reasoning to handle inter-document relations without labeled training data.
This work shows how knowledge graph embeddings can approximate probabilistic inference in Statistical EL, providing runtime and soundness guarantees and empirical evaluation. It demonstrates efficient approximation with guarantees for a statistical extension using KG embeddings.
DTKG provides a dual-track framework for KG-verified reasoning in multi-hop QA, distinguishing parallel fact-verification and other reasoning patterns. It leverages relational KG structure to improve accuracy on complex multi-hop questions.
ReaLM tackles misalignment between KG embeddings and LLM token space using residual quantization to bridge the gap, enabling better semantic transfer from structured KG representations to LLMs for KG completion tasks.
Core-based Hierarchies for Efficient GraphRAG introduces a core-based hierarchical organization for documents in a knowledge graph used by RAG. It addresses Leiden clustering limitations on sparse graphs and enables better global sensemaking.
A Community Survey on SHACL and ShEx surveys RDF validation practices, collecting user experiences across academia and industry to identify benefits, limitations, and desired improvements to guide future research and standardization.
CacheRAG introduces a semantic caching system for retrieval-augmented KGQA that caches retrieval plans to avoid re-planning. This reduces schema hallucinations and improves retrieval coverage in RAG-based KGQA systems.