Showing 21 papers for 2026-02-12
This work proposes a human-interpretable graph-to-text encoding that injects graph structure into text prompts to bridge LLM reasoning with graph problems. The encoding aims to preserve permutation invariance and complex relational structure so LLMs can perform graph reasoning more effectively.
We study sparsification as regularization for GNNs to reduce memory and compute costs in large graphs, using Erdős-Rényi-based strategies for adaptive sparsification. The approach is demonstrated on real-world grid reliability tasks (N-1 contingency) across three datasets of varying sizes, showing improved efficiency with competitive accuracy.
RiemannGL argues that Riemannian geometry provides a principled basis for graph representation learning, proposing a unifying framework rather than a collection of isolated methods. It treats non-Euclidean graph data in a geometric space to improve representation and learning.
SynergyKGC introduces an adaptive framework to reconcile topological heterogeneity in knowledge graph completion, addressing structural resolution mismatch across dense and sparse regions. It fuses pre-trained entity semantics with topology-aware synergy to improve robust relational reasoning.
We propose a method that leverages LLMs to accumulate and organize knowledge about graph-property–architecture relations, guiding automated GNN design. The approach mitigates knowledge gaps and noisy prompts and yields more effective GNN designs.
MOTGNN enables interpretable GNNs for multi-omics disease classification by integrating heterogeneous modalities with interpretable components and addressing class imbalance. It provides insights into modality contributions and biological relevance.
We propose Information-preserving Graph Neural Simulators (I-GNS) built on Hamiltonian dynamics to capture long-range interactions and reduce error accumulation in autoregressive rollouts.
EDT-Former uses entropy-guided dynamic tokens to align Graph-LLMs for molecular understanding, addressing fixed-token bottlenecks and neglect of stereochemistry. It replaces static tokens with dynamic tokens, improving efficiency without extensive LLM fine-tuning.
We present localized graph-based neural dynamics (GBND) for terrain manipulation, representing terrain states as a graph of particles and learning local dynamics to predict deformation and enable manipulation.
We introduce Intersecting Block Graph (IBG), a low-rank factorization of large directed graphs via intersecting bipartite components to efficiently approximate any graph, enabling faster learning on large graphs.
MalMoE tackles encrypted traffic detection under graph drift by using a Mixture-of-Experts architecture to adapt to evolving flow statistics and topology, improving robustness of detection.
KORAL integrates knowledge graphs to guide LLM reasoning for SSD operational analysis, addressing fragmented data and environmental factors; enables more informed diagnostic reasoning.
VulReaD employs a security knowledge graph as backbone for CWE-aligned vulnerability reasoning and detection, moving beyond binary predictions to semantic CWE-level explanations.
FeatureBench provides a benchmark for agentic coding with LLM-powered agents, covering more tasks beyond PR bug-fixing and offering automated, updatable evaluation coverage.
GraphSeek enables next-generation graph analytics with LLMs by planning over a Semantic Catalog instead of directly generating graph queries from NL, enabling scalable analytics on heterogeneous, evolving property graphs.
HeDA uses a knowledge graph to synthesize evidence from thousands of publications to map bio-ecological mediation of cascading heatwave risks, enabling improved inference of systemic risk pathways.
We unify deductive and abductive reasoning in knowledge graphs using a masked diffusion model, enabling joint retrieval of satisfying entities and generation of plausible hypotheses.
Structured sentiment analysis is framed as transition-based dependency graph parsing, leveraging transition systems from parsing to efficiently extract structured opinions.
This work argues for global-structure-aware summarization to improve coherence for long documents, proposing methods that model global structure beyond sentence-level pruning, potentially using LLMs with efficient architectures.
GeoGR is a generative retrieval framework for spatio-temporal POI recommendation, addressing sparsity and complex cross-category spatio-temporal dependencies, and aligning retrieval with large language model-based reasoning.
This paper proposes a boundary-aware multi-behavior dynamic graph transformer for sequential recommendation. It jointly captures evolving graph topologies and multi-behavior user interactions to address limitations of prior GNN- and transformer-based models. The boundary-aware design aims to better represent dynamic structural changes and cross-behavior signals to improve recommendation accuracy.