Showing 22 papers for 2026-07-21
We propose a graph-neural-network predictor for node-level structural manipulability in gate-level netlists. A topology-driven score is defined from path participation, k-core embedding, symmetry, and centrality and learned via node-level regression on directed graphs. Experiments on ISCAS85 and EPFL benchmarks compare several GNN architectures.
This work investigates using Graph Neural Networks as metamodels for supply chain optimization. It provides an initial dataset and architectures, and discusses how GNN surrogates can generalize across topologies while modeling both structure and parameters. The authors outline key directions for future research.
Node4All introduces a universal node representation learner that works across arbitrary graph datasets without dataset-specific optimization. It builds on the Channel Graph Transformer (CGT) and complementary ideas to promote cross-dataset generalization. The framework aims to avoid per-dataset tuning and enable ready-to-use representations for new graphs.
We provide a Weisfeiler-Leman characterization of global-attention graph transformers for MILPs. The analysis shows that a broad class of hierarchical transformers with global and cross attention is bounded by 1-WL in distinguishing MILP instances. This reveals limitations in expressiveness and guides design choices for MILP representations.
Sobek introduces streaming equivariant tensor-product convolutions that avoid materializing edge-specific intermediates. By reassociating computation, it reduces memory and bandwidth while preserving the operator's expressiveness. This makes larger, wider GNNs feasible on GPUs and reframes the bottleneck as a scheduling problem.
This survey reviews GNN-based link prediction methods, offering a new taxonomy that organizes techniques by architectures and graph structures. It discusses applications, evaluation settings, and practical challenges, and outlines directions for future research.
The paper studies how deep a GNN must be on sparse graphs, using the context of the sparse contextual stochastic block model. It links depth to the Kesten-Stigum threshold and reveals a depth-dependent phase transition in classification performance. The results provide practical guidance for designing deep GNNs in sparse regimes.
Taurus presents a single-machine system for out-of-core GNN inference on billion-scale graphs. It addresses RAM constraints with optimized I/O and memory management and supports exact full-graph inference. The approach delivers practical performance gains over existing distributed setups and reduces the need for large clusters.
mHC-GNN applies manifold-constrained hyper-connections to GNNs, expanding node representations across multiple parallel streams. Mixing matrices are constrained to the Birkhoff polytope via Sinkhorn-Knopp normalization, mitigating over-smoothing and boosting deep GNN expressiveness. The authors provide a layer-wise residual lower bound to support the approach.
DIB-OD preserves an invariant core for robust heterogeneous graph adaptation via a decoupled information bottleneck and online distillation. It aims to disentangle task-relevant invariant knowledge from domain-specific noise to improve cross-domain transfer. The framework supports pretrained GNNs for heterogeneous domains and enhances cross-domain robustness.
LATTICE is a graph-based self-supervised framework for multimodal spatial omics integration. It learns spot-level representations by harmonizing five aligned modality blocks (Visium RNA, scMultiome RNA/ATAC, spatial ATAC, and spatial CUT&RUN), yielding cross-modal embeddings for joint analyses.
Beyond semantic equivalence argues that LLM uncertainty should reflect logical relationships among answers, not just semantic agreement. Semantic entropy can misrepresent uncertainty when outputs are diverse but logically compatible. The paper proposes logical-graph-based uncertainty measures to better quantify uncertainty and reduce false hallucination flags.
PEARL provides a training-free framework that repairs noisy LLM-generated scientific reasoning graphs into auditable, semantically valid structures. It addresses edge labels, orientation, and source anchors to improve verifiability.
Debate-on-Graph equips LLMs with a debate-style mechanism to reason over knowledge graphs, helping identify and filter noisy or erroneous content. The approach yields more reliable and adaptive reasoning under uncertain KG data.
SAGA generates large-scale temporal benchmarks with a four-phase pipeline that decouples structure from semantics. Skeleton-first generation creates scale-free graph skeletons, while the dispatcher assigns semantic attributes to produce richly annotated temporal graphs with ground-truth anomalies.
DeLIVeR treats evidence retrieval as a reinforced exploration task driven by a Planner LLM that decomposes claims into targeted questions and traverses a knowledge graph for precise evidence. The retrieval policy is optimized with Group Relative Policy learning to improve veracity recognition.
MMGraphRAG presents an interpretable multimodal knowledge graph framework for vision-and-language tasks, unifying textual and visual modalities to enhance retrieval-augmented generation. The MMKG supports fine-grained cross-modal semantics and offers a blueprint for building and querying multimodal graphs.
RADD decouples retrieval and reranking for multi-modal KG completion with a retrieval-augmented discrete diffusion approach. A relation-aware multimodal KGE retriever serves as both global retriever and disambiguator, improving MMKGC performance.
SmartRAG proposes a fully on-device retrieval-augmented generation framework organized into Perception, Memory, Focus, and Thinking modules. It targets privacy, low latency, and offline capability for mobile intelligent assistants.
VL-KnG constructs spatiotemporal knowledge graphs from monocular egocentric video to act as persistent scene memory for embodied question answering. It is a training-free framework that processes video in chunks and maintains memory without 3D reconstruction to support long-horizon reasoning.
This paper tackles the challenge of adapting large language models to structural data such as knowledge graphs and web graphs. Existing approaches either convert graphs to text, incurring token overhead, or encode graphs with fixed-size representations that may lose detail. GRIP introduces in-parameter graph reasoning by fine-tuning LLMs, enabling graph-aware reasoning inside the model to reduce token overhead and scale to large graphs.
Given the widespread use of LLMs, robust attribution of authorship and detection of generated text is increasingly important. Rather than relying on surface linguistic features, this work introduces reasoning graphs to capture the internal reasoning structures of LLM-produced texts, which tend to persist even under paraphrase or obfuscation. By analyzing how the text's reasoning is constructed, the approach aims to provide more robust authorship signals for attribution.