Showing 10 papers for 2026-04-27
Mochi uses a meta-learning training framework to align pre-training and inference for graph foundation models, aiming to unify tasks and improve training efficiency beyond traditional reconstruction-based pretraining. It challenges the assumption that representations learned via link prediction can be easily repurposed for downstream tasks through a separate unification step, showing limitations of simple alignment in both synthetic and real-world experiments.
This paper defines distance-misaligned training in Graph Transformers, where the model’s communication across graph distances does not match where label-relevant information resides, potentially hurting performance. It analyzes a contextual stochastic block model benchmark with mixtures of local and long-range signals to illustrate how such misalignment affects task outcomes.
WG-SRC provides a white-box signal-subspace probe for graph datasets by replacing learned message passing with a fixed graph-signal dictionary, enabling direct interpretation of which signals drive predictions. It serves as an operational fingerprint for datasets, aiding prediction and dataset diagnosis by revealing feature-level graph-learning mechanisms.
FixV2W introduces a lightweight approach that leverages knowledge graph embeddings and longitudinal trends to improve CVE-CWE mappings in the NVD, addressing inconsistent and incomplete mappings. The method systematically analyzes mappings to enhance accuracy for vulnerability management and risk assessment.
The paper combines physics-informed graph neural networks with extreme-value analysis to forecast long-range extreme rainfall in Thailand. By modeling gauge stations as a graph, it captures spatiotemporal teleconnections and offers explainability through these connections, after preprocessing relevant climate indices.
The work proposes a parameter-efficient conditioning mechanism for graph-based simulators to generalize across materials, using granular flows as a running example. It enables conditioning with a small number of parameters to adapt constitutive behavior, improving generalization to unseen geometries and materials.
Causal Concept Graphs (CCG) place a directed acyclic graph over sparse latent features in LLMs to model causal dependencies among concepts for stepwise reasoning. The approach combines task-conditioned sparse autoencoders for concept discovery with differentiable structure learning (DAGMA-style) to recover graphs and introduces the Causal Fidelity Score to evaluate the impact of graph-guided interventions.
Graph-to-Vision investigates multi-graph understanding and reasoning with Vision-Language Models, introducing the first comprehensive benchmark to evaluate VLMs on multi-graph reasoning beyond single-graph tasks. It outlines evaluation protocols and methods to enhance multi-graph reasoning abilities in VLMs.
The LLM+Graph workshop summary synthesizes progress on integrating large language models with graph data management and graph ML, outlining key directions, challenges, and proposed solutions for scalable algorithms and systems.
ASPIRE introduces adaptive spectral graph collaborative filtering to overcome low-frequency explosion bias in graph filters. It uses bi-level optimization to learn adaptive graph filters rather than relying on manually tuned hyperparameters, improving the effectiveness of spectral collaborative filtering.