Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 13 papers for 2026-08-31
Nothing cleared the bar today. Papers read in full today: 3.
Proposes Probabilistic Dependency Graphs (PDGs) to fuse edge existence and orientation information from large language models with Bayesian network structure learning. In a PDG, each edge carries a distribution over directed, undirected, and absent states and is combined via weighted averaging. Evaluated on 26 benchmark networks using ensembles of BNSL methods.
Investigates aligning Reactome ontology with LLM-based semantic embeddings to scale curation of biological pathways. The work analyzes whether relationships among textual descriptions reflect higher-order biological relationships and proposes methods to bridge textual embeddings with ontological structure.
LLM-Augmented Causal Discovery: Probabilistic Fusion of Edge Existence and Orientation
Introduces hypergraph adaptive wavelet operators to learn solution maps for parametric PDEs beyond pairwise graphs. The method handles unstructured meshes and point clouds, capturing high-order interactions on hypergraphs to improve accuracy and stability in time-dependent simulations.
Proposes K-GAT, a knowledge-conditioned topology generation method for multi-agent systems. It uses evidence and domain knowledge to condition the evolving collaboration topology, reducing redundant interactions and improving verification in knowledge-intensive tasks.
Proposes a self-explainable multi-label graph neural network that jointly models label correlations and evidence sharing across labels. Unlike post-hoc explainers, it provides training-time interpretation and explicit attributions for correlated evidence. The approach improves both interpretability and predictive performance on multi-label graph tasks.
Develops temporal memory-aware online test-time adaptation for dynamic graphs. The approach equips dynamic GNNs with memory of past graph dynamics to adapt at test time under distribution shifts, improving robustness and efficiency during streaming inference.
Present DeepSeq3, a hierarchical graph representation learning framework for scalable analysis of sequential circuits. It abstracts circuits into fine-grained combinational subgraphs between flip-flops and a high-level Super-Node Graph that encodes register-level temporal dynamics, enabling scalable EDA tasks on large nets.
Expert Knowledge & Machine Understanding: Bridging Reactome's Ontology with LLM Semantic Embeddings
CheXtriev is a graph-based, anatomy-aware retrieval framework for chest radiographs. It uses graph transformers to extract region-specific features and model spatial context, yielding anatomy-centered representations that improve retrieval, especially for rare findings.
Proposes a framework that propagates construction-time triple quality signals, derived from clinical guidelines, into medical KGQA. The approach links knowledge-graph construction-time quality control with inference-time evidence usage to improve QA reliability.
Proposes QGPINNs, a physics-informed neural network framework for nonlocal differential equations on quantum graphs. Each edge is modeled by a neural network, with a unified graph loss enforcing governing equations and vertex transmission conditions, including continuity and Kirchhoff-Neumann constraints.
Propagating construction-time knowledge quality into medical question answering: A framework grounded in clinical guidelines