Showing 29 papers for 2026-07-29
We introduce CondPSE, a learned structural encoding approach for graphs. It applies a learnable polynomial graph filter bank to Gaussian node probes with conditional modulation, producing topology-informed signals that supplement message passing and go beyond 1-WL limitations. The encoder can be pre-trained and reused across downstream graph models to enhance performance on topology-sensitive tasks.
We analyze diffusion-based conformal prediction on graphs and identify a shortcoming: uniform low-pass diffusion assumes homophily and enlarges prediction sets on heterophilic graphs. We propose a heterophily-aware diffusion strategy that adapts to graph structure, achieving smaller, yet reliable, prediction sets on heterogeneous graphs.
This work questions whether learned restriction maps in Sheaf GNNs truly contribute to performance. It formalizes two estimands to separate checkpoint dependence from task-level value, by analyzing fixed-predictor reliance and protocol-relative retraining that removes map capacity. Results reveal when learned maps are genuinely beneficial versus incidental.
We present a contrastive representation learning framework for longitudinal disease trajectories modeled as temporal graphs. Nodes represent patient observations over time, and edges capture temporal continuity and trajectory similarity. Structure-aware random walks guide contrastive learning to produce robust representations for downstream clinical tasks.
We present the first systematic application of GraphRAG to raw earthquake catalogs, constructing knowledge graphs from tabular records across multiple catalogs. This enables retrieval-augmented generation for open-ended questions about earthquake sequences.
AdvSynGNN introduces a structure-adaptive graph neural network framework that uses multi-resolution structural synthesis and contrastive objectives to learn geometry-aware initializations robust to structural noise and non-homophilous topologies. A transformer backbone modulates attention to adapt to heterophily, improving performance under challenging Topologies.
We propose MapTab, a diagnostic benchmark to evaluate long-horizon, multi-criteria multimodal reasoning in multimodal LLMs. It tests models on route-planning tasks that require grounding map images and integrating route attributes.
We study vulnerabilities of topology-driven GNNs and propose PEANUT, an attack that perturbs eigenvector alignment to craft small structural changes that cause large output shifts. This reveals a critical attack surface for topology-aware message passing.
We model the ionosphere as a dynamic graph over pierce points whose topology evolves with satellite positions. Because satellite trajectories are predictable, we condition forecasts on future graph structure (ephemeris conditioning), improving predictive accuracy.
We study memory management for RL over temporal knowledge graphs under partial observability. We treat short-term-to-long-term memory transfer as per-item keep-or-drop decisions learned via Q-learning, enabling scalable, variable-sized buffers.
We propose EdgeRefine, a Jaccard-based sampling method to balance privacy and utility under edge differential privacy. By perturbing edges selectively rather than injecting global noise, EdgeRefine aims to improve privacy with preserved graph utility.
CityGuard presents a privacy-preserving identity search framework for decentralized surveillance using a topology-aware transformer. It includes a dispersion-adaptive metric learner and spatial conditioning to adjust margins, enabling bias-resilient re-identification without sharing raw imagery.
We identify calibration issues in prompt-tuned VLMs and propose ARGTCA, which represents class-attribute pairs as nodes in a symbolic attribute graph and uses a Graph Attention Network to reason over attribute relations for calibrated prompts.
We introduce TRWH, a Text-driven Random Walk Heterogeneous GNN that fuses LLM-generated textual profiles with heterogeneous graphs via strategic random walk augmentation to improve semantic-aware sparse recommendations.
HiSkill presents a hierarchical skill graph that organizes interaction trajectories into a directed graph, enabling structured skill reuse and composition across tasks for LLM agents.
CHARM introduces a multimodal graph foundation model that uses hierarchical context modeling to enable zero-shot transfer across multimodal graphs, improving cross-domain performance without task-specific labels.
DPR-GM introduces a domain-prior-regularized graph modeling approach for anomaly detection in cyber-physical systems, improving stability and reducing false alarms when labeled anomalies are scarce.
TraceBound provides a diagnostic protocol for adaptive knowledge-graph retrieval, exposing a compact query profile, post-failure hints, and trajectory counters to diagnose and improve retrieval.
We propose a dual-level graph representation that jointly learns atomic environments and coordination geometries (polyhedra) to improve crystal property prediction.
MedJudgeRAG integrates retrieval-augmented generation with a dynamic knowledge graph to perform option-wise evidence judgments for medical MCQA, combining retrieved documents and KG cues.
HVM-GraphRAG introduces a holistic-view multimodal GraphRAG framework for complex documents. It addresses unreliable cross-modal evidence indexing and expensive graph traversal by using a holistic view to guide graph construction and evidence retrieval, enabling more accurate retrieval-augmented generation across distant regions and modalities.
We introduce IRIS, a method to obtain reusable identity representations from frozen LLMs for entity alignment across knowledge graphs. By extracting deep semantic embeddings that remain stable across descriptions, IRIS can match the same real-world entity more accurately without heavy fine-tuning and distinguish it from similar entities.
Sheet As Token proposes a graph-enhanced representation for multi-sheet spreadsheet understanding. By connecting data across sheets with a graph, it preserves sheet-level semantics and mitigates fragmentation from chunking into rows or columns, enabling more coherent reasoning for LM-based analysis.
Neuro-Symbolic Meta-Policies for Temporal Knowledge-Graph Memory under Partial Observability introduces a neuro-symbolic meta-policy that learns which symbolic memory heuristic to apply at each decision point while keeping execution symbolic. Using RoomKG with RDF graphs and temporal annotations, it combines memory encoding and retrieval to improve decision making in partially observable environments.
EviDAG presents a browser-based system for auditable causal DAG authoring linked to biomedical literature. It enables reproducible literature snapshots and provenance-rich causal reasoning with LLM-based support.
This work compares RAG and GraphRAG for page-level retrieval QA on a math textbook. Using 477 QA pairs tied to specific pages, it benchmarks five embedding-based RAG models, BM25, and GraphRAG on retrieval accuracy and answer quality, offering insights into when GraphRAG is advantageous.
NeuroSymActive presents a differentiable neural-symbolic framework for KG question answering that supports active exploration. By integrating graph-structured facts with neural reasoning, it aims to achieve efficient, end-to-end trainable multi-hop reasoning over knowledge graphs.
Aethel provides a reproducible graph-retrieval framework for multi-hop financial diligence. It combines bipartite Personalized PageRank graph retrieval with a coreference-aware Bipartite Coreference Teleportation layer and an orchestrated specialist-agent architecture, modeling corpora as entity-passage graphs and propagating relevance through explicit relations.
AuthentiCity introduces a multi-source provenance-aware 3D city knowledge graph and benchmark spanning five cities. The dataset supports provenance tracking, spatial reasoning, and machine learning for urban digital twins.