Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 30 papers for 2026-08-26
Nothing cleared the bar today. Papers read in full today: 11.
The load-bearing idea is real rather than a new GNN-style wrapper: shared continuous latent positions provide identifiable heterogeneity even when corresponding edges carry zero correlation, and the proof turns uniform OOS embedding error plus latent-point separation into perfect matching. The text does test the no-edge-correlation premise in RDPG simulations and shows that matching can restore power for shuffled graph tests, but its strongest practical variant (λ=0.5) is not the variant analyzed, and it is often nearly indistinguishable from the much simpler Procrustes baseline (ProcMatch).
The exact-recovery theorem is narrow—fixed-dimensional continuous, well-separated RDPGs with identical latent positions and a restrictive seed/unseeded scaling—and the experimental baseline set omits modern uncorrelated latent/graphon-matching competitors.
The load-bearing contribution is the effective-distance criterion on the bridge/2-edge-connected-component condensation tree: after gathering free vertices into a pivot-adapted configuration, a tasked agent can reach a pivot iff the number of holes meets this effective distance. Together with the 2-edge-connected guarantee and single-pivot NP-hardness reduction, this is a meaningful structural account of when intermediate-resource access is feasible. The supplied text gives substantial constructive proofs, but it stops during the flowtime-hardness proof and contains no experimental section, so it does not let us verify the advertised 74–89% PPP success rate, orders-of-magnitude claim, baseline tuning, or whether PPP's pivot-prioritization is materially better than standard prioritized planning plus assignment.
The strongest result is tailored to a narrow one-shot, single-type interchangeable-pivot model, while the empirical algorithmic claims are uncheckable from the supplied incomplete text.
We systematically evaluate 12 instruction-tuned open-weight LLMs on direct causal-edge judgments across six benchmark graphs, testing six prompts and four confidence signals (verbalized, logit-based, cross-prompt, cross-model). We find that LLMs offer calibrated but not fully reliable causal judgments; consistency across prompts and models helps, but trust remains conditional on the confidence source.
GNNBleed analyzes inference attacks that reveal private edges in graphs under realistic access to GNN models. It demonstrates how model outputs can leak edge information and discusses potential mitigations to protect edge privacy.
MEOR formalizes candidate-set interference in continual KG embedding and introduces a host-level regularization that matches newcomer pressure with answer-relative excess outranking, preserving historical answer rankings when new entities join.
The article introduces graph-dependent shrinkage priors for Bayesian trend filtering to better capture dependencies encoded by graphs. It argues that classical trend filtering is brittle with missing data and lacks uncertainty quantification, and proposes graph-aware priors to improve smoothing, robustness, and uncertainty assessment.
This study tackles cross-domain data-to-text generation without in-domain training data. It analyzes how to perform multi-task D2T generation across diverse data forms (tables, knowledge graphs, charts, time series) without domain-specific training text or references, leveraging zero-shot and generalizable methods.
We tackle complex network reconstruction from heterogeneous sources using Wasserstein distributionally robust optimization and algorithm unrolling. By matching distributions across sources, the method robustly fuses data to reconstruct networks without being biased by source divergence.
G2I proposes a two-stage greedy framework that leverages explainable explanations on graphs to generate intervention hypotheses from predictions. It moves beyond node-level explanations by producing network-level interventions and demonstrates improvements in actionable hypotheses for public health and social science decisions.
We show that, in a localized atomic-orbital basis, the molecular single-particle Hamiltonian (after a shift to make it positive semidefinite) is the Laplacian of a cellular sheaf on a regular cell complex built from the molecular graph. This reveals a structural connection between E(3)-equivariant Hamiltonian learning and sheaf cohomology, enabling new equivariant learning architectures for electronic structure.
We study KGQA where LLM-grounded answers must be grounded to a KG under partial knowledge; propose a constrained entity-selection approach that enforces grounding constraints to improve answer accuracy even when the KG is incomplete.
Quantum maximum entropy inference extends GIS and gradient-descent-based learning to quantum graphical models, addressing non-commutativity. The paper introduces quantum iterative scaling (QIS) and analyzes convergence, providing algorithms to perform maximum-entropy inference and Hamiltonian learning in quantum settings.
FlowNeg introduces a FlowNet-based, context-conditioned hierarchical generative model to sample hard negatives for knowledge graph embeddings. Given a positive triple and a corruption context, it selects the corruption type and entity to produce diverse, informative negatives without needing to normalize over all entities, improving learning efficiency and quality.
TaLK distills text-attributed graph datasets by coupling a language model with a graph-aware kernel, enabling efficient dataset distillation for TAGs. The method reduces training costs while preserving key semantic and structural information for downstream tasks.
Using autonomous LLM research loop, the paper has an expert-designed crystal graph network for band-gap prediction, with an autonomous coding agent that designs and trains the model on MatBench; achieves strong results without external pretraining.
We construct ViralRecipesTrans, a dataset of procedurally aligned execution-flow graphs extracted from cooking videos and mapped to creators; we formalize procedural style metrics and investigate how to discover and generate creator-style procedural graphs.
MolGA studies adapting pre-trained 2D graph encoders to molecules by incorporating rich submolecular knowledge (atoms, bonds). It proposes a flexible adaptation framework that combines pretrained encoders with domain knowledge to improve molecular property prediction and generalization across molecular domains.
Domain-specific knowledge graphs (DKGs) often suffer from limited coverage and quality compared with General Knowledge Graphs (GKGs). This work investigates how high-quality GKGs can be systematically leveraged to supplement DKGs, addressing coverage gaps and enrichment opportunities.
This thesis/exposition surveys Ollivier-Ricci curvature for metric spaces and connects it to classical Ricci curvature via 1-Wasserstein distance and optimal transport. It covers key results and proofs, including extensions to graphs (Lin–Lu–Yau) and implications for graph neural networks.
MGQL provides executable, small-step semantics for Graph Query Language (GQL) to support rigorous reasoning about the ISO standard. It preserves features like bag semantics, schemas, and multi-graph queries, enabling faithful implementation and verification.
We extend GraphGANFed to enable conditioning on user-defined objectives in a federated setting, allowing generation of graph-structured molecules that optimize specified metrics while preserving data privacy. The approach blends conditional graph generation with federated learning to tailor molecular design without sharing raw data.
LION introduces a Clifford neural paradigm for multimodal-attributed graphs, leveraging geometric algebra to fuse topology and multiple modalities in a principled way. The approach improves representation and downstream task performance by better aligning contextual information across modalities within graphs.
We propose Adaptive Influence Graphs to attribute failures in multi-agent LLM systems by organizing observability traces around components, actions, and dependencies; the approach enables targeted debugging and improves failure localization using LLMs.
FedV-KGQA enables multi-hop question answering over vertically partitioned KGs in a federated setting by sharing entities while keeping relation subsets local. It combines local graph enrichment and KG embeddings to support cross-organizational multi-hop reasoning without centralized data sharing.
We propose FreqDiff, a Frequency-aware Diffusion framework for temporal knowledge graph extrapolation. It addresses the issue that conditioning on subject histories can obscure query-specific evidence, by incorporating frequency-aware conditioning to better isolate signals relevant to forecasting future facts.
Open-Ended Deep Research (OEDR) agents often struggle with long-horizon tasks. This work advocates separating knowledge exploration from outline structure using two graphs, providing explicit supervision for discovering missing relations and triggers to improve long-horizon research workflows.
GATNextHop studies whether a Graph Attention Network can approximate shortest-path routing and generalize across network topologies. Trained on synthetic graphs and tested on real ISP-topology graphs, it benchmarks the GNN’s ability to predict next hops and generalize SPF decisions beyond the training topology.
The paper targets long-form audio meeting understanding (LAMU) by identifying limitations in current speech QA and memory. It introduces LongAudioQA and GRGA, a graph-based model that encodes heterogeneous audio features into a multi-dimensional graph, and uses agent planning to retrieve relevant information and generate answers to improve handling of long-range context.
The paper presents the Unified Path Planner (UPP), a graph-search algorithm that dynamically balances safety and optimality via adaptive heuristic weighting. It uses a local inverse-distance safety field and auto-tunes parameters during search to maintain bounded heuristics while achieving superior obstacle clearance, enabling rigorous evaluation of safety-optimality trade-offs.
Molecular LLM Agents examine how to build LLM-driven agents capable of perceiving and acting upon molecular objects across strings, graphs, 3D conformations, spectra, simulations, and wet-lab data. The work outlines an architectural framework for molecular agents, emphasizing perception, domain-specific tool grounding, and feedback loops to enable scientific autonomy.