Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 35 papers for 2026-08-27
Nothing cleared the bar today. Papers read in full today: 13.
The load-bearing contribution is the decomposition of historical rank into old-candidate rank plus newly admitted candidates that outrank the answer, followed by a one-sided regularizer that only suppresses newcomer pressure exceeding score-blind matched old cohorts. The paper does test this mechanism better than a typical benchmark-only paper: paired comparisons, maximum/unmatched-reference controls, and ablations show that the reported ENTITY gain (H_cur +0.0057 over replay) is almost entirely a reduction in the proposed interference term rather than improved old-universe ranking. This is a careful and useful diagnostic intervention, but its ingredients closely resemble rank-aware negative control plus matching/calibration rather than a field-changing reframing.
Evidence is limited to constructed five-snapshot entity-growth streams, eight seeds, mostly frozen old embeddings, and comparisons centered on replay/LKGE/IncDE rather than a broad, current suite of strong continual-KGE and rank-calibration baselines; moreover, the claimed entity-interference prior work cannot be independently verified from the manuscript.
The load-bearing contribution is the solvability result: 2-edge connectivity suffices, while on arbitrary connected graphs the number of empty vertices must meet an effective-distance threshold determined by chains of bridge-separated components and singleton branching vertices. The supplied text gives a substantive constructive argument for that theorem and an NP-hardness reduction for station makespan, so this is more than a new MAPF acronym; however, the claimed SAT solver, PPP algorithm, and the reported 74–89% experimental result are not present in the supplied full text, so the practical optimization claim cannot be audited here.
The theory concerns a narrowly specified one-shot, undirected, anonymous-parking model, and without the missing algorithmic/experimental sections it remains unclear whether PPP beats modern MAPF baselines such as CBS/SAT implementations, LaCAM*, or high-density MAPF methods under fair tuning.
The load-bearing result is diagnostic rather than architectural: across nine datasets, short one-body rules plus learned temporal confidence functions match or exceed many much more elaborate TKG forecasters, and remain runnable on all tested TGB 2.0 graphs. The ablations substantiate that learned, rule-specific recency is the main gain over static confidence, with frequency adding a smaller increment; importantly, on WIKI recurrent rules alone equal the full system, directly reinforcing the paper's claim that these benchmarks often reward simple temporal repetition rather than multi-hop reasoning.
The strongest practical claim is limited by single-step evaluation and uneven comparisons—20 prior methods are excluded, several results are imported or fail under particular hardware/time budgets, and the key ablations cover only WIKI and ICEWS14.
Query-Side Attacks on GNN-Based KGQA: Tracing Failures from Entity Linking to Answer Generation. The paper analyzes robustness of GNN-based KGQA pipelines under adversarial question perturbations and introduces a stage-isolation protocol to identify where failures originate, enabling targeted defenses.
We reveal a storage-retrieval gap in KG memory: instead of retrieving subgraphs at query time, offline compile the KG into a bank of LoRA adapters, one per entity, forming a parametric knowledge layer. Queries are answered by injecting weights rather than text, eliminating query-time context costs. On MetaQA, these adapters encode factual knowledge that generalizes to unseen questions.
The paper studies stealthy backdoor attacks against Graph Foundation Models on text-attributed graphs under graph-language alignment. It argues that existing attacks targeting either the graph or text alone are insufficient in aligned settings. It proposes backdoor mechanisms exploiting the cross-modal alignment to stealthily control outputs.
ReactBench: A Benchmark for Topological Reasoning in MLLMs on Chemical Reaction Diagrams. It introduces ReactBench to test structural/topological reasoning in multimodal LLMs on chemical diagrams, highlighting that current models struggle with branching, convergence, and cycles beyond semantic understanding.
We systematically evaluate privacy risks in LLM-enhanced GNNs with a unified five-stage framework (dataset preparation, victim model, attacker, attack scenarios, and defense). Experiments reveal privacy vulnerabilities where sensitive graph or attribute information can be inferred from outputs. The work highlights risks and provides guidance for safer integration of LLMs with graph models.
DeltaGNN introduces information flow control to regulate message passing in GNNs. By controlling how information propagates, it mitigates over-smoothing and over-squashing that limit depth. The approach enables deeper GNNs to better capture long-range node interactions.
We investigate why coupling a text teacher does not fully benefit graph learning. The paper proposes a multimodal setup combining a self-supervised transfer that allows a GNN encoder pretrained on one dataset to operate on another with different feature dimensionality, and an alternating E-step/M-step optimization that updates a language-model module and the GNN module separately. This avoids costly joint retraining and improves cross-dataset transfer.
The work tackles multi-source complex network reconstruction with Wasserstein distributionally robust optimization and algorithm unrolling. It fuses heterogeneous source-domain data while preserving distinct geometries and achieving robustness to distribution shifts. The approach improves reconstruction performance in data-scarce target domains.
This paper studies KGQA with LLMs under partial knowledge, focusing on constraining entity selection among candidate answers to improve grounding in the knowledge graph. It proposes a framework that uses partial information to select and refine candidate entities, aiming to enhance robustness of LLM-based KGQA.
The paper investigates creator personas in procedural tasks by building ViralRecipesTrans, a dataset of procedurally aligned execution graphs tied to creators. It analyzes stylistic patterns and develops methods to characterize and generate procedural personas.
The paper extends rotary position encodings to graphs using Wave-Induced Rotary Encodings (WIRE). It rotates attention tokens according to the graph Laplacian spectrum to inject structural information into the attention mechanism. This improves graph learning on synthetic and real tasks.
MetaSieve targets faster relational deep learning by SQL-based metapath selection. It reduces subgraph size by exploiting the join and aggregation capabilities of relational databases to focus on informative metapaths. This leads to significant training speedups while preserving performance.
Ollivier-Ricci Curvature of Riemannian Manifolds and Directed Graphs with Applications to Graph Neural Networks. This work surveys Ollivier-Ricci curvature and its connections to classical Ricci curvature, extending results to graphs and discussing implications for GNN design and generalization.
SNAP-KG proposes a streaming node assignment method, for newly arriving entities that have no graph connections yet, projecting their features into semantic communities to enable downstream entity resolution and link prediction. Unlike transductive multi-view clustering that assumes fixed graphs, SNAP-KG is designed for streaming KG integration with scalable assignment.
We propose self-evolving LLM agents for hardware kernel optimization, leveraging an experience-driven workflow and an Experience Graph Memory to store decisions, feedback, and later use evidence. This enables agents to learn from past optimization runs rather than starting from scratch. The approach aims to automate and improve kernel optimization across iterations.
Clinical Graph-JEPA presents a knowledge-graph construction and refinement framework for predictive patient-state reasoning. It combines multi-agent relation proposal, ontology-aware normalization, deterministic evidence scoring, and JEPA-based latent refinement to produce robust clinical KG. The system aims to improve cognitive decision support from noisy clinical data.
This paper proposes Edge-Local and Qubit-Efficient Quantum Graph Learning for the NISQ Era. It proposes a hybrid quantum-classical architecture with a variational feature extraction layer and edge-local quantum processing to reduce circuit depth and qubit requirements, enabling unsupervised graph learning on near-term quantum devices.
The work incorporates cognitive load and cross-domain knowledge transfer into multi-domain knowledge tracing. It models how managing learning across domains affects cognitive effort and how knowledge states transfer between domains. The goal is more accurate dynamic student-state tracking across multiple domains.
Adaptive Influence Graphs for Failure Attribution in Multi-Agent Systems uses adaptive graph-based traces to attribute failures in complex multi-agent runs, helping LLMs reason about causes. It improves debugging efficiency by organizing traces around components, actions, and dependencies.
HMGCLIP proposes heterogeneous multi-granularity contrastive learning for e-commerce representations. It constructs a heterogeneous hypergraph to capture fine-grained attributes across modalities and aligns them with contrastive objectives. The method improves fine-grained attribute understanding beyond global embeddings.
FedV-KGQA introduces a federated framework for multi-hop QA over vertically partitioned knowledge graphs. Organizations share entities but keep disjoint relation sets, and the method combines local graph enrichment with KG embeddings to enable cross-silo reasoning without centralizing data.
Panning for Gold: Expanding Domain-Specific Knowledge Graphs with General Knowledge. The work argues that domain-specific KGs have limited coverage and proposes a systematic approach to enrich them by leveraging information from general knowledge graphs.
Denoising the Future: Context-Aware Spectral Diffusion for Temporal Knowledge Graph Extrapolation. It introduces FreqDiff, a frequency-aware diffusion framework for extrapolating future facts in temporal knowledge graphs, conditioning diffusion on subject histories and query-specific evidence.
Cross-Domain, Multi-Task Data-to-Text Generation without In-Domain Training Data. The study explores data-to-text generation across domains without in-domain training data or test references, aiming to generalize across domains, generation goals, and input types such as tables, KGs, charts, and time series.
MOTIF: Motivation-guided Topology Inference for Cold-start Multimodal Recommendation. MOTIF combines Semantic Motivation Reasoning, Knowledge-enhanced Graph Reconstruction, Weighted Graph Contrastive Learning, and Semantic-Structural Alignment, aided by offline LLM reasoning, to tackle cold-start multimodal recommendation.
Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence. The paper discusses a shift from sole reliance on individual models to system-level intelligence, outlining prompts, contexts, harnesses, and loops as engineering paradigms for coordinating heterogeneous components.
STDSH-MARL is a spatio-temporal dual-stage hypergraph MARL framework for human-centric corridor traffic signal control. It uses a centralized training and decentralized execution paradigm to model multimodal traveler interactions. The dual-stage hypergraph captures complex spatio-temporal dependencies to optimize signal decisions.
Don't Just Listen, Try Planning: Graph-based Retrieval-Generation Agent for Long-form Audio Meeting Understanding. It introduces the LongAudioQA dataset and the GRGA agent, which models heterogeneous audio features as a multi-dimensional graph and uses planning to guide retrieval and answer generation.
Balancing Safety and Optimality in Robot Path Planning: Algorithm and Metric. The Unified Path Planner (UPP) is a graph-search algorithm that dynamically balances safety and optimality via adaptive heuristic weighting, using a local inverse-distance safety field and on-the-fly parameter tuning to maintain bounded heuristics while improving clearance.
Optimizing Expert-Designed Crystal Graph Networks for Band-Gap Prediction with an Autonomous LLM Research Loop. The work shows an autonomous LLM-driven loop that optimizes expert-designed crystal graph networks for band-gap prediction on MatBench, achieving strong results without external pretraining, driven by self-guided data and model exploration.
Molecular LLM Agents: From Architectural Design to Scientific Autonomy. This work surveys how molecular LLM agents must perceive, reason about, and act on chemical objects across strings, graphs, 3D conformations, spectra, and simulations, outlining a framework for perception grounding, tool grounding, and feedback for scientific autonomy.
COCI: Conference Organisers and Content Identifier. It demonstrates an AI-based framework to extract fine-grained structured metadata from raw Calls for Papers texts, using a multi-stage pipeline that leverages large language models for parsing heterogeneous CfP documents.