Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 71 papers for 2026-08-25
Nothing cleared the bar today. Papers read in full today: 14.
The load-bearing idea is to turn size generalization into a covering-and-Lipschitz-certificate problem: if the target is uniformly approximable with a bounded certificate and training examples form an appropriate graph-pseudometric cover, low regularized empirical loss implies uniform accuracy on arbitrarily large graphs. The paper does substantiate this mathematically, including finite-Lipschitz constructions for normalized-sum, mean, and min/max MPNNs, a sharp diagnostic that unnormalized-sum/degree-sensitive classes fail the finite-cover condition on unbounded-degree graphs, and a constructive K+1-example Bellman–Ford result; experiments are supportive but narrow and the proposed regularizer only modestly beats ordinary ℓ1.
The headline guarantee is strongly conditional on an often nonconstructive ε-cover, bounded-certificate realizability, and reaching near-global regularized optima; beyond fixed-step Bellman–Ford, it does not establish that ordinary sampled training and gradient descent will reliably learn algorithms of unbounded effective depth.
The load-bearing finding is concrete and practically consequential: native SID collisions reach 30.52%, conventional Hit@10 exceeds its proposed item-level score by as much as 103.36%, and RK-Means changes from SID-level leader to corrected-score last place on three datasets. The paper tests that diagnosis across four datasets and four collision-prone tokenizers, and its ZCR ablation establishes lower quantization reassignment cost than MQL4GRec-style greedy reassignment; it is a strong diagnostic paper, not another tokenizer architecture.
CCE is only a uniform-random tie-breaking interpretation of an ambiguous generated SID—not uniquely the faithful item-level metric—and ZCR retrains on altered IDs, so its downstream gains conflate collision removal with changing the generator's prediction code and are demonstrated only when every fixed-prefix group fits the final codebook.
The load-bearing contribution is a reasonably clean factorial diagnosis: hold the teacher narrative and backbone fixed, vary SID versus title representation and reasoning, then separately test SIDReasoner's costly eight-objective alignment. It finds that titles and full SID alignment markedly improve judged groundedness (vanilla SID 1.58 versus Title 4.07; aligned SID 3.97) but reasoning SFT lowers or does not change Recall@10 in every controlled setting; dense judge/embedding rewards partly restore accuracy, implicating reward sparsity/objective mismatch rather than descriptive-trace quality. This is a useful negative result for anyone tempted to treat fluent recommendation rationales or SID grounding as evidence of functional reasoning, but it is a careful diagnosis rather than a new recommendation capability.
The causal scope is limited by one small 1.7B backbone, three roughly 3.5K-item Amazon catalogs, answer-conditioned GPT teacher traces, asymmetric title-to-BM25 versus constrained-SID decoding, and offline single-held-out-item evaluation; therefore it cannot establish that high-quality traces generally fail in production or under end-to-end, target-unconditioned reasoning supervision.
The load-bearing idea is not merely the RWR architecture but the multigrid-style diagnosis: bottlenecked latent attention learns low spatial frequencies while finite-depth message passing lacks domain-scale communication, so their combination improves gradients and wall-shear quantities more than bulk fields. The full text supports that diagnosis reasonably well through matched MGN/RWP configurations, training-time spectra, latent-count sweeps, scarce-data curves, and million-cell tests; however, the interleaving-specific gain is narrow once both operators are included, and the claimed mechanisms are explicitly only consistent explanations rather than causally isolated results.
Its strongest explanatory claim rests on one steady RANS case and single-seed public-benchmark results, while it neither disentangles bottleneck, attention, loss, and spectral-bias causes nor broadly compares against existing hybrid PDE surrogates with reference implementations.
Search Broadly, Seek Evidence on Both Sides, Decide Narrowly: Evidence-Admissible GraphRAG for Longitudinal Clinical Event Verification.
Compositional Chain-of-Relations for Faithful Knowledge Graph Question Answering with Large Language Models.
EchoTrace proposes a role-aware, phase-wise diagnostic framework to diagnose recursive risks in LLM-powered recommender systems, where LLM-generated signals influence future training data and recommendations. It identifies bias, hallucination, and data-quality risks across pipeline phases and roles.
BIRDNet mines Boolean implication relationships from tabular data with a sparse-exception binomial test and encodes the resulting typed graph as the connectivity of a layered neural network. Each hidden unit corresponds to a mined rule and connects only to the two features involved, yielding a sparse, interpretable model.
GeoRisk-RAG is a hierarchy-aware framework to improve reliability of retrieval-augmented generation in geospatial hazard contexts by distinguishing geographic validity from semantic similarity and enabling selective answering. It leverages geographic granularity (town/city/state) to decide when to answer, abstain, or escalate, and to calibrate risk-aware retrieval.
The boundary-value Friedkin-Johnsen dynamics study treats fully stubborn agents as boundary nodes and interior agents as part of a discrete boundary-value problem, yielding a Green function formulation to compute source-specific influence-based centralities on directed weighted networks.
We propose a discrete Bayesian network model with hierarchical Dirichlet shrinkage, introducing a hierarchy over node-parent conditional probabilities to shrink high-dimensional parameters and enable posterior sampling.
This work proposes hierarchy-aware semantic losses for knowledge graph link prediction using box embeddings to encode class hierarchies and enforce subclass relations during representation learning, extending hierarchy-aware GNNs beyond biological regression tasks.
TRicci introduces a curvature-inspired edge sparsification for dynamic graphs based on temporal Forman-Ricci curvature. By pruning less informative edges, it improves scalability for dynamic graph learning while preserving essential structure.
From Association to Causation: Improving Retrieval Precision of Retrieval-Augmented Generation via Causal Relations and an Attention Mechanism.
The paper presents a unified framework for systematically comparing lightweight strategies to align biomedical text and knowledge graphs. With a frozen text encoder and a frozen KG embedding model, it learns only a lightweight projection between the two spaces to enable grounding, evidence retrieval, and KG completion.
GRAFT builds a graph of scientific papers whose edges are typed by four facets: problem, method, result, and contribution. By moving beyond single similarity scores, it enables explaining why papers are related and supports exploratory retrieval through a facet-aware, graph-based representation derived from facet items and citations.
We argue for decoupling temporal granularity in long-video RAG; propose Density-Aware Graph Construction (DAGC) to allow coarse indexing for locating relevant regions and fine-grained evidence for reasoning. This decoupling improves scalability and reasoning precision for long-form video content.
Hypergraph Embedding Index (HEI) organizes documents by combinations of highly activated latent embedding dimensions, enabling inverted-index style candidate generation while preserving the ranking capabilities of dense embeddings. This approach yields efficient dense-vector retrieval with competitive accuracy.
MileGPO proposes Milestone Inference with Local Evidence to improve graph-based policy optimization for long-horizon LLM agents. It derives process-level credit from grouped on-policy rollouts via Milestone Discovery and Local Evidence, enabling better credit assignment.
When Persona Simulations Are Informative: Graph-Structured Signals for Pluralistic Opinion Sensing.
GTA-RAG proposes Graph-Trajectory-Augmented RL for multi-turn retrieval-augmented reasoning, using graph trajectories to provide denser supervision beyond sparse final-answer rewards, improving evidence retrieval across documents.
SLogic: Subgraph-Informed Logical Rule Learning for Knowledge Graph Completion learns rule confidence conditioned on local subgraph context, enabling instance-specific rule weighting and improved KGC accuracy and interpretability.
Beyond Observed Auxiliary Relations introduces environment-conditioned modeling for multi-behavior recommendation; it handles missing and noisy auxiliary signals by conditioning on the environment/context and learning robust behavior interactions.
DuoGNN proposes a topology-aware GNN with homophily and heterophily interaction-decoupling; it decouples processing for different relation types to mitigate over-smoothing and over-squashing, improving robustness across datasets.
We propose Confidence-Aware Dual-Teacher Learning for few-shot node classification on Text-Attributed Graphs. The method uses two teachers (e.g., LLMs and GNNs) with node-wise confidence to decide who should teach which nodes, reducing cost by avoiding unreliable LLM supervision for certain nodes.
DeltaGNN introduces information-flow control in GNNs to balance short- and long-range information; it mitigates over-smoothing and over-squashing, enabling deeper, more expressive networks.
From Mastery Profile to Simulated Response: Stochastic Student Knowledge Graphs (SSKG) for Faithful LLM Student Simulation.
We propose a reproducible, graph-centered evaluation framework to ground healthcare LLMs in a causal knowledge graph, stress-testing intervention-oriented reasoning, mechanisms, harms, evidence, and uncertainty in a cardiovascular pilot. The four components help assess reasoning beyond single-answer accuracy.
We present a graph-based framework to analyze how humans and LLMs reason in scientific fact-checking, beyond merely verdicts. The framework traces reasoning paths and compares them with human expert trajectories, highlighting when LLMs follow different yet valid reasoning routes. This aids in diagnosing faithfulness and transparency.
We present a three-stage pipeline for conversational music recommendation that combines multi-modal retrieval with constrained LLM reranking. It builds decay-weighted centroids across seven embedding spaces (track- and user-level CF-BPR, Qwen3 metadata/lyrics/attributes, CLAP audio, SigLIP visual) and augments them with BM25 lexical signals and artist substring matches to form candidates, before a constrained LLM re-ranker selects final results. The approach yields personalized, multimodal recommendations in conversational settings.
ReCoG introduces Reciprocal Co-Evolution for multimodal graph learning, where graph topology and multimodal node attributes are learned jointly in a feedback loop. This captures cross-modal interactions more effectively than decoupled approaches.
Two-level domain-decomposition AdaGrad (DD-AG2m) for scalable GNN training decomposes the graph into subdomains and uses second-order AdaGrad updates to reduce computation, memory, and communication in distributed settings.
HI-MeshGraphNets develops hierarchical multi-scale graph neural networks to efficiently learn physics on large, high-fidelity meshes. This architecture enables long-range interactions with fewer layers, reducing computational cost and memory usage while maintaining accuracy.
SAFE-G is a structure-aware evidence-guided generation framework for knowledge-based visual question answering (KB-VQA). It emphasizes faithful use of external knowledge by aligning evidence with explicit structures and constraining the reasoning process of the model. The approach improves the reliability and coherence of answers in knowledge-intensive VQA tasks.
Adaptive Item-based Collaborative Structures via Noise Rescheduling in Diffusion for Generative Recommendation introduces noise scheduling to inject item-based CF priors into diffusion training. This enables the model to better incorporate item relationships alongside user histories, boosting recommendation quality.
CoAL-RAG is a complexity-aware legal retrieval-augmented generation method that uses a multi-dimensional evaluation based on question essence and retrieval consistency to enable adaptive retrieval strategies, balancing answer quality and efficiency in high-risk legal scenarios.
Clinical Graph-JEPA integrates multi-agent relation proposal, ontology-aware normalization, deterministic evidence scoring, and JEPA-based latent refinement to construct and refine patient-state knowledge graphs for cognitive decision support. It addresses extraction errors, ontology mismatch, missing relations, and temporal ambiguity.
Unified Path Planner (UPP) balances safety and optimality in robot path planning via adaptive heuristic weighting, using a local inverse-distance safety field and real-time parameter tuning to achieve provable suboptimality bounds while maintaining superior clearance.
VisAdj learns adjacency matrices from node-link images; it introduces an attention-sparse neighbor sampler to propose a high-recall candidate edge set and performs joint edge inference, enabling topology recovery from visuals.
NL2SHACL-Bench introduces a benchmark suite for translating natural language requirements into SHACL shapes, addressing evaluation beyond string similarity since semantically equivalent shapes can differ in serialization and structure. The benchmark emphasizes structural and semantic fidelity in NL2SHACL generation.
We introduce self-supervised graph representation learning for in-the-wild wearable and smartphone emotion recognition using SSL graph masking and subgraph sampling, leveraging both labeled and unlabeled data in a limited-resource setting. The approach improves robustness and accuracy for emotion classification
MasterSet provides a large-scale benchmark for must-cite citation recommendation in AI/ML literature, targeting the identification of foundational papers, direct experimental baselines, and core dependencies whose omission would misrepresent a contribution’s novelty. It addresses the rapid growth of AI/ML venues and enables automated evaluation.
SchemaRouter introduces a field-aware tool routing layer for heterogeneous agentic RAG systems. By representing tools with their input/output fields and parameters, it reduces over-fetching and under-fetching and improves tool usage efficiency.
GenRec is Netflix’s LLM-backed recommendation ranker built on an in-house foundation model. It follows a two-phase framework: Phase 1 adapts an open-source LLM to Netflix data to gain catalog and user-behavior understanding, Phase 2 adds post-processing ranking to produce final recommendations.
Beyond Similarity: Heterogeneous Graph Learning for Multi-Objective Food Substitution in Charitable Food Agencies.
Toward Effective and Reliable LLM Agents via Dynamic Ontology.
GraphMed-LT introduces a patient-specific graph memory with latent clinical thought refinement to support multi-turn medical conversations, enabling coherent reasoning and evidence aggregation across turns by structuring patient data in a graph.
FashionKG-RAG proposes a knowledge graph–enhanced retrieval-augmented generation approach for fashion question answering. It addresses limitations of existing fashion KGs that focus on product attributes or item-level relations by integrating broader knowledge to reduce hallucinations and improve decision support in fashion QA.
Entity Representation Learning Through Onsite-Offsite Graph for Pinterest Ads builds graphs using both onsite user activity and offsite conversions to learn richer embeddings for ads and improve conversion prediction.
ONOTE is a hypergraph-grounded omnimodal reasoning framework for computational music science, unifying auditory, visual, symbolic, and physical representations of musical events. It addresses structural consistency across notation systems and resolves pitch, timing, and instrument-specific constraints.
PolyUQuest proposes a verifiable, structure-aware web RAG built on heterogeneous graphs, unifying hyperlink topology, DOM hierarchy, and cross-page entity-relationship knowledge. A two-tier router dispatches queries to direct block retrieval, cross-page graph traversal, or multi-hop retrieval to match structural needs and enable verifiable outputs.
Hierarchical Book Organization for Learning-Resource Discovery uses Dual-Path Graph Convolutions to capture both hierarchical book categories and the divergence between authoritative descriptions and crowd reviews. This approach improves semantic organization and resource discovery.
MAVEN-T presents reinforced heterogeneous distillation for real-time multi-agent trajectory prediction. Lightweight predictors are trained through distillation from heavier models to enable real-time deployment under dense interactions and multimodal futures.
The paper presents a machine learning-based method to automatically tune the penalty parameter in minimum bisection formulated for quantum annealers. It aims to improve partition balance and solution quality without manual parameter tuning.
Agentic-Kube is a graph-enhanced multi-agent RL framework for real-time Kubernetes scheduling. It enables cooperative agents to optimize multi-objective pod placement, balancing cost, fault resilience, and utilization while mitigating gradient interference.
RSTGCN: Railway-centric Spatio-Temporal Graph Convolutional Network targets station-level train delay prediction; it builds a railway-aware spatio-temporal graph and uses a spatio-temporal GCN to forecast average delays at stations.
Molecular LLM Agents survey outlines architectural design and autonomy for molecular science agents that operate across symbolic strings, molecular graphs, 3D conformations, spectra, simulations, and wet-lab measurements. It highlights the need for chemically faithful perception, an LLM-centered agent framework, domain-specific tool grounding, and feedback loops.
Mycelial Search (Myco) is a graph-structured metaheuristic for continuous optimization, featuring active tips, community-weighted flow, adaptive cord plasticity, and anchor-based injection. Candidate solutions form a dynamic graph partitioned by Louvain to distinguish within-community from cross-community information exchange, guiding search.
DAGSmith introduces dependency-aware rewriting for dbt-style SQL pipelines. Since modern analytics rely on large DAG-structured pipelines of interdependent SQL models, traditional single-query optimizers fall short; DAGSmith rewrites and optimizes across the dependency graph to improve execution.
TSWAP presents a deployed eight-language conversational wellness advisor grounded, via retrieval-augmented generation, in a verified knowledge base of Thai traditional medicine and certified wellness providers. It uses an open-weight LLM (Qwen3.6-35B-A3B on vLLM) with a ~30.6K-chunk Thai index by a hybrid dense-sparse retriever; a first-turn query classifier forces tool-based retrieval for entity lookups; a safety layer enforces medical scope and Thai emergency routing across eight languages.
GraphSVR combines graph convolutional learning to capture inter-station spatial dependence with support vector regression for robust spatiotemporal forecasting of urban air pollution. It addresses nonlinearity, nonstationarity, and anomalies caused by traffic, weather, and measurement errors, delivering robust forecasts.
The Abstention Protocol CoreSec applies a PAM-style abstention algebra for root-cause analysis in large datacenter fabrics. Telemetry agents are composed with control flags to yield deterministic decisions and explicit abstention when evidence is ambiguous, improving RCA robustness.
GrOIL introduces a seven-stage, graph-grounded pipeline that converts domain documents into a complete, auditable OWL TBox without any unconstrained generation step. Documents are encoded as Unified Discourse-Hypergraphs (UDH) that capture entity participation and discourse structure, enabling corpus grounding, vocabulary consistency, axiom-level expressivity, and end-to-end provenance.
SAGE is a Stability-Aware Graph-Based Ensemble feature selection system for postpartum depression risk prediction that combines local explainable AI and a genetic optimization component to produce stable, interpretable features despite class imbalance.
First-principles atomistic structure and dynamics of polyethylene during high-pressure radical polymerization are studied using machine-learning force fields, achieving near-quantum accuracy in large-scale simulations and revealing microstructural evolution.
This survey reviews neural-symbolic reasoning for knowledge graphs from a query perspective, summarizing how neural methods can complement symbolic reasoning to handle incomplete and noisy KG data. It discusses representative frameworks for query-driven inference, reasoning, and learning over KGs, with a focus on future directions.
HiFiNet presents a hierarchical fault identification framework for wireless sensor networks using edge-based classification and graph aggregation, with a two-stage process to improve fault detection accuracy while reducing energy consumption by exploiting spatio-temporal correlations.
We apply Graph Representation Learning to analyze differential cryptanalysis of Lightweight IoT ciphers SIMON/SIMECK. The approach identifies high-probability differential clusters efficiently, aiding resilience evaluation and visualization.
Automated Construction of FAIR Digital Object Knowledge Graphs from Flat Cultural Heritage Records.
This work presents a Neighbor-Embedded Graph Neural Network-based Crowd Delivery Traffic Management in Smart City, leveraging neighbor embeddings to coordinate crowd delivery and reduce congestion while optimizing vehicle assignment.
Composable Trust Infrastructure for Manufacturing Knowledge Graphs argues that SHACL validation, PROV-O provenance, bi-temporal versioning, and graph-native decision objects can be composed via shared correlation identifiers to achieve composable trust.