Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 34 papers for 2026-08-12
We introduce Power Law Graph Attention (PLGA), replacing the fixed bilinear form of scaled dot-product attention with a learned, input-generated bilinear operator G_LM built from a positive tensor A_LM using power laws. PLGA exactly generalizes SDPA, collapsing to it when G_LM equals the identity, and the work provides theoretical statements and empirical observations about its behavior, including inference-level phenomena. This offers a unified attention mechanism that spans standard attention and its power-law generalization.
ProTAGAD is a foundation-model approach for anomaly detection in Text-Attributed Graphs (TAGs) that decouples topological structure and textual semantics into separate prototypes. By leveraging these decoupled prototypes, it detects anomalies that manifest in either structure or semantics, with applications to security, moderation, and cyber-threat detection. The method emphasizes robust cross-modal anomaly reasoning in TAGs.
GARLIC presents a Graph Attention-based Relational Learning framework for irregularly sampled multivariate time series in ICU data. It uses a learnable exponential-decay encoder to impute missing values, builds time-lagged summary graphs to capture inter-sensor dependencies, and fuses global information with graph attention for accurate predictions and interpretable insights. The approach targets improved ICU forecasting with better handling of missingness.
This work develops a Graph Neural Network–guided Genetic Algorithm for optimizing Physical Internet supply chains under cost uncertainty. It formulates deterministic and min-max regret models for a three-echelon factory-hub-retailer network and uses a GNN to guide the GA in solving the coupled discrete-continuous decisions. The result is more robust, cost-aware supply chain planning.
The paper investigates whether computational reducibility can yield transferable models for graph combinatorial optimization. It introduces a GNN encoder with a GCON module and energy-based unsupervised losses, achieving competitive results when trained per task, and then leverages reductions to transfer knowledge across tasks for improved generalization. The work highlights pathways toward more universal neural solvers for CO problems.
Lost in Aggregation identifies fundamental expressivity limits of Message-Passing GNNs. By proposing an information-complexity property for aggregations, it proves that MP-GNNs induce only a polynomial number of graph equivalence classes, while the space of non-isomorphic graphs is super-exponential; even two Color Refinement steps can yield exponential distinctions, underscoring inherent limitations.
BiScale-GTR introduces fragment-aware graph transformers for multi-scale molecular representation learning. It acknowledges that fragments have reusable identities but their behavior depends on the surrounding atomic environment, so representations must be both fragment-reusable and context-sensitive. The model combines multi-scale fragment information to improve molecular property prediction.
Foundations of Equivariant Deep Learning unifies graph and sheaf neural networks through order-equivariant neural networks (OENN). It extends geometric deep learning to generalize beyond simple group actions by leveraging equivariant vector bundles, providing a principled framework for symmetry-aware architectures that bridge graph and topological computations.
NL2SHACL-Bench presents a benchmark suite for translating natural language requirements into SHACL constraints. It addresses the challenge that semantically equivalent shapes can differ in serialization, proposing evaluation methods beyond string similarity and providing datasets and baselines for NL-to-SHACL research.
Towards Researcher Agents for Knowledge-Graph QA introduces an agentic text-to-SPARQL system that self-improves over rounds. After each validation step, it proposes and tests changes to prompts, rules, and tool orchestration, enabling a researcher-like adaptive approach to KG question answering.
GraphThink integrates a task graph to guide LLM thinking and a scene graph for environmental memory in long-horizon embodied planning. This dual-graph approach reduces planning hallucinations and enables robust event-driven replanning in complex environments.
KGCache proposes an in-memory cache for one-hop KG neighborhoods to amortize repeated retrieval in KG reasoning with LLMs. It is compatible with both iterative traversal (ToG) and one-shot planning (RoG) paradigms, reducing latency and improving efficiency.
VDGR-RAG presents a unified reasoning framework for hierarchical enterprise knowledge by combining vectors, directories, graphs, and reflective reasoning. It harmonizes diverse retrieval modalities to improve enterprise QA over structured, multi-level documents.
SodaMem introduces an evidence-grounded temporal graph memory for LLM agents. It extracts typed FactEvents with provenance spans, stores mention and occurrence times, and tracks validity windows to support currency, provenance, and ordered temporal reasoning.
Neurosymbolic Discovery of Algebraic Graph Constructions investigates automatically discovering short algebraic descriptions (e.g., Cayley graphs, lexicographic products) from raw graph data, bridging data-driven graph discovery with symbolic algebraic representations for interpretability.
ForestBench introduces a unified graph-based evaluation framework for multi-agent collaboration, mapping diverse MAS traces into collaboration graphs. This shared representation enables fair cross-method evaluation of collaborative behavior and performance.
PolicyKG provides an agentic LLM pipeline that translates institutional policies into SHACL graphs. It classifies policy sentences into obligations, permissions, or prohibitions, lifts them into deontic logic, and emits SHACL constraints via a state-machine with validators and a grounding vocabulary registry.
An Explainable GNN Framework for Component-Level Anomaly Diagnosis proposes a graph-based anomaly detection approach that attributes faults to specific components rather than only sensors, delivering explanations to improve reliability and safety in industrial processes.
Coupled Graph–Policy Distillation for Personalized Medication Safety in Older Adults with Multimorbidity (ATLAS) builds a medication safety framework where guideline evidence forms a medication-safety graph. It uses targeted questions to update patient state and distills relevant relations into a patient-specific conflict graph, guided by a risk-first multi-agent policy.
SkillConsist develops a bidirectional graph-alignment approach to detect inconsistencies between declared agent skills and actual behaviors. By aligning skill metadata, instructions, and resources in two directions, it identifies misalignment and potential unsafe or unintended behavior.
This paper introduces LGNNIC, a system architecture that uses SmartNICs to accelerate distributed GNN training. By offloading inter-node communication and placing graph data across remote memory nodes, LGNNIC reduces network congestion and improves scalability on multi-node clusters.
This work demonstrates benchmarks for LLM-driven kernel optimization under selection pressure. In Metal-Sci and Metal-ZK tasks, frontier LLMs generate and evaluate kernels inside an evolutionary loop, favoring winners whose improvements may reflect benchmarking artifacts rather than true generalization.
This paper presents STAIR, an end-to-end agentic planning framework for incident response. It addresses the limitations of static playbooks and unstable long-horizon planning by unifying incident state management, action alignment, and feedback from execution, enabling more adaptable automated responses.
Proposes an ontological perspective to define decentralization across computer systems. It argues lack of universal definition hinders applications in security, distributed computing, and decentralised AI, and proposes ontological criteria and relations to unify understanding.
Beyond Pixels argues for using DOM downsampling as a lightweight alternative to GUI screenshots for LLM-based web agents. The method preserves essential structural information while reducing token cost, enabling scalable and robust agent reasoning across web tasks.
Introduces probabilistic circuits to represent distributions over rule sets for KG completion, enabling much smaller rule vocabularies. The approach achieves a 70-96% reduction in the number of rules needed to reach peak baseline performance while maintaining accuracy.
Proposes El Agente Gráfico, a semantic execution runtime for scientific agents that uses typed execution graphs to enforce admissible scientific state transitions, record provenance, and confine model judgement to explicit decision points. Using the same top-level LLM and task-specific rubrics on s
Proposes Experience Memory Graph for one-shot error correction in LLM agents, addressing compounding errors without brittle prompt reflections. It provides a reusable memory mechanism that supports generalization across tasks with reduced cost.
Presents ProbSPARQL for querying knowledge graphs with multi-dimensional, uncertain numeric data, addressing native uncertainty representation gaps in RDF/SPARQL. The approach enables querying and reasoning over uncertain measurements to support reliability and triage workflows.
Introduces UNICON, a foundation model of numerical intelligence that generalizes across disciplines by acquiring and applying knowledge from numerical context. The work frames numerical reasoning as a cross-domain capability and presents a foundation model that learns to manipulate numerical information beyond textual contexts.
Explores how LLMs perform ranking tasks via pairwise comparisons and aggregation into a total order, addressing cognitive demands and potential biases. The paper discusses design choices for robust triaging rankings and stability under different prompts.
Develops a unified framework for Riemannian deep learning, including reusable modules, manifold-specific architectures, and geometries. It generalizes batch normalization to manifold-valued data and avoids Euclidean reliance, enabling neural networks on non-Euclidean spaces.
Introduces SCALE, a framework that aggregates scientific concepts using LLMs and embeddings to extend fine-grained taxonomies. It tackles fragmentation and variability of author keywords, proposing a method to create more stable, precise concept units for knowledge organization.
Proposes methods to make LLM-based recommendations more comprehensible by fine-tuning models to generate recommendation reasons. It also discusses aligning explanations with user preferences and evaluating explainability in recommender systems.