Showing 34 papers for 2026-08-11
SkillConsist investigates inconsistencies between agent skills by bidirectionally aligning skill graphs built from metadata, instructions, and resources. It addresses risks from undisclosed or misapplied skills that lead to unsafe behavior or incorrect skill selection, and advances detection beyond prior text-only or declared-scope checks, leveraging a bidirectional graph alignment approach.
CONFER presents a graph-based conflict-aware evidence negotiation framework for weakly supervised multimodal emotion recognition. Each modality expert is a node with predictive belief, boundary-based uncertainty, and reliability estimated from historical performance and current sample uncertainty; the model negotiates compatibility and reliability to handle self-report unreliability and cross-modal conflicts.
This work treats circuit localization as a graph-learning problem, where a transformers computation graph's edges are the pathways; they propose Graph Circuit Learning (GCL) as a supervised method to identify sparse subgraphs that reproduce a targeted behavior across models.
We address input heterogeneity by SIGIL, mapping any attributed graph to a fixed-dimension representation. A structural interaction graph is constructed where nodes are input feature dimensions and edges are weighted/typed interactions; this enables fully-inductive generalization to unseen graphs.
We tackle Trajectory-User Linking by learning embeddings from a multi-relational knowledge graph that connects POIs, times, and semantics across trajectories. The approach leverages shared structural knowledge to better relate anonymous trajectories to the correct users, addressing limitations of feature-only methods.
HOPPER introduces learnable hop extraction for Linearized Graph Sequence Models to decouple information depth from processing depth, alleviating over-smoothing and enabling long-range information to be retained.
LEED introduces a local metric, Local Embedding Evolution Distance, to quantify over-smoothing at the node level, enabling more fine-grained analysis than global metrics. The metric guides architecture design and even virtual node selection to address over-smoothing in GNNs.
This work argues that item semantic IDs (SIDs) added to LLM vocabularies should be initialized with semantic priors instead of random kernels to preserve item semantics. It proposes strategies to initialize SIDs to improve generalization and recommendation quality.
We explore automatic discovery of compact algebraic graph descriptions from raw adjacency data, aiming to express graphs as Cayley graphs or lexicographic products. The approach seeks to reveal structural properties beyond raw adjacency, bridging graph construction with algebraic descriptions.
PreGress presents a ranking-native pre-training and prompting framework for graph node ranking tasks, enabling scalable, task-agnostic ranking without retraining for each downstream task. It learns general ranking notions during pre-training and retrieves tailored prompts at inference.
We propose fast GNN-based estimators for sparse matrix condition numbers, with an O(nnz + n) feature construction, and two schemes for decomposing and predicting the condition number. The approach aims for efficient deployment and accurate estimation.
We show that LayerNorm inherently removes degree signals in degree-sensitive tasks, and argue for placing degree scales after LayerNorm based on a simple algebraic fact. The proposed placement improves performance on degree-sensitive tasks without changing other components.
We propose Graph Wavelet Compressed Sensing (GWCS) to offline-compress graph signals using spectral graph wavelet representations, enabling scalable recovery with scale-aware neural reconstructions. This reduces data requirements for scientific machine learning and inverse problems.
CGRL aims to improve node-level OOD robustness by causal-guided representation learning that models latent geometry and disentangles causal from spurious features. It handles node classification under distribution shifts and leverages environment-aware causal mechanisms.
We present an agentic text-to-SPARQL system where a researcher agent, after validating results, proposes and tests changes to prompts, rules, and tool orchestration to improve question answering over knowledge graphs.
GraphThink integrates a task graph and a scene graph to provide structured knowledge for robust planning and environmental memory, reducing planning hallucinations and enabling better long-horizon embodied task planning.
KGCache introduces an in-memory cache for one-hop KG neighborhoods to amortize repeated retrieval in KGQA workloads, compatible with iterative traversal (ToG) and one-shot planning (RoG) paradigms.
VDGR-RAG integrates vector retrieval, directories, graphs, and reflection to unify reasoning over hierarchical enterprise documents, improving domain routing and leveraging document structure for better QA.
SodaMem builds an evidence-grounded temporal graph memory that records typed events with provenance spans, stores mention and occurrence times, and supports currency and ordered temporal reasoning for long-running conversations.
ForestBench provides a general evaluation framework that maps multi-agent system traces to unified collaboration graphs, enabling fair, cross-method comparison without model-dependent inference.
PolicyKG is an end-to-end LLM-based pipeline that converts natural-language institutional policies into machine-readable SHACL constraints. It processes a policy PDF, classifies each sentence as an obligation, permission, or prohibition, lifts these labels into first-order deontic logic, and emits SHACL constraints, all under a LangGraph state machine with per-stage validators. The Corpus Adapter provides a YAML vocabulary registry to ground policy concepts and ensure robust translation.
An explainable Graph Neural Network (GNN) framework is proposed for component-level anomaly diagnosis in industrial multi-sensor settings. It uses multivariate time series data to attribute anomalies to system components rather than single sensors, providing human-understandable root-cause explanations. By modeling the interactions among sensors, it improves interpretability compared with sensor-centric anomaly detection.
ATLAS introduces a coupled graph–policy distillation framework for patient-adaptive medication safety in older adults with multimorbidity. The framework structures guideline evidence as a medication-safety graph and uses targeted questions to update the patient state, distilling relevant relations into a patient-specific medication conflict graph (PMCG). A risk-first multi-agent policy guides safe medication choices by integrating graph reasoning with patient data to tailor recommendations.
GALA+ is a graph-augmented LLM agentic framework for root cause analysis and incident response in microservices. It leverages service dependency graphs to bound exploration, refine diagnoses, and generate actionable incident responses. The approach addresses multi-telemetry heterogeneity and the hallucination tendencies of LLM-based RCA, improving reliability and producing actionable guidance.
KGCaRe proposes a hybrid approach for explainable complex conditional question answering by combining automatic knowledge-graph construction and context retrieval with LLMs. It augments retrieval-augmented generation with both unstructured and structured knowledge extracted from documents and knowledge graphs to improve reasoning and answer accuracy. The system uses neural retrieval plus symbolic reasoning over LLM-generated contexts to provide explanations.
Probabilistic Circuits for Knowledge Graph Completion with Reduced Rule Sets. Rule-based KG completion typically requires tens of thousands of rules. This work learns a probability distribution over rule sets using probabilistic circuits to enable small, cooperative rule sets. It achieves a 70–96% reduction in the number of rules needed to reach baseline performance, enabling scalable and interpretable reasoning.
M$^3$Prune: Hierarchical Communication Graph Pruning for Efficient Multi-Modal Multi-Agent Retrieval-Augmented Generation. We propose M3Prune to reduce token overhead in multi-modal RAG by pruning the hierarchical communication graph among agents. The method decreases cross-agent communication and computational costs while preserving performance. This enables more efficient deployment of large-scale multi-agent RAG systems.
El Agente Gráfico: A Semantic Execution Runtime for Scientific Agents. The paper presents a semantic execution runtime for scientific agents that uses typed execution graphs to enforce admissible scientific state transitions, record provenance, and confine model judgment to explicit decision points. This design ensures valid state progress across heterogeneous computations and provides provenance record. It also uses the same top-level LLM and task rubrics to maintain consistency.
Experience Memory Graph: One-Shot Error Correction for Agents. The work introduces Experience Memory Graph (EMG) for one-shot error correction in LLM agents. By storing experiential memory, it mitigates compounding errors in long-horizon tasks and enables rapid recovery without costly iterative prompt-based reflection. The memory framework supports generalization to new tasks.
ProbSPARQL: Querying Knowledge Graphs with Multi-dimensional, Uncertain Numeric Data. The paper addresses querying knowledge graphs where numeric data are multi-dimensional and uncertain due to sensor-derived measurements, highlighting the lack of native uncertainty handling in RDF/SPARQL. ProbSPARQL proposes representations and query semantics for uncertain numeric data to support reliability modeling and decision making.
ViSR-KGC: Visual Subgraph Reasoning with Vision-Language Models for Multimodal Knowledge Graph Completion. In multimodal KGC, ViSR-KGC leverages visual evidence and vision-language models to reason over visual subgraphs, improving completion accuracy when relation-specific evidence is scarce. It preserves graph structure better than text-only prompting by considering visual-grounded subgraphs.
Studying People to Study AI: Expert Perspectives on the Epistemic Fit and Barriers of Human Research in AI Safety & Ethics. The study reports expert surveys and interviews with AI Safety & Ethics researchers across technical and governance domains. It highlights gaps in the acceptance of human-subject research and identifies practical, ethical, and regulatory barriers to conducting human studies in AI safety and ethics.
HPSO: Particle Swarm Optimization with Hypergraph-Based Topology. The work proposes a particle swarm optimization (PSO) that uses hypergraph-based topology, extending standard ring and star topologies to hypergraphs to enable more complex particle interactions. This leads to improved exploration and convergence, especially for high-dimensional or coupled optimization problems.
GraphReview: Scientific Paper Evaluation via LLM-based Graph Evidence Expansion. GraphReview formulates paper evaluation as inference-time graph evidence expansion over a semantic paper graph, allowing the model to reason with both internal quality signals and cross-paper links. The graph jointly captures intrinsic quality and synchronic connections, enabling unified, evidence-based evaluation across related work.