Showing 18 papers for 2026-05-09
SPARK proposes a self-play reinforcement learning framework for reasoning over scientific literature grounded in knowledge graphs. It uses asymmetric rewards derived from knowledge graphs to guide both problem generation and solution, addressing the weak explicit relational structure in text. By grounding signals in structured relations, SPARK aims to produce verifiable, multi-modal reasoning questions.
Text-Graph Synergy introduces a bidirectional verification and completion framework to enhance retrieval-augmented generation. It couples textual and graph-based evidence to verify relevance and complete reasoning paths, mitigating the information island problem caused by asymmetric evidence and pruning. The approach improves factual grounding and multi-hop reasoning in LLM-based RAG systems.
Knowledge-Graph Paths as Intermediate Supervision for Self-Evolving Search Agents builds on self-evolving search by providing relational context through knowledge-graph paths during self-play. This intermediate supervision reduces the incidence of invalid or unverifiable questions produced by Proposers and helps Solvers reason over richer relational structures, accelerating learning.
Sheet as Token proposes a graph-enhanced representation for multi-sheet spreadsheet understanding, treating sheets as interconnected tokens rather than fragmenting into rows or blocks. The approach builds a sheet-level graph to preserve global semantics across heterogeneous schemas and enables scalable language-model-based spreadsheet understanding. It improves cross-sheet reasoning and data analysis tasks.
Which Are the Low-Resource Languages of the Semantic Web? presents a methodology to quantify language distribution in Linked Open Data graphs and identify low-resource languages. The poster analyzes language coverage across LOD KGs and proposes metrics to define linguistic resource scarcity in semantic-web datasets.
In Data or Invisible: Toward a Better Digital Representation of Low-Resource Languages with Knowledge Graphs focuses on language coverage of Linked Open Data Knowledge Graphs, identifying key variables that shape language distribution and outlining strategies to improve representation of low-resource languages in KGs. It highlights measures such as the number of Wikipedia articles per language edition and the count of language-tagged entities to guide better digital representation.
Graphlets as Building Blocks for Structural Vocabulary in Knowledge Graph Foundation Models proposes using graphlets as elementary structural primitives to form a discrete structural vocabulary for Knowledge Graph Foundation Models. This helps address irregular, non-Euclidean KG topologies by enabling consistent local pattern encoding, improving generalization across diverse graphs.
Knowledge Graphs, the Missing Link in Agentic AI-based Formal Verification demonstrates that generating SVAs from natural-language specs without grounding to RTL is fragile. It argues that incorporating knowledge graphs can serve as the missing link to connect specifications with low-level micro-architectural details, improving grounding and reliability in AI-assisted formal verification.
Patch-Effect Graph Kernels for LLM Interpretability reframes mechanistic interpretability as a graph-learning task by representing activation patching as patch-effect graphs over model components. It introduces three graph-construction methods to build these graphs and uses graph kernels to compare interventions across prompts and tasks, enabling scalable interpretability analyses.
COPYCOP: Ownership Verification for Graph Neural Networks introduces CopyCop, an algorithm to determine whether two GNNs share training lineage or one has copied embeddings, even if the models have different architectures and embedding dimensions. The method remains effective under adversarial embedding transformations and outperforms existing approaches.
Robustness of Graph Self-Supervised Learning to Real-World Noise: A Case Study on Text-Driven Biomedical Graphs evaluates how Graph Self-Supervised Learning methods perform when constructing graphs from text in biomedical domains, where noise is prevalent. The study provides comprehensive robustness assessments, highlighting vulnerabilities and suggesting directions for noise-aware training and evaluation protocols.
A Unified Benchmark for Evaluating Knowledge Graph Construction Methods and Graph Neural Networks introduces a dual-purpose benchmark to jointly evaluate KG construction quality and downstream GNN performance, addressing how graph quality affects learning outcomes. The benchmark emphasizes noisy, fragmented, and semantically inconsistent graphs, and provides tasks and metrics for both construction and learning evaluation.
Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs introduces LATTE, a framework that coordinates language agent teams via adaptive task graphs which evolve during collaboration. The approach aims to reduce error propagation, inter-agent conflicts, and wasted tokens or file operations by balancing structure and flexibility in team tasks.
Towards Metric-Faithful Neural Graph Matching develops a theoretical framework linking encoder geometry to the accuracy of neural graph matching approximations of Graph Edit Distance. The work shows how representation geometry influences GED estimation and provides design guidelines for building metric-faithful matchers.
New Bounds for Zarankiewicz Numbers via Reinforced LLM Evolutionary Search determines exact Zarankiewicz numbers for several parameter settings and provides lower bounds for many others using reinforced LLM evolutionary search. The results tighten the known bounds for K_{s,t}-free bipartite graphs and advance extremal graph theory.
Disentangled Generative Graph Representation Learning proposes disentangled generative graph representation learning to address entangled latent factors caused by broad masking of graphs. The method seeks improved robustness and explainability by separately modeling distinct generative factors in graph representations.
ReMAP: Neural Reparameterization for Scalable MAP Inference in Arbitrary-Order Markov Random Fields introduces ReMAP, a neural reparameterization framework that optimizes a differentiable relaxation of the MRF energy on a per-instance basis. It avoids supervised training and amortization, enabling scalable and accurate MAP inference for arbitrary-order MRFs.
Mochi: Aligning Pre-training and Inference for Efficient Graph Foundation Models via Meta-Learning proposes Mochi, a Graph Foundation Model that uses meta-learning to align pre-training with downstream inference, addressing task unification and training efficiency. The framework argues that standard reconstruction-based pretraining and post-hoc unification can be limiting and demonstrates improvements via joint meta-learning.