Showing 21 papers for 2026-05-29
This paper studies designing active tether-net systems for space debris capture, leveraging graph-learning aided mixed-combinatorial optimization to navigate continuous, integer, and categorical design variables. It highlights how complex design and control choices interact under constraints and proposes a framework to optimize capture performance and robustness.
This work presents an evolutionary refinement of generative graph topologies using a hybrid WGAN-GA approach. It targets reducing deviations in degree and spectral distributions between generated and real graphs, while preserving class-specific structural patterns and graph sizes, by evolving graph topologies to better match real data.
The paper introduces gated graph attention with a learnable temperature, where gating reduces the influence of unreliable feature dimensions and the temperature adaptively sharpens or smooths attention. This enhances robustness across standard graph attention networks.
The authors develop a data-centric theory for graph foundation model transfer, using a graphon-based continuum limit for dense graphs. They show that transfer behavior depends on domain properties and provide explicit characterizations for different tokenizations, explaining when transfer is beneficial or detrimental.
The work proposes Horizon-Aware Graph Neural Networks for long-horizon forecasting, training a multi-horizon emulator that learns state-to-state transitions from a single historical curve. This addresses nonlinear dynamics and error accumulation in long-range predictions, enabling more stable long-term forecasts.
iLoRA introduces Bayesian Low-Rank Adaptation with latent interaction graphs for microbiome diagnosis. It infers a latent interaction graph from input data to produce input-conditioned LoRA updates, enabling joint learning of prediction and latent interactions rather than treating them separately.
OOD-GraphLLM addresses out-of-distribution generalized drug synergy prediction by leveraging graph-based large language models to handle shifts caused by novel compounds and varying molecular structures, aiming to improve generalization beyond in-distribution data.
The paper proposes Better Later Than Sooner, a neuro-symbolic knowledge graph construction method with ontology-grounded post-extraction correction to improve retrieval-augmented generation for complex, multi-hop QA requiring symbolic operations.
PassNet explores scaling LLMs for graph compiler pass generation, arguing that pass-level generation—integrating structured graph transformations into compiler pipelines—addresses tail workloads better than isolated kernel generation.
Temporal Motif-aware Graph Test-time Adaptation develops a test-time adaptation framework leveraging temporal motifs to cope with evolving blockchain transaction patterns, addressing adversarial evolution and out-of-distribution shifts in anomaly detection.
SAHG introduces a Sector-Anisotropic Hyperbolic Graph Model for social bot detection, leveraging hyperbolic geometry to capture hierarchical, scale-free structures and sector-specific anisotropy in social graphs for improved detection.
What drives performance in molecular MPNNs? An operator-level factorial benchmark decomposes MPNNs into operator families (initialization, fusion, update) and benchmarks 84 configurations on MoleculeNet to uncover how each operator affects performance.
Size Transferability of Graph Transformers with Convolutional Positional Encodings studies how graph transformers with GNN-based positional encodings transfer across graph sizes, linking GTs with manifold neural networks and offering transferability insights.
Self-supervised Adversarial Purification for Graph Neural Networks proposes a dedicated purifier module that cleanses inputs before classification, enabling robustness without conflating accuracy and adversarial robustness within a single classifier.
Rel-MOSS tackles imbalanced relational deep learning on relational databases by addressing minority entity under-representation, proposing methods to improve learning when the relational data is highly skewed.
RewardFlow introduces topology-aware reward propagation on state graphs to estimate state-level rewards for agentic RL with LLMs, enabling finer-grained credit assignment along trajectory graphs.
UI-KOBE proposes knowledge-oriented behavior exploration for lightweight graph-guided GUI agents deployed on mobile devices, enabling reliable end-to-end task execution with limited model capacity.
mcp-proto-okn presents a Python-based Model Context Protocol server enabling natural-language access to open scientific knowledge graphs, supporting graph routing, SPARQL, ontology expansion, multi-graph querying, and transcripts.
Graph-Enhanced Policy Optimization in LLM Agent Training analyzes credit assignment in group-based RL for multi-step LLM agents, proposing graph-based policy optimization over an online state-transition graph to better credit states according to their structural roles.
The Compressive Knowledge Graph Hypothesis investigates which graph facts matter for scientific hypothesis generation by perturbing local KGs (density, ontology richness, topology, control structure) and evaluating model outputs across several LLMs, finding KG utility to be selective and model-dependent.
FundaPod introduces a multi-persona agent pod platform that uses a knowledge graph memory to support AI-assisted fundamental investment research. Unlike typical finance LLM work that focuses on trading signals or NLP predictions, FundaPod emphasizes gathering evidence, identifying business drivers, comparing viewpoints, and generating investment memos. The platform aims to produce transparent, reusable, and verifiable investment plans and to advance cumulative expertise in fundamental investing through collaborative reasoning among agents.