Showing 37 papers for 2026-04-21
We develop a physics-informed graph attention network to predict multi-label phase diagrams for complex alloys. The model learns element-aware representations and incorporates thermodynamic information to enable rapid mapping across composition–temperature space, accelerating alloy design.
The paper proposes graph transformer-based pathway embeddings for cancer prognosis. By constructing robust pathway representations from omics data, it improves interpretability and predictive performance over traditional gene-centric approaches.
We explore whether LLM-derived graph priors can improve coordination in multi-agent systems. The work investigates using semantically informed priors from large language models to guide agent coordination, aiming to reduce brittleness and data requirements.
We introduce SigGate-GT to tame over-smoothing in graph transformers. Sigmoid-gated attention alleviates attention-sink and entropy degeneration caused by softmax, improving depth performance on long-range reasoning tasks.
This work studies topological contractions as a learning tool on graphs, simplices, and cells. They define Contraction Homology to analyze persistence of contraction sequences and show theoretical properties that can improve graph representations.
LoReC rethinks using LLMs for graph data analysis. It identifies why vanilla prompting underperforms and proposes directions to better integrate graph structure into LLM-based workflows.
We propose a causally constrained probabilistic forecasting framework for multivariate time series anomaly detection. By embedding causal structure into forecasting and providing probabilistic outputs, it enables root-cause localization and more interpretable anomalies.
The paper investigates generalization of fine-tuned small LMs for graph structural inference along graph-size and graph-family distribution axes. Through controlled experiments with three instruction-tuned models, it maps out the limits and factors affecting transfer.
We model the ionosphere as a dynamic graph over ionospheric pierce points with ephemeris-conditioned topology. Because satellite trajectories are predictable, we condition forecasts on future graph structure to improve predictions of ionospheric irregularities.
We propose a metapath classification framework centered on bridges, using R-GCN-VGAE to classify metapaths and quantify bridges' roles in disaster resilience. The method informs maintenance decisions under budget constraints.
We compare continual pretraining versus Graph Retrieval-Augmented Generation (GraphRAG) for injecting UMLS-derived biomedical knowledge into LMs. The study constructs a large biomedical knowledge graph and evaluates which approach yields better knowledge integration and downstream performance.
We present a scalable and adaptive parallel training framework for graph transformers on large graphs. The approach addresses memory and bandwidth bottlenecks and achieves efficient distributed training.
We introduce a graph-enhanced mitigation (GEM) framework that uses physically informed graph neural networks to perform quantum error mitigation. By incorporating physical structure, GEM improves the estimation of observables on noisy devices at scale.
This study empirically evaluates Reinforcement Learning with Verifiable Rewards (RLVR) under low data and compute regimes using open-source setups. It assesses effectiveness and provides guidance for practicing RLVR when resources are scarce.
A survey of GNNs designed for heterophilic graphs, where connected nodes may have different labels. It reviews models, training strategies, datasets, and highlights challenges and future directions.
The paper investigates how mobility-based graphs and temporal granularity affect GNN-based COVID forecasting. It shows that graph sparsification and suitable time resolution can improve performance relative to strong baselines.
We propose physics-informed GNNs that encode detector geometry and physics observables to estimate transverse momentum in CMS triggers. Four graph construction strategies are compared to balance accuracy and hardware constraints.
We provide a convergence analysis for GNDEs with time-varying parameters in the infinite-node limit and introduce Graphon Neural Differential Equations to study size transferability.
PFΔ provides a benchmark dataset to evaluate power flow computations under variations in load, generation, and network topology. The dataset facilitates testing of contingency analysis and topology optimization workflows.
Torch Geometric Pool introduces a library of 20 pooling operators with a unified SRCL interface for PyTorch Geometric. It standardizes outputs and enables easier benchmarking and reuse.
This paper proposes using a Graph Neural Network to model sea ice dynamics by treating ice blocks as nodes and their interactions, including collisions, as edges. It starts with a one-dimensional setting as a tractable proof-of-concept and introduces the Collision-captured Network (CN) that blends data assimilation with graph-based dynamics to improve efficiency over traditional solvers.
The work develops an LLM-guided, query-aware inference system for GNNs on large knowledge graphs. Since GNN inference queries differ in size and complexity, the system uses an LLM to tailor subgraph selection and computation per query, achieving faster, query-sensitive inference while preserving accuracy.
The authors argue that in text-rich graphs, text content is the primary medium through which relationships are manifested, not merely node attributes. They propose RAMP, a graph kernel that leverages LLMs to perform message passing directly on raw text, enabling more faithful reasoning on text-rich graphs.
ARCS automates autoregressive circuit synthesis to produce SPICE-ready designs in milliseconds. It uses a hybrid pipeline with a graph VAE and a flow-matching model, plus SPICE-based ranking, achieving 99.9% simulation validity across 32 topologies with only 8 SPICE evaluations—far fewer than genetic algorithms.
ST-Sheaf GNN reframes spatio-temporal forecasting as learning information flow over locally structured spaces via sheaf theory. It embeds graph topology into local vector spaces using adaptive local structures to capture heterogeneity in responses to disruptions, improving predictive performance.
The paper discusses replicating AI research using LLM agents and executable knowledge graphs as representations. It argues that current methods miss latent technical details and signals, and proposes executable KGs with structured, multi-granular retrieval to improve reproducibility and knowledge reuse.
This survey reviews how graphs can be integrated with LLMs for reasoning and retrieval, offering a taxonomy of design choices, and clarifying when graph-augmented LLMs are advantageous across tasks like reasoning, retrieval, generation, and recommendation.
The authors propose using knowledge graphs to improve ML interpretability in manufacturing, connecting domain data, model results, and explanations. A selective retrieval mechanism fetches relevant triplets to present explanations in a form accessible to users.
An exascale workflow for atomistic graph foundation models built on HydraGNN, trained on 16 datasets with multi-task heads and scalable data pipelines. Demonstrates large-scale training on Frontier and hyperparameter optimization, yielding PaiNN-based leading models.
The Amazing Agent Race introduces a benchmark of DAG-like tool-use puzzles with 1,400 instances, requiring agents to navigate Wikipedia, perform multi-step tool chains, and aggregate results. It contrasts sequential and compositional variants to test tool-use strength.
Diagnoses LLM-based cross-encoder rerankers in cold-start movie recommendation, using Serendipity-2018 with 500 users across seeds to identify coverage and exposure gaps and proposes practical mitigations.
Evaluates whether LLMs can support post-publication research evaluation by comparing their recommendations to expert judgments and citation indicators, using articles from H1 Connect as test cases.
GraphRAG-Router learns to route retrieval over GraphRAGs and LLMs with reinforcement learning to adaptively select retrieval strategies based on query complexity and cost, achieving cost-efficient reasoning.
Proposes Tensor Manifold-Based Graph-Vector Fusion for AI-native literature retrieval, using tensor manifold theory to fuse graph and vector representations while addressing storage, semantic dilution, and AI-native integration.
Self-Distilled Reinforcement Learning for Co-Evolving Agentic Recommender Systems introduces a self-distillation mechanism to RL-based ARS, moving beyond simple Reflexion-style memory towards distillation-based learning to improve long-term interaction quality.
CPGRec+ is a balance-oriented framework for personalized video game recommendations that extends CPGRec by accounting for disparities in player-game interactions and addressing over-smoothing in recommendations.
MegaRAG introduces multimodal knowledge graph-based retrieval augmented generation to improve reasoning over long-form content by leveraging a knowledge graph to provide structured, entity-centric context.