Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 246 papers for ICML 2026
AgentConductor investigates how to adapt the density of communication topology among agents in LLM-driven multi-agent systems and refine it during task execution using feedback. By making topology density task-aware and updating it within an instance, the approach reduces redundant communication and improves competition-level code generation performance.
SAOT proposes Structure-Aware Optimal Transport to preserve global relational structure in self-supervised continual graph learning. It addresses the drift of inter-node correspondences that arises when learning from a graph sequence without labels, going beyond node-level consistency objectives. The approach improves stability and retention of relational structure across tasks.
GFFMERGE introduces a closed-form framework for merging Graph Neural Network-based force fields by exploiting the linearity of message passing. The merging is formulated as a convex embedding-alignment problem with an analytical solution, enabling efficient model merging and benchmarking.
SLAE presents a strictly local all-atom protein representation by modeling each residue's local atomic environment using only atom types and interatomic geometries. It employs a multi-task autoencoder that jointly reconstructs coordinates, recovers sequence, and regresses energy, enforcing physically meaningful features.
GraphFlow proposes wGraph to represent workflows as a unified graph for efficient LLM-agent serving. Each node corresponds to a workflow element (step or operation), and edges encode semantic and data dependencies, enabling better capture of deep relationships and generalization beyond templates.
FAHNES introduces a feature-aware hierarchical graph/hypergraph generation framework that jointly models topology and node/edge features. It builds multi-scale representations via node coarsening and next-scale prediction to scale generation to large graphs and hypergraphs.
Causal Preference Elicitation provides a Bayesian expert-in-the-loop framework for causal discovery with active querying. It models expert judgments with a three-way likelihood over edge existence and direction, uses particle-based posterior approximations, and selects queries via expected information gain.
SE3Set develops SE(3) equivariant hypergraph neural networks for molecular representation learning. By constructing hypergraphs that capture high-order interactions via a fragmentation method incorporating chemical and 3D spatial information, SE3Set advances beyond pairwise graphs.
Rethinking GNNs and Missing Features analyzes limitations of missing node features in real-world domains; shows that high sparsity can limit the information loss and complicate fairness comparisons under MCAR assumptions; proposes a robust evaluation and practical remedy.
Physics-informed coarsening for multigrid GNN surrogates in solid mechanics couples an encoder-processor-decoder backbone with physics-informed coarsening to respect nonlinear elasticity, plasticity, and transient dynamics, enabling accurate and scalable surrogates.
Beyond Trajectory-Level Attribution introduces GraphGPO, a graph-based group policy optimization method for agentic RL. It enables step-level credit assignment by attributing contributions within groups and leveraging graph structure to improve learning over mere trajectory-level signals.
Failure is Feedback proposes History-Aware Backtracking for agentic traversal in multimodal graphs. It treats subgraph retrieval as sequential decision-making and uses history of failures to backtrack and recover from dead-ends, improving retrieval effectiveness.
MOC proposes Multi-Order Communication for LLM-based MAS, reconstructing communication beyond first-order neighbor responses to widen the evidence receptive field and reduce information dilution across multi-hop paths.
PromptDyG enables test-time prompt adaptation for dynamic graphs, addressing distribution shifts and evolving complexity that offline training on historical snapshots cannot capture.
Graph Neural Networks Are Not Continuous Across Graph Resolutions reveals that GNNs can produce different latent representations for structurally-equivalent graphs at different resolutions, due to a structural obstruction in common information-propagation schemes, and provides a principled remedy.
HieraMAS introduces hierarchical agent collaboration with intra-node LLM mixtures and inter-node topology; within each functional role, multiple heterogeneous LLMs form a 'supernode' to diversify capabilities, improving coordination across MAS.
NaviAgent presents graph-driven bilevel planning to scale tool orchestration for LLMs; decouples task planning from tool execution using a graph of tool relations, enabling planning across hundreds or thousands of tools.
Discriminative Attribute Graph Clustering uses topology-guided contrastive learning to produce discriminative node representations in attribute-rich graphs, addressing noisy edges and redundant multi-view features.
MultiHal provides a multilingual dataset for knowledge-graph grounded evaluation of LLM hallucinations, using knowledge graphs to provide structured facts and enable cross-lingual assessment of factuality.
Full-Spectrum GNNs generalize spectral GNNs to second-order signals by lifting node signals to the node-pair domain and applying multi-dimensional spectral filters, offering greater expressivity and scalability.
This work tackles high-order brain connectivity by learning a multi-scale hypergraph to capture interactions among multiple brain regions beyond pairwise links. It aims to improve early neurodegenerative disease classification (AD, PD) by modeling higher-order dependencies rather than relying on predefined hyperedges or fixed edge weights.
Introduces CLINIC for graph out-of-distribution detection, leveraging contextual affinity and twin concordance to capture semantic structure beyond topology. It aims to identify anomalous graphs by exploiting context-aware relationships.
Proposes Structure-Centric Graph Foundation Models (SCGFM) that treat topology as the main transferable knowledge. Graphs are modeled as metric measure spaces; introduces learnable geometric bases defining a shared structural coordinate system; alignment via Gromov-Wasserstein distances yields structure-aligned latent representations that accommodate heterogeneous graphs.
Analyzes discriminative Mixture-of-Experts on graphs and discovers two issues: expert discrimination loss causing homogenization and routing collapse to a few experts, hindering diverse semantics; addresses unreliable routing.
Proposes D^3: Dynamic Directional Graph-Constrained Data Scheduling for LLM Training. Argues training data interactions affect optimization; models directional influences among samples with a graph; prioritizes train-units with greater influence to improve learning efficiency.
Rel-MOSS tackles class imbalance in relational deep learning on relational databases by proposing relation-centric minority synthetic oversampling to under-represented minority entities.
Graph-Link bridges semantic-structural gap in Text-to-SQL by constrained subgraph induction; reframes schema linking from independent retrieval to constrained subgraph induction to preserve bridge tables for multi-hop joins.
Introduces DyGMoCE, a dynamic graph transformer with a mixture of curvature-aware experts to handle diverse and evolving topologies (hierarchical, grid-like, cyclic). It uses multiple curvature spaces to better represent local geometries; dynamic routing among curvature-aware experts improves representations.
Revisits vector quantization for graph representation learning; show codebook collapse occurs when training VQ jointly with GNNs on graph reconstruction; discuss implications and potential remedies.
Proposes principled zero-shot ranking using tournament graphs; each k-document comparison yields a complete tournament of pairwise preferences; uses tournaments to perform k-wise reranking in a principled manner.
Large-scale MD simulated with direct interatomic modeling via dilated message passing; a scalable, accurate framework for interatomic interactions; stacks multiple message-passing layers with dilation to capture long-range interactions efficient for full-atom MD.
CRAMER enables editing recommenders by control via request-aware masking, allowing adaptation to immediate user requests without retraining or LLM prompting; reduces computational overhead.
Graph Masked Auto-Encoder studies show that reconstruction outcomes can be similar even with different neighborhood context usage; proposes C2-GMAE to enforce consistency and coverage in graph reconstruction by better leveraging neighborhood context.
Proposes graph alignment-based benchmarking for GNNs; frames as self-supervised learning; creates datasets using synthetic and real graphs with increasing difficulty to rank models and positional encodings.
GP2F: Cross-domain Graph Prompting with Adaptive Fusion of Pre-trained Graph Neural Networks. Observes GPL effectiveness under domain shifts; proposes adaptive fusion of pretrained GNNs for cross-domain prompting.
Selecting Samples on Graphs: A Unified Dataset Pruning Framework for Lossless Training Acceleration. Proposes a unified framework combining intrinsic and extrinsic signals to prune dataset on graphs while preserving training losslessness; robust across pruning ratios and distribution.
DANCE: Dynamic, Available, Neighbor-gated Condensation for Federated Text-Attributed Graphs. Addresses TAG-FGL with high LLM overhead; propose graph condensation to reduce computation; dynamic neighbor gating condensation for federated text-attributed graphs.
Unsat Core Prediction through Polarity-Aware Representation Learning over Clause-Literal Hypergraphs. Proposes polarity-aware representations on clause-literal hypergraphs to capture clause-level and higher-order interactions along with polarity.
From Distribution to Geometry: Stable Graph Generalization via Invariant Barycenters. Proposes DIGL, geometrizing invariance using optimal transport to improve OOD generalization of GNNs.
Deep Multi-view Graph Clustering via Attribute-aware Bidirectional Structural Refinement and Pseudo-label Guided Multi-level Fusion. Proposes ABSR and PGMF for DMGC; ABSR strengthens good connections and suppresses conflicting ones; PGMF uses pseudo-labels for multi-level fusion across views; yields improved clustering.
STPGC introduces scalable topology-preserving graph coarsening by borrowing concepts of graph strong collapse and edge collapse from algebraic topology. It provides three algorithms—GStrongCollapse, GEdgeCollapse, and Neighborho...—to preserve topology while reducing graph size, enabling efficient GNN training on coarsened graphs.
DCD-GNN presents dual parallel channels: a static structure-preserving channel and a dynamic adjacency learning channel, enabling adaptive propagation over latent relationships. It also performs multi-scale representation fusion to capture information at different granularities.
A consensus anchor-guided hypergraph framework for scalable incomplete multi-view clustering leverages hypergraphs to capture high-order correlations beyond pairwise relations and accounts for distributional shifts between observed and missing views to reduce bias in consensus learning.
Demonstrates that ensemble-consensus-based semi-supervised learning improves molecular property prediction where labeled data are scarce. It avoids fragile label-preserving augmentations and leverages agreement among multiple models to guide learning.
ST-TGExplainer targets temporal GNN interpretability by disentangling stability patterns (historic interactions) and transition patterns (new, first-time interactions). It helps identify which past and emerging interactions drive predictions, enabling faithful explanations.
Introduces Dune, a directed unitary GNN with edge features that retains the stability guarantees of Unitary Convolutions while incorporating edge directionality. This approach mitigates oversmoothing and gradient issues in deep models for directed graphs.
LLM-MatLogic defines Executable Exchange Contracts to handle scoped negation in knowledge-graph question answering. The executor applies scope-local masks during multi-hop propagation to reliably enforce exclusions, improving performance on exclusion-heavy queries.
Variational Bayesian Flow Network (VBFN) for graph generation combines Bayesian flow ideas with variational methods to model joint distributions over discrete node and edge attributes. It addresses coupling between node and edge attributes and provides a robust alternative to diffusion/flow-matching approaches.
Examines theoretical limitations of embedding-based link prediction by analyzing rank bottlenecks in the output layer that constrain expressivity. The paper discusses bounds on embedding dimension needed to fit training data in knowledge graph embeddings.
HONet integrates a structure-preserving hypergraph encoder with a differentiable fixed-constraint quadratic programming layer to solve Exact Cover tasks data-efficiently. The Fixed Polytope paradigm, guided by Geometric Consistency Loss, shapes the optimization landscape to enforce known constraints.
Studies MaxECC on edge-colored r-uniform hypergraphs, aiming to color vertices to maximize satisfied edges. The paper provides an improved 1/(r+1) approximation and introduces a novel dependent-round algorithm.
Proposes One-Step Graph-Structured Neural Flows (GSNF) for irregular multivariate time series classification. It introduces auxiliary trajectory self-supervision to better capture inter-variable interactions within a single-step mapping.
Demonstrates that simple 2-hop node connections in graph rewiring for fluid simulation can match or surpass long-range connections, improving information propagation in GNN-based simulators.
Argues plain Transformers are strong link predictors, often outshining GNNs with heavy structural encodings. They show that removing elaborate structure encodings can still yield competitive or superior performance.
FlatLand proposes personalized graph federated learning in Lorentz space to tailor graph models to individual users.
Hierarchical Anchor Graph Learning for Multi-View Clustering (HAG-MVC) builds a pyramid of anchors across views to capture multi-granularity structure, improving scalability and clustering performance.
Which Algorithms Can Graph Neural Networks Learn? studies neural algorithmic reasoning in GNNs, evaluating whether MP-based models can learn discrete algorithms and discussing guarantees and limits of expressivity.
GFMate introduces test-time prompt tuning for Graph Foundation Models to improve cross-domain generalization, addressing prompt entanglement with source-domain pretraining.
Investigates what makes a relational graph suitable for relational deep learning; shows schema-derived graphs suffer information overload and semantic fragmentation; proposes task-dependent balance to prune task-irrelevant structure while preserving semantics.
CausalRAG2 designs hierarchical causal knowledge graphs for retrieval augmented generation, addressing reliance on entity-centric matching and enabling explicit causal modeling across modules for scalable reasoning.
GoS introduces Graph of States to solve abductive tasks with LLMs. It uses explicit state graphs and backtracking control to avoid evidence fabrication, context drift, and other failure modes, enabling robust abductive reasoning with large language models.
We study tool-calling in FHIR with RL, creating agents that navigate FHIR resource graphs to answer clinical questions. Our method improves resource selection and adherence to traversal constraints on the FHIR-AgentBench benchmark.
We propose ReDoS-MotifGNN (RMGNN), a motif-based heterogeneous GNN for detecting ReDoS vulnerabilities in regex patterns, including extended features like lookarounds and backreferences. The model balances static efficiency with dynamic detection accuracy and reduces false positives.
X-EviProbe provides a post-hoc, parameter-free method to quantify evidential uncertainty for frozen GNNs. It probes latent space and native outputs to build class-wise Dirichlet evidence, separating epistemic and aleatoric uncertainty without retraining.
GiFlow is a Graph-Informed Flow Matching framework for spatiotemporal imputation. It alleviates error propagation of recurrent and graph neural approaches by using flow-based matching guided by graph information, improving efficiency and accuracy.
HG-Rec introduces Hyperbolic RQ-VAE for generative recommendation, embedding latent codes in hyperbolic space to better capture hierarchical relationships. It improves residual quantization with a differential-length codebook strategy to enhance codebook utilization.
We propose graph-aware LoRA generation to enhance LLMs for graph tasks. By injecting information at weight level, the approach enables whole-graph information encoding and improves transferability across datasets.
We present Uncertainty-Modulated Dual-Path Diffusion for robust text-attributed graph learning. The method addresses structure–semantics mismatch and dual-source noise in node text and graph structure, yielding more robust representations.
We propose a dynamic cutoff for MLIPs instead of a fixed radius, enabling stable long-timescale molecular dynamics. The dynamic cutoff improves accuracy and reduces memory consumption without sacrificing stability.
MiSe-GNN introduces a dual-head architecture for error-driven graph augmentation on mesh-based PDE surrogates. It adaptively adds long-range edges prioritizing regions with high error, improving accuracy in challenging physics like shocks.
Graph-GRPO trains Graph Flow Models with reinforcement learning under verifiable rewards. It derives an analytical expression for transitions and aligns GFM sampling with task objectives via online RL, improving training stability and objective adherence.
From Retrieval to Translation: Translating Query into Graph-level Clues for Retrieval-Augmented Generation, KG-Translator.
S^3GNN revisits spectrum-based global mixing and local message passing for long-range graph learning. It analyzes spectral filtering and Jacobian stability to enable strong long-range dependencies.
Learning Adaptive Topology with FiLM-Guided Distillation for Tertiary Structure-Based RNA Design, ATL-FGD.
Scaling the Prior: Size-Consistent Geometric Diffusion for 3D Molecular Generation.
What Preferences Can—and Cannot—Predict in Multi-Agent Online Learning.
Same Graph Cross-Task Transfer in GNNs: Protocols and Predictors.
Progressive Graph Structure Adjustment for Homophily Shift Adaptation, PSAHS.
Revisiting Asymmetries in Black-box Link Stealing against Graph Neural Networks.
Position: Graph Condensation Needs a Reset—Move Beyond Full-dataset Training and Model-Dependence.
We propose SuperHype, an exact hypergraph generator based on graph-superposition decomposition. It tackles intractability and information loss in hypergraph synthesis by enabling high-quality generation with preserved structural patterns and graph-level validity.
We introduce SeRAG, a token-free hierarchical indexing framework for retrieval-augmented generation. It replaces textual summaries with an information-theoretic knowledge taxonomy and builds a multi-perspective graph of semantic, logical, and sequential dependencies, minimizing structural entropy to produce a topologically faithful encoding tree.
UniRTL unifies code and graph modalities for robust RTL representation learning. It leverages the control data flow graph (CDFG) to capture structural information and the code modality to encode semantics, enabling complementary signals and improved generalization.
We propose MeLA, a meta-objective-based framework for robust training of Hypergraph Neural Networks. It uses the hypergraph Laplacian to enable gray-box, structural, and feature perturbations within a unified objective, improving resilience to adversarial perturbations.
RADE introduces Random Add-Drop Edge, a stochastic augmentation that both deletes and adds edges during training. This regularizer addresses overfitting and over-squashing by maintaining informative connectivity, with theoretical justification and empirical gains.
We present Multi-scale Explainer for Graph Neural Networks. Unlike single-scale explainers, it identifies influential structures at multiple graph scales to capture multi-level semantics and improves stability and reliability of explanations.
GOCM offers a single-step graph outlier synthesis framework based on an Origin Consistency Model. The Origin Consistency mechanism enables efficient outlier generation, addressing class imbalance without costly multi-step diffusion, and boosting outlier detection performance.
C-FREE proposes contrast-free representation learning for molecular graphs by fusing 2D graphs with ensembles of 3D conformers. It avoids hand-crafted augmentations and heavy generative objectives, enabling effective pretraining with limited labels and improving downstream predictions.
We analyze hierarchy–uniformity in hypergraph contrastive learning and identify a Hierarchy–Uniformity Conflict that causes semantic flattening. We propose mechanisms to preserve semantic depth while maintaining uniformity, leading to richer hyperedge representations.
MC-HNN introduces a Multi-Channel Hypergraph Neural Network to avoid rank collapse and to model latent hyperedge semantics. It uses multi-channel message passing to maintain high-rank representations and capture latent hyperedge information, improving performance on hypergraph tasks.
TriForces is a model-agnostic three-stream framework that separates composition and structure information and employs self-supervised learning to preserve transferable representations across chemistries. This design enhances the transferability of atomistic neural potentials.
FusionCell is a dual-modality predictor that cross-attends to fuse layout geometry and netlist topology for standard-cell performance prediction. It captures layout-dependent effects and structural connectivity to predict delay and power more accurately than geometry- or topology-only predictors.
HInT injects hypergraph structure into the structural layers of decoder-only LLMs to improve table understanding. It leverages the observation that row and column relations concentrate in a subset of layers and attention heads, enabling more accurate table reasoning during autoregressive decoding.
Geometry-Guided Generative Representation for Functional Brain Graphs models functional brain connectomes as lying on a low-dimensional latent geometry. It uses joint topological and spectral priors to generate graph representations and improve brain-graph modeling with limited data.
NonZero proposes interaction-guided exploration for multi-agent MCTS. It performs surrogate-guided selection over a low-dimensional nonlinear representation, ranking single-agent deviations by predicted gain and two-agent deviations by interaction effects to enable tractable search under budgets.
We learn the interaction prior for protein-protein interaction prediction with a model-agnostic approach that encodes a biology-inspired L3 rule: multiple length-3 paths between proteins suggests interaction. The classifier combines this prior with generic representations to improve PPI prediction.
A Tale of Two Graphs introduces DualGraph memory to separate knowledge exploration from outline structure in open-ended deep research. This decoupled memory supports dedicated exploration and planning traces, improving long-horizon reasoning and evidence synthesis.
Backjump-on-Graph empowers LLMs with reinforced retrospective exploration for knowledge-graph reasoning. It addresses context-window limits and schema gaps by allowing backtracking along the KG to recover from dead ends, enabling more robust KG-based QA.
Can Computational Reducibility Lead to Transferable Models for Graph Combinatorial Optimization? We develop a GCON-based encoder with energy-based unsupervised losses that performs well across CO tasks when trained per-task, and demonstrate transferability by leveraging reducibility ideas to cross-task generalization.
Identifying Common Hubs in Multiple Gaussian Graphical Models introduces JIC-HD, a method to recover common hub variables across related GGMs without fully estimating each graph. It enables joint hub detection across sub-populations and reveals shared structure.
Electronic health records pose challenges for modeling many diseases, especially when outcomes are rare and share risk factors. We propose a Bayesian hypergraph inference framework that represents latent, risk-factor-modulated disease pathways. In this view, risk factors act on hyperedges, enabling interpretable structure with principled uncertainty quantification over disease risk.
Relational deep learning often uses full-resolution graphs to preserve relational semantics, but many methods rely on fixed schemas. We propose FROG, a learnable full-resolution graph structure learning framework for relational DL on relational databases. It enables end-to-end optimization of graph structure, improving relational modeling beyond fixed graphs with empirical gains.
Crystal property prediction is data-hungry with scarce labeled data. We introduce latent-diffusion pretraining to learn crystal representations and then fine-tune for property prediction. The pretraining framework improves data efficiency and predictive accuracy in crystal property tasks.
We introduce Collaborative Likelihood-Ratio Estimation over graphs, where each graph node observes data from two node-specific densities p_v and q_v. Our non-parametric collaborative framework leverages graph structure to share information across nodes and estimate density ratios r_v(x)=q_v(x)/p_v(x).
We present FLAG, a diffusion-based framework that reframes spatial gene expression prediction as structured distribution modeling. It aligns latent diffusion representations on a graph to preserve gene coordination and spatial patterns. The approach tackles the high-dimensional Gene Dimension Curse by jointly modeling gene expressions and their spatial relationships.
We study whether common substructures in graphs are transferable under a foundation-model paradigm. Moving beyond discrete substructures, we analyze transferability from the perspective of functional behavior and geometry. We propose a Riemannian Graph Foundation Model with Neural Vector Bundles to capture transferable structures and provide theoretical insights.
BLIPs introduces Bayesian Learned Interatomic Potentials to provide predictive uncertainty for MLIPs. This enables better handling of out-of-distribution data and data-scarce regimes, guiding active learning and improving alignment with quantum calculations.
AgentWebBench is a benchmark for evaluating how user agents synthesize answers by interacting with website-specific content agents. It covers four tasks, including ranked retrieval and answer synthesis, to assess multi-agent coordination in the Agentic Web setting.
Debate2Create (D2C) formulates robot co-design as structured multi-agent debate guided by physics-based evaluation. A design agent and a control agent debate, with judges giving multi-objective feedback to steer exploration. On MuJoCo locomotion benchmarks, D2C achieves top performance among LLM-based and black-box baselines with substantial gains.
We propose a graph-centric attention pipeline for code analysis where a project is parsed into a code graph, an LLM encodes nodes with semantic and structural signals, and a GNN is trained under sparse supervision for initial malicious-code detection. The GNN outputs guide further interpretation to localize malicious behavior.
RADAR introduces redundancy-aware diffusion to generate and adapt multi-agent communication structures over time. This enables finer-grained topology exploration and robustness while controlling token usage in large language model driven systems.
We propose graph-based conformal compression to produce compact subgraphs that preserve statistical validity. Compression is framed as selecting the smallest subgraph that captures a prescribed probability mass, reducing to a weighted densest-k-subgraphs problem on hypergraphs.
We revisit Pre-Propagation GNNs and introduce robust diffusion operators along with hidden-state re-propagation to enhance expressivity. These improvements reduce the gap with standard message-passing GNNs, especially on heterophilic graphs.
GraphP-FL is a personalized federated graph learning framework with dynamic structure awareness and Fisher information elastic alignment. It includes self-supervised topology reconstruction to adapt graphs across clients and Fisher-information-based alignment to preserve important parameters during aggregation.
We rethink graph coarsening with a non-selfishness principle that considers collective neighborhood interference during coarsening. This yields more efficient and structure-preserving coarse graphs compared to selfish pairwise matching.
Anchor-guided hypergraph condensation distills large hypergraphs into compact synthetic ones. It uses dual-level discrimination to jointly optimize structure generation and condensed features, addressing misalignment in decoupled training.
We propose IN2R, a framework that rectifies inter-modal noisy correspondence through graph-based intra-modal reasoning. By leveraging intra-modal neighbors, it reduces single-point fragility and discretization error in cross-modal retrieval.
FedSSA tackles heterogeneity in graph federated learning by semantic and structural alignment. It infers class-wise node distributions with a variational model, clusters clients accordingly, and aligns local distributions with cluster representatives, addressing feature and structural heterogeneity.
Quantile-free uncertainty quantification in graphs is achieved with QpiGNN, which uses quantile regression to directly optimize coverage and interval width without calibration. This yields reliable prediction intervals without requiring quantile-specific tuning.
We present a brain-function based ADHD detection model that combines short-term and long-term encoding via a memory graph network. A novel brain map sequence from short-term windows and a short-term state encoder capture immediate patterns, while memory graph reasoning integrates longer-term dependencies.
GRASP introduces agentic graph reasoning for LLMs by interleaving on-demand neighbor retrieval with a deterministic Code Interpreter, enabling autonomous probing and computation over large graphs beyond context window limits. It uses staged reinforcement learning to train the agent to select probes and solve subproblems, mitigating structural hallucinations and numerical errors.
TANGO frames graph representation learning as dynamical systems with a learnable energy landscape, where node features descend along the energy gradient to ensure stability via a Lyapunov function. It adds a tangential component, learned via message passing, that updates features without increasing energy, offering flexible dynamics.
Uni-DHC provides a unified framework for deep hypergraph clustering that goes beyond homophily by learning propagation strategies adaptable to heterophilic relations; it relaxes fixed propagation and jointly learns representations to better capture high-order relationships in hypergraphs.
The work proposes watermarking GNNs for ownership protection by embedding ownership signals via explanations, addressing gaps where backdoor-based methods can cause ambiguity and be vulnerable to data poisoning. Results show robust verification without harming model integrity.
Executable Agentic Memory (EAM) builds a structured knowledge graph for GUI agents to plan via retrieval and execution rather than free-form generation; uses a memory construction pipeline with state-aware DFS and action-group mining to compress multi-step routines, and a value-guided planning mechanism to improve search efficiency.
RSF-GLLM decouples differentiable reasoning from answer generation in multi-hop KG QA; the Recurrent Soft-Flow module propagates continuous relevance scores with a GRU-guided updater and a dynamic gate to traverse semantically distant bridge nodes, improving robustness to lexical gaps.
H2CL introduces heterogeneity-aware hypergraph contrastive learning to cope with noises from heterogeneous node–hyperedge interactions; it designs contrastive objectives and augmentations that respect heterogeneity, yielding robust representations on hypergraphs.
The paper rethinks Graph Transformers as graph signal denoisers, showing that incorporating a block-diagonal prior acts as a desirable structure for denoising graph signals; provides theoretical insights and design guidelines for better propagation operators.
SEMIR presents Semantic Minor-Induced Representation Learning to segment small, sparse structures in large images by decoupling inference from the native grid, learning task-aware representations that emphasize minority structures and reduce computation compared to full-resolution inference.
The study analyzes the expressive power required for GNNs to recover optimal solutions of linear SDPs, showing limitations of standard GNNs in solving such problems and suggesting the need for more expressive architectures to capture SDP structure.
L2G-NET introduces Local-to-Global Spectral GNNs by factoring the graph Fourier transform into subgraph operators linked by Cauchy matrices, enabling scalable spectral filtering that combines local locality with global spectral information.
Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents proposes MRAgent, modeling memory as a Cue–Tag–Content graph and adding active reconstruction to revise memory during reasoning, enabling dynamic retrieval and synthesis of past evidence.
iLoRA is a Bayesian low-rank adaptation framework with latent interaction graphs for microbiome diagnosis; it infers a latent interaction graph from input to condition LoRA updates and jointly learns prediction and interaction structure.
Adaptive Node Feature Selection for GNNs proposes training-time feature pruning by measuring feature contributions via validation performance changes, enabling data-, model-, and task-agnostic feature reduction during training.
Mesh-Based Simulations with Spatial and Temporal Awareness advocates ML surrogates for CFD that incorporate spatial-temporal awareness and respect stiff dynamics and local flux continuity, addressing shortcomings of node-wise supervision and explicit Euler time-stepping.
E2Former-V2 delivers on-the-fly equivariant attention by combining algebraic sparsity with hardware-aware execution; introduces Equivariant Axis-Aligned Sparsification (EAAS) using an SO(3)→SO(2) basis change to avoid dense edge computations in 3D data.
Fixed Aggregation Features can Rival GNNs shows that training-free FAF transforms graph tasks into tabular problems; with well-tuned MLPs on FAFs, they rival or surpass GNNs on many benchmarks, offering interpretability and deployment simplicity.
Bipartite Graph Attention-based Clustering for Large-scale scRNA-seq Data proposes a bipartite attention mechanism to cluster cells efficiently at scale, reducing quadratic attention complexity and enabling effective clustering for millions of cells.
DTKG frames multi-hop QA with Dual-Track knowledge graph-verified reasoning, combining parallel fact-verification and chained reasoning; addresses limitations of purely LLM-based or path-based methods by jointly leveraging LLM verification and KG paths to improve accuracy and efficiency.
Beyond Explicit Edges introduces INSES, a dynamic framework for robust reasoning over noisy and sparse KGs by LLM-guided navigation to prune noise and embedding-based similarity expansion to reach non-explicit relations, enabling reasoning beyond explicit edges.
We introduce Equivariant Noise GNNs (ENGNNs), a framework that aims to reconcile universal expressivity with practical performance in graph neural nets. By leveraging the equivariant use of noise within GNN layers, ENGNNs enhance expressive power without prohibitive computation, enabling more general graph tasks.
We frame efficient scaling of GNNs around IO-aware layer implementations. We classify common layers into SpMM-based convolutions, reduction-based aggregations, and attention-based layers, and design GPU kernels to reduce data movement and improve scalability on large graphs.
Conformal Path Reasoning (CPR) provides trustworthy KGQA with path-level calibration using Conformal Prediction. It addresses invalid calibration and weak score discrimination in prior methods to produce prediction sets with coverage guarantees.
VecMol models 3D molecules as continuous vector fields over space, with vectors pointing toward nearby atoms and implicitly encoding geometry–chemistry coherence. This vector-field representation alleviates heterogeneity in discrete graph co-generation and improves learning of spatial structure.
SFCLTA introduces Spectral Fusion Contrastive Learning with Topology-Adaptive Graph Augmentation to better handle heterophilic graphs. It tackles potential distribution shifts from augmentations and enables robust representations by fusing spectral information with topology-aware augmentations.
TIDE proposes Tri-Component Information Decomposition, decomposing info into feature-specific, structure-specific, and joint components. This decomposition helps detect out-of-distribution shifts in graphs, improving OOD robustness.
We study learning-based approaches to approximate Uniform Facility Location (UniFL), a hard combinatorial problem. We compare GNN-based heuristics with classical approximation methods, discussing differentiability, scalability, and potential guarantees.
Toward Effective Multimodal Graph Foundation Models, we propose a Divide-and-Conquer approach to better model interactions across modalities in Multimodal-Attributed Graphs. The work identifies limitations of existing MGFM approaches in explicit cross-modal interaction modeling and modality integration, and offers a scalable strategy to capture rich cross-modal semantics.
Clustering as Reasoning reframes Chain-of-Thought graph learning as a clustering process, offering a k-Means interpretation of iterative reasoning on graphs. It proposes a unified framework to enable step-by-step semantic-topological interaction and improve interpretability and effectiveness of graph-based CoT methods.
GiG is a planning framework that uses graph-informed action generation guided by LLMs to decompose high-level intents into sub-goals in embodied tasks. It emphasizes coherence across long horizons and constrains actions to environmental dynamics, addressing issues like state-transition hallucinations and context-window limits.
DREAM introduces Dual-Standard Semantic Homogeneity with Dynamic Optimization to learn robust GNNs under label noise. It distinguishes reliable vs unreliable nodes and leverages graph topology to maintain semantic consistency under noisy labels.
This work investigates open-world graph learning with unlabeled data uncertainty and fresh perspectives on handling unlabeled nodes that may belong to known or unknown classes. It discusses strategies beyond node-level OOD detection and conventional open-world assumptions to manage uncertainty in TAGs.
CGFM introduces Condition-Aware Graph Flow Matching to model full distributions of complex fluid systems across varying conditions. It combines diffusion models and flow matching to generalize across the condition space and handle large irregular geometries with sharp gradients.
LEGO is an LLM-enabled hierarchical optimizer for tensor computation graphs, enabling structure-aware search and compositional synthesis. It uses a Top-Down construction to decompose graphs into sub-problems for parallel exploration and Bottom-Up mutation to fuse verified sub-plans for local/global performance.
EpiTwin develops spatiotemporal graph transformers to reconstruct epileptic sEEG signals with improved spatial awareness for sparse electrode layouts. It leverages electrode spatial information in encoding and modeling to enhance reconstruction accuracy on irregular, sparse sampling.
The paper analyzes truncated positional encodings for GNNs, noting spectral vs random-walk PEs are equivalent in expressive power when complete. However, truncation is common due to cost; the work studies the implications and practical recommendations.
The authors present an OT-guided stochastic control framework for solving graph combinatorial optimization via exact multilinear relaxation of QUBO. They treat the objective as an energy function and perform sampling to optimize in a highly nonconvex landscape, leveraging optimal transport ideas.
The work models spectral energy shifts to detect both increasing and decreasing spectral variation, including camouflaged anomalies in spatio-temporal graphs. It introduces a node-level spectral energy formulation compatible with message passing for detecting camouflaged anomalies.
The paper introduces ICGNN to robustify GNNs under label noise by exploiting graph structure to identify and mitigate mislabeled nodes via influence contradiction. It uses structure-aware mechanisms to distinguish reliable vs unreliable labels and improve robustness.
InvGNN introduces invertible node representations by designing invertible GNN layers; stacking them yields fully invertible GNN models. The approach enables reconstructible node representations and potential new applications requiring invertibility in graph learning.
We propose AllScAIP, an attention-based, energy-conserving MLIP that scales to around 100M training samples, addressing long-range interactions in large systems via an all-to-all node attention mechanism. The model avoids extra physics terms by integrating long-range physics into the attention, achieving improved LR accuracy and scalability.
HSMAD addresses graph anomaly detection under heterophily. It shows down-weighting heterophilic edges yields more concentrated spectral energy, aiding discriminative spectral representations. It also notes limitations of existing methods and proposes spectral-manifold learning aligned with heterophily.
We study adaptive recurrent message passing for test-time computing on graphs. We derive step dependence as a necessary and sufficient condition for convergence of adaptively computed recurrent processes in graphs, enabling efficient test-time adaptation of pre-trained graph backbones. The method offers a theoretical foundation for adaptive recurrent GNNs.
The paper analyzes brain graph learning and identifies spectral unreachability: standard coupled encoder-pooling oversmooths node representations, destroying high-frequency topological signals needed to delineate module boundaries. It argues for approaches that preserve or recover high-frequency information and proposes directions to overcome this limitation.
We formalize multi-agent graph generation as an energy minimization problem grounded in non-equilibrium thermodynamics and score-based diffusion. This framework guides the self-organization of agents into task-adaptive communication graphs, enabling efficient collaboration.
We introduce masked discrete diffusion for generative representation learning on hyper-relational knowledge graphs (HKGs). We address fact generation where multiple components of a fact may be missing, beyond simple link prediction. The method learns to generate valid hyper-relational facts by diffusing discrete components.
RL4RLA presents a general reinforcement learning framework to automate discovery of randomized linear algebra (RLA) algorithms, tackling sparse rewards and large search spaces with curriculum design and graph-based search. It demonstrates applicability across diverse RLA tasks.
Neural QAOA^2 is an end-to-end differentiable framework that jointly performs graph partitioning and initializes QAOA parameters for quantum combinatorial optimization on limited qubit devices. It aligns partitioning objectives with quantum goals and provides topology-aware parameter initialization to improve optimization.
SAGE-NAS marries an LLM-driven Semantic Agent with a Graph-Based Evaluator in a closed-loop framework to address semantic-physical misalignment in NAS. The Semantic Agent proposes architectures guided by semantic plausibility, while the Graph Evaluator enforces gradient-flow and training dynamics constraints.
We revisit the role of positive samples in graph contrastive learning. Surprisingly, competitive performance is achievable even without positives, prompting a reexamination of what drives representation learning in GCL and implications for designing future methods.
We propose a Graph Foundation Model for multimodal graphs with cross-modal alignment and modality-aware expert fusion, enabling universal pattern learning across graph modalities (structure, text, images) and robust generalization to open-world tasks.
T-GINEE introduces a tensor-based framework (CP tensor) for multi-graph representation learning with generalized estimating equations, modeling cross-network correlations explicitly and integrating task-specific loss. It captures inter-layer dependencies beyond simple aggregation.
Capacity-Agnostic Parameter Isolation for Continual Graph Learning (CAGNN) proposes a neuron-inspired architecture that constructs task-specific subnetworks using graph context and decouples them during training/inference, achieving efficient continual learning with increasing capacity.
BOCLOAK presents optimal transport-guided adversarial attacks on GNN-based bot detectors, examining realistic constraint scenarios and demonstrating vulnerabilities; it informs defenses for robust bot detection.
Weaver extends decoder-only LLMs with graph-aware reasoning by mapping graph distances to rotary positional embeddings, effectively weaving graph structure into token-level attention to enable multi-hop reasoning without encoders.
DeR-Mamba decouples multi-domain spatiotemporal entanglement in dynamic GNNs to mitigate representation drift and noise amplification, improving robustness of DGNNs. It provides a principled decoupling strategy for robustness.
We introduce Diffusion Value Function (DVF), a factored value function for graph-based multi-agent RL, assigning each agent its own value component for improved credit assignment and stable learning in GMDPs.
GraphFLEx is an unsupervised framework for learning structure in large expanding graphs, addressing scalability by restricting edge formation and avoiding full-relearns upon new nodes.
Backward oversmoothing investigates why deep GNNs are hard to train, analyzing optimization dynamics that cause or prevent oversmoothing, and offering insights into training strategies to mitigate it.
SynGR explores cross-modal synergy for generative recommendation, proposing a framework that leverages multimodal signals beyond alignment-centric fusion to capture emergent item properties in generation.
ReCoG introduces Relational and Compact Context Graph Learning for few-shot molecular property prediction. It addresses two core challenges—insufficient structural context modeling and redundant auxiliary context learning—by constructing relational, compact context graphs to enhance context exploration and information utilization for molecule representations. The goal is to improve predictive performance with limited labeled data.
This work studies decentralized multi-agent reinforcement learning for networked service systems with affinity in the presence of Byzantine nodes. Each agent learns an actor-critic policy over an unbounded traffic-state space and exchanges parameters with neighbors. The paper provides stability analysis and protection design to mitigate Byzantine threats while maintaining performance.
Hierarchical Multi-Scale Graph Neural Networks address heterophily by mitigating hub-dominated aggregation and oversmoothing/oversquashing. The approach introduces a hierarchical, multi-view spectral framework (HMH) with HAAR-inspired filters to improve learning on heterophilous graphs. This yields scalable, robust representations for complex graph structures.
E-mem proposes episodic context reconstruction for LLM agents to preserve narrative integrity across long horizons. Moving beyond memory preprocessing, the framework uses multi-agent collaboration to reconstruct episodic context, enabling System-2 like deliberation and improved reasoning over extended interactions.
CausalDPO extends Direct Preference Optimization by incorporating causal reasoning to reduce environmental confounders in generative recommendations. This causal augmentation mitigates spurious correlations, improving out-of-distribution generalization for LLM-based recommendation systems. The result is more robust alignment with user preferences.
Size Transferability of Graph Convolutional Networks across Sparsity introduces GWCN, a generalized graphon-based GCN. By stretching graphons to construct non-trivial limits, GWCN enables size transfer across graphs with arbitrary sparsity while preserving topological structure. This broadens theory and practice of scalable GCNs.
Boundary Embedding Shaping with Adaptive Contrastive Learning targets graph structural disentanglement near class boundaries. By focusing contrastive learning on boundary nodes, the method reduces structural noise from irrelevant neighbors and sharpens decision boundaries for more reliable node classification.
This work presents a lightweight, interpretable transformer by unrolling a mixed graph optimization algorithm for traffic forecasting. It uses an undirected spatial graph and a directed temporal graph to model correlations, predicting future signals under smoothness assumptions on both graphs. The model emphasizes transparency and efficiency.
JanusPipe enables efficient pipeline-parallel training for conservative machine learning interatomic potentials (MLIPs). It addresses the double-backward pattern inherent to conservative MLIPs, delivering scalable distributed training for accurate atomistic simulations. The design enables higher throughput and better scalability.
GI-GCN introduces Global Interacted Graph Convolutional Networks using Dominant Sets. At each layer, it leverages Dominant Set-derived solutions to model global node interactions and adaptively modulate node weights, enhancing graph classification performance. This enriches local aggregation with global context.
Adapting to Evolving Graphs proposes a scalable framework for dynamic graph coarsening that updates the coarsening map incrementally as graphs change. This enables efficient multi-scale representation learning on evolving networks without full recomputation.
Federated Graph Learning via Structure-Aware Fusion uses a Kalman framework with learnable dynamics to align structural information across clients. This structure-aware fusion reduces aggregation drift caused by graph heterogeneity while preserving privacy.
Information-Geometric Adaptive Sampling for Graph Diffusion reframes graph diffusion as a trajectory on a Riemannian manifold. By using the Fisher-Rao metric and the Drift Variation Score, it provides geometry-aware adaptive sampling to improve diffusion efficiency.
A Graphop Analysis of Graph Neural Networks on Sparse Graphs unifies generalization and universal approximation via a compact metric on the space of all graphs. The framework encompasses both sparse and dense regimes, offering broad theoretical insights.
Generalist Graph Anomaly Detection via Prototype-Based Distillation (ProMoS) is an unsupervised framework that models normality with prototypes to detect anomalies across unseen graphs. It enables robust, transferable anomaly detection without labeled data.
PFT (Phonon Fine-tuning) directly supervises second-order force constants to align MLIPs with DFT phonon properties. It scales to large supercells via stochastic Hessian-column sampling, improving vibrational property predictions while retaining efficiency.
HIAL proposes semantics-aware hypergraph active learning by modeling influence over high-order context rather than relying on clique expansions. It is a training-free framework that selects informative hyperedges for labeling via influence maximization.
TN-SHAP-G provides graph-structured tensor-network surrogates for efficient Shapley value and interaction estimation. By mirroring graph topology in a learned surrogate, it enables scalable attribution in complex graph models.
GLAD presents bidirectional structure-attribute alignment via latent graph diffusion models to robustly complete missing node attributes. The framework aims for stable training and multi-modal alignment beyond adversarial setups.
Principled Synthetic Data enables the first scaling laws for LLMs in recommendation by generating high-quality synthetic data with a curated pedagogical curriculum. This approach aims to provide predictable scaling behavior for LLM-powered recommender systems.
We propose enhancing RTL quality estimation by incorporating the control/data flow graph (CDFG) structure, in addition to token-based embeddings. The approach leverages explicit structural semantics alongside LLM-derived representations to predict area and delay without full synthesis. Experiments show improved accuracy and faster feedback compared with token-only methods.
GESC is a complex-valued graph network that adds gauge-consistent U(1) transport and projection-based self-interference cancellation to attention-based message passing. For each transported neighbor message, GESC removes the component parallel to the target before computing attention, and applies a gauge-invariant, sign-aware gate. The authors prove gauge-equivariance and demonstrate improved robustness on heterophilous graphs.
DICT models distributional uncertainty to improve trustworthiness in graph learning, by formulating a unified objective that captures perturbation-induced shifts in graph topology and node distributions. The framework seeks robustness to perturbations, fairness concerns, and reliability by explicitly accounting for distributional changes. Experiments show improved robustness and calibrated uncertainty estimates.
We address offline LLM-based recommendation with partial token-transition support by introducing Conservative Generative Flow Networks (GFlowNets). The paper formalizes why Sub-Trajectory Balance fails in offline settings, identifying three sources of non-identifiability and proposes remedies to constrain probability mass to supported regions. Empirical results show better diversity and calibration.
GAUSS uses graph-based structure and semantic signals to quantify uncertainty in long-form LLM generation for critical domains. It goes beyond paragraph-level confidence or fact extraction by modeling the relationships among facts and their narrative structure. This leads to better calibrated uncertainty estimates for long-form outputs in domains like clinical, legal, and policy drafting.
We propose EB-1WL and EB-GNN, edge-based message passing that passes messages along edges and triangles. The framework yields higher expressivity than node-centric 1-WL, and scales to larger graphs. Theoretical results show EB-1WL strictly surpasses 1-WL in distinguishing power, with experiments validating performance gains.
We develop a data-centric theory for graph foundation model transfer using a graphon-based continuous limit for dense graphs. The key result is an explicit decomposition of cross-domain output shift for Lipschitz backbones into components depending on domain properties and tokenization (set-based or message-passing). This guides when transfer is positive or negative and how to adapt.
DiP-G introduces discrete prompting for GNNs, aligning prompt optimization with the inference-time discretization of graphs. Traditional prompting uses continuous adjacencies during adaptation but discretization is applied at inference, causing train-test mismatch. DiP-G aims to minimize this mismatch and improve downstream performance.
We propose LGKDE to learn kernel density estimates for graphs by encoding each graph with a GNN into a discrete distribution and optimizing a maximum mean discrepancy-based objective to learn kernels. This improves graph-level anomaly detection by better capturing structural and semantic variations than fixed kernels.
CCLRec proposes consensus-driven contrastive learning to align GNN structural representations with LLM-derived semantic knowledge for graph recommendation. By jointly optimizing contrastive losses over structure and semantics, it reduces supervisory gaps and boosts recommendation accuracy.
SAND learns shape-aware molecular representations from 2D graphs for virtual screening. It uses a rank-preserving contrastive objective with differentiable Spearman correlation, yielding representations whose 2D similarity tracks 3D shape similarity, enabling efficient shape-based screening without 3D conformations.
An inference-time optimization framework for experiment-grounded protein ensemble generation. It optimizes latent representations to maximize ensemble log-likelihood to match experimental data, addressing fixed horizons and initialization sensitivity in diffusion-based approaches.
We present Graph Alignment via Dual-Pass Spectral Encoding and Latent Space Communication. The method tackles oversmoothing and latent space misalignment by using two spectral encoding passes and a communication mechanism between graph latent spaces, enabling reliable cross-graph node correspondence under noisy/heterogeneous structures.
Message Tuning Outshines Graph Prompt Tuning from a Prismatic Space perspective. The authors define Prismatic Space Theory to measure the adaptation capacity of graph prompts and show that message-tuning approaches can offer greater adaptability and interpretability than standard graph prompt tuning in GFMs.
VDW-GNNs introduce vector diffusion wavelets for geometric GNNs. The new wavelets capture vector-valued signal structure on graphs, improving performance on synthetic point clouds and real datasets like wind fields and neural activity. Theoretical results show desirable frame properties and stability.
View Space formalizes a new representation axis induced by graph structure, enabling inductive learning across graphs with heterogeneous features. Graph View Transformation maps graphs into a common view space, supporting unified representations across datasets without retraining.
We develop a Cartesian-3j framework for machine learning interatomic potentials, introducing the Cartesian-3j symbol and Cartesian Clebsch-Gordan coefficients as direct Cartesian analogues of Wigner-3j symbols, enabling a Cartesian tensor-based equivariant formulation for MLIPs.
Rotary Position Encodings for Graphs. We apply rotary position encodings to graphs via Wave-Induced Rotary Encodings (WIRE), rotating attention tokens according to the graph Laplacian spectrum to inject structural information. The method recovers standard RoPE on grids and offers desirable theoretical properties.
Sparse Topology-Aware Pairwise Scoring for Large-Scale Multi-Agent Reinforcement Learning. We propose SOPS to scale MARL communication by decoupling decision-making from global topology, using sparse, topology-aware pairwise scores to guide information sharing in large agent networks.
MedMamba combines multi-scale convolutional embeddings, adaptive graph learning, and state-space modeling to capture local-global dynamics and nonstationarities in medical time series, along with modeling latent channel interactions for improved classification.
We propose CL-GCL, a Comprehensive and Lightweight Graph Contrastive Learning framework. It addresses semantic distortion from random augmentations and class entanglement from one-to-one sampling by leveraging graph coarsening to preserve community-level semantics and manifold learning to capture local geometry.
We present FlashSchNet, an IO-aware SchNet-style GNN-MD framework designed to optimize GPU memory traffic between high-bandwidth memory and on-chip SRAM. By building around four IO-aware techniques, it delivers faster, more accurate molecular dynamics while retaining accurate many-body interactions.
We introduce ERAlign, an energy-based representation alignment framework for aligning GNN and LLM representations on text-attributed graphs. Using energy-based constraints reduces representation drift and improves generalization across TAG tasks.
We propose MIMO-LP, a Multi-Input Multi-Output framework to accelerate subgraph-based link prediction. It processes a batch of query node pairs with shared subgraph computations and multiple outputs to speed up inference without sacrificing accuracy.
We propose BSCT, the Bond Smoothness Characterization Test, an efficient benchmark that probes the smoothness of the potential energy surface by controlled bond deformations. BSCT helps detect non-smooth regions and guides ML interatomic potential architectures toward physically realistic behavior.
We introduce FedGCM, a federated graph learning framework with Group-oriented Conflict Mitigation. It uses tailored gradient surgery to align inconsistent optimization objectives across subgraphs, addressing homophily heterogeneity and improving convergence and generalization.
We present ExCyTIn-Bench, the first benchmark to evaluate an LLM agent on cyber threat investigation tasks derived from investigation graphs. It uses a controlled Azure tenant with a SQL environment and 57 security log tables to simulate real-world threat investigations.
We propose M^2GenCO, a multi-task learning framework that treats problem types as tasks for meta-learning and uses diffusion-based generative solving for graph-based combinatorial optimization problems. This enables multi-task pretraining and rapid adaptation to new COPs.
We propose Graph-R1, the first agentic Graph-RAG framework trained end-to-end with reinforcement learning. It features lightweight knowledge hypergraph construction and learned retrieval decisions, reducing cost and enabling better integration with long-context reasoning.
We address certified robustness of GNNs against poisoning with a novel semi-supervised learning approach. Our method provides poisoning certificates in graph settings where traditional partition-and-aggregate schemes fail due to label and structure sparsity.
We introduce CalPro, a prior-aware evidential conformal prediction framework for protein structures. It outputs Normal-Inverse-Gamma distributions via a graph neural network, includes a differentiable calibration surrogate, and uses split-conformal evaluation for shift-robust uncertainty quantification.
We propose GraphPFN, a prior-data fitted graph foundation model that pretrains on synthetic graph data to support in-context learning. It adapts the PFN paradigm to graphs, similar to TabPFN and G2T-FM.
We analyze oversmoothing in deep GNNs through bifurcation theory, characterizing it as convergence to a homogeneous fixed point. We show that replacing monotone activations with non-monotone/topological priors can break this stability and improve expressivity.
We present HGMem, a hypergraph-based working memory to improve multi-step RAG for long-context complex relational modeling. It consolidates high-order correlations among primitive facts to support coherent multi-step reasoning in long-context tasks.
We extend the Strong Expressive Lottery Ticket Hypothesis to relational and temporal GNNs, showing that sparse subnetworks can preserve relational WL expressivity. The results provide probabilistic guarantees and a lower bound on the expressivity preservation.
We prove exact learnability results for executing graph algorithms with GNNs under bounded-degree and finite-precision constraints. The approach first trains an ensemble of MLPs to execute local instructions, then uses the trained ensemble as the GNN update function at inference. NTK theory provides generalization insights.
We propose QHNetV2, SO(3)-equivariant Hamiltonian predictors leveraging SO(2) local frames to achieve global SO(3) equivariance without Clebsch–Gordan tensors. The network uses SO(2)-equivariant operations to update off-diagonal and other Hamiltonian components efficiently.
We introduce GFedCL, a graph-based federated continual learning framework that explicitly incorporates spatial and temporal information across clients. This improves adaptation to evolving data distributions, such as regionally varying COVID-19 prevalence.
We propose FedRGL, robust federated graph learning under label noise. Specifically, FedRGL leverages globally aggregated information to mitigate the impact of noisy labels on the global model.
We present TACTIC, a CTDE framework with three components: (i) VQ-VAE-based trajectory abstraction that learns discrete task-semantic classes; (ii) semantic-conditioned sparse coordination graphs that prune edges according to variance-based payoff sensitivity; and (iii) a pretrained, frozen trajectory-class pre
We reveal a new attack surface in multimodal LLM-based recommender systems: synchronized cross-modal poisoning can steer fused representations during fine-tuning, even when cross-modal consensus mitigates single-modality attacks. We formalize cross-modal interactive poisoning and propose VENOMREC, a framework that performs Exposure Alignment to enable targeted promotion via multimodal signals. The work highlights threat characterization and implications for defense and robustness in multimodal recommender systems.
We propose R-GFM, a Graph Foundation Model built on a Riemannian Graph-of-Graphs, where structural scale is treated as a first-class modeling primitive. Unlike fixed-hop subgraph sampling that fixes a receptive field, R-GFM constructs a multi-scale GoG to better capture diverse structural contexts across tasks. This scalable design addresses scale mismatch and enables scale-aware pretraining for varied graph domains.
KBQA-R1 presents a reinforcement framework to strengthen LLMs for knowledge base question answering by addressing hallucinated queries and rigid template-like reasoning. It shifts from purely text imitation to a verification-driven paradigm that aligns generated logical forms with knowledge graph schemas. The approach integrates schema-aware checks and executable form generation to improve KBQA reliability.
We study cooperative multi-agent reinforcement learning on graphs, where each agent's local state and action update depend only on its 1-hop neighborhood. The global state-action space grows exponentially with the number of agents, making globally optimal policies intractable to compute. We propose scalable local policies that exploit the graph structure to coordinate effectively without enumerating the full joint space.
We introduce order-equivariant neural networks (OENN) that unify graph neural networks and sheaf neural networks through a face-poset viewpoint. We characterize all linear order-equivariant maps, build OENN layers, and prove a universal approximation theorem for continuous order-equivariant maps, extending the UAT beyond standard GNNs. The framework is demonstrated on graph and sheaf models, showing how to extend UAT to this broader class of architectures.
FRIGID proposes a diffusion-based molecular generation framework conditioned on mass spectra, addressing scalability bottlenecks in both training and inference. It uses intermediate fingerprint representations and fixed chemical formula constraints to guide the generation process. The approach enables scalable, MS-conditioned de novo structure elucidation for high-throughput molecular discovery.