Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 25 papers for 2026-09-23
Nothing cleared the bar today. Papers read in full today: 7.
The load-bearing contribution is a diagnostic one: under a same-population, fixed-entity evaluation, isolated beneficial owners obtain exactly 0.5 AUC, and several common negative-sampling or endpoint-varying choices can report spurious gains. The full text supports that narrow isolated-node claim with a symmetry proof and large-scale stratified measurements, including replication across leak sources; however, the broader WL-class story is empirical rather than a proved recovery bound, and the stronger-model evidence is only a SIGN/SGC-style model plus hand-engineered SEAL-like features, not an actual trained subgraph GNN.
The theorem constrains automorphism orbits, not WL colour classes, and a global WL class neither equals a stabilizer orbit in general nor bounds every isomorphism-invariant structural rule; thus the paper overstates its general distinctiveness ceiling beyond the exact isolated-node case.
Realtime-Venus introduces a proactive full-duplex interaction system built on two separately trained 9B models: Realtime-Venus-Omni for audio-visual interaction and Realtime-Venus-Audio for spoken dialogue. Each model acts as a complete conversational frontend that tightly integrates continuous perception, dialogue management, and natural speech generation, enabling asynchronous delegation between perception and response. The system targets natural, timely interaction in digital-physical contexts.
Regularized Recursive Self-Improvement (RRSI) proposes automating how LLM agents refine their own harnesses—prompts, control flow, tooling, memory, and context management—via iterative edits. While this RSI can boost performance on in-distribution tasks, it risks memorizing training tasks and degrading out-of-distribution generalization. The paper introduces a regularization mechanism to curb overfitting and demonstrates improved robustness across benchmarks.
OmniEdu is an open family of educational foundation models for K-12 learning and teaching. Its instruction-tuning corpus aggregates over 100 educational resources and general instruction sources and focuses on four capabilities: subject competence, curriculum grounding, diagnosis, and instructional support. The models are designed to understand curricula and tailor guidance to learners.
WorldCrafter introduces a video world model that maintains a camera-queryable implicit 3D-aware memory to support long-horizon, cross-view consistency. The key idea is to let the requested view-point shape how memory is compressed into the video generator's limited token budget, and to train a memory encoder and a pose-conditioned readout alongside the video generator. This yields more reliable, view-consistent video understanding and interactive generation.
GameHorizon Suite provides a unified data and evaluation framework for gameplay across multiple temporal horizons and diverse model families. It addresses limitations of existing benchmarks by covering multiple games, including language instructions and lower-variance evaluation, and comprises three components centered on data collection, tasks, and standardized evaluation.
Graph Domain Adaptation Does Not End with Representation Learning argues that adapting representations alone may not exhaust predictive evidence in the target domain, since there is more than one graph-propagation path. It proposes and analyzes additional components and paths to exploit in domain adaptation. The paper provides theoretical and empirical insights into richer graph-aware adaptation.
Structured multi-agent workflows exchange intermediate messages whose content can reveal private state even when the final output is safe. The authors identify selection-channel leakage: after authorization fixes, a private-state-aware choice among semantically valid realizations creates an additional inference channel. They propose the selection-invariant communication compiler (SICC), which constrains the post-authorization representation kernel rather than prescribing fixed templates.
TailSpec-EASE is a knowledge-graph-regularized linear recommender for Web long-tail discovery. It integrates a knowledge graph into a lightweight linear model (EASE-style) to improve recommendations for rare items, while keeping global closed-form solvability. The approach offers efficient, scalable long-tail discovery.
We present the ABAI submission to COLIEE 2026 Task 1 for case law retrieval, accompanied by a controlled study of why it underperformed. The task suppresses the cited passages themselves, which removes much lexical overlap a retriever would rely on. Our four-stage pipeline combines multi-view BM25 over citation-context windows with reciprocal rank fusion, neural re-ranking, graph-based features from entity communities and a graph attention network, and a LightGBM meta-learner over 34 features.
Hybrid Quantum-Classical GNN for Banking IT Root Cause Analysis (HQ-RCA) uses a DynEdge-based classical backbone with a variational quantum circuit head. On 13 months of anonymized bank IT data, HQ-RCA matches the standalone DynEdge baseline, demonstrating a competitive quantum-classical approach. The work suggests potential practical value of quantum components in industry-scale RCA.
Empirical Auditing of Edge-Private Graph Generators provides a framework to measure privacy leakage by testing distinguishability of outputs from neighboring inputs. It compares direct-edge, local-structural, and GNN-based attacks across two generators and networks. Results reveal that leakage is mechanism- and network-dependent, with learned representations revealing more information than local statistics.
Towards Hierarchical GNNs for multi-grid power flow studies generalization across operating scenarios by enabling hierarchical latent communication between two reduced graphs in a GENCO-based corrective network. The paper compares Kron-based transport, same-anchor quotient, and a flat backbone, across 200 new scenarios per grid. Results show improvements in generalization.
MAGIC introduces Mixed-Granularity Agent Graphs via Incremental Construction with Dense-Reward Reinforcement Learning to tailor collaboration topology for tasks. Granularity is chosen locally per functional role, combining fine- and coarse-grained agent groups while building graphs incrementally under dense rewards. This improves coordination and efficiency in multi-agent systems.
EMERGE provides resolution-agnostic point cloud generation using equivariant graph-based diffusion. The method leverages graph-equivariance to handle varying resolutions and preserve geometric structure during diffusion. This advances 3D generation for non-grid topologies.
The work presents a size-agnostic deep reinforcement learning framework for 1D bin packing on item-compatibility graphs. It is end-to-end and does not rely on size-specific encodings, enabling generalization across different instance sizes. Experiments show competitive performance against standard baselines.
DefaultGNN uses a dual-perspective GNN over buyer-seller transaction networks to predict corporate defaults, addressing sparse financial statements by leveraging transaction data. Using six years of electronic tax-invoice data, the approach demonstrates that transaction networks reveal risk propagation channels beyond what financial statements show. This improves default prediction accuracy and robustness.
SambaGraph introduces a dataset and benchmark for soccer tactical response modeling using action-reaction spatio-temporal graphs. From 2022 FIFA World Cup data, it constructs 4,070 action-centered episodes as 23-node graphs with attack/defense views, response labels, and evaluation splits. This enables developing and evaluating models that reason about how actions elicit defensive or offensive responses.
CacheDyG decouples temporal propagation from optimization in dynamic graph learning to reduce training and parameter costs. It stores and reuses historical propagation results, avoiding repeated re-computation for largely unchanged structures. Experiments show substantial speedups with minimal or no loss in accuracy.
Towards Adaptive Federated Graph Clustering proposes a global community-aware contrastive learning framework for federated graph clustering, addressing client subgraph heterogeneity. By aligning communities across clients with contrastive objectives, it improves clustering in unlabeled, distributed graphs. The method enhances adaptability to diverse client data while preserving privacy.
DISCO (Diffusion-Induced Spatial Attention) addresses overlapping community detection by combining a diffusion-based structural prior with spatial attention. It preserves long-range dependencies and sharpens community boundaries by leveraging influence spreading dynamics. Experiments show improved detection of overlapping communities.
FeatLens introduces feature-guided dynamic code graph construction and retrieval for repository-level code generation. To generate a target function, an LLM must identify reusable dependencies such as existing functions, APIs, and cross-file definitions, which existing retrieval methods struggle to provide efficiently. FeatLens uses feature-oriented methods to build a dynamic repository graph that captures these dependencies, enabling cost-efficient retrieval of relevant code components.
This paper proposes Graph Edge Sparsification (GES), a learning-based method to sparsify Euclidean TSP instances by exploiting geometric structure and combinatorial optimization to prune edges, accelerating exact solutions. It also introduces a Dual-GNN multilevel coarsening framework to preserve important structure while reducing problem size. Together, GES and multilevel coarsening aim to boost large-scale TSP solver efficiency.
TopoSIGN is a topology-guided pre-training and prompt learning framework for signed graphs. It uses a structural encoder based on the magnetic signed Laplacian to capture both positive and negative relations and supports transfer across tasks. This enables better graph representations for signed-graph downstream tasks.
PreGS is a parameter-transfer-based multi-expert GNN framework for node classification. It pretrains multiple structural branches (experts) and transfers parameters among them to encourage diversified yet stable representations. This reduces training overhead while maintaining or improving accuracy.
ArticleMiner constructs knowledge graphs from scientific publications guided by ontologies. It tackles the challenge that tables alone do not expose the full scientific facts, by aligning headers, captions, units, and methods to ontology terms. The system supports publication-level extraction beyond table recovery.
Despite progress, LLM-based and knowledge-graph-augmented radiology report generation (RRG) methods still suffer from defects. LLM-only models lack structured medical prior knowledge, leading to hallucinations and limited interpretability, while static one-round knowledge fusion uses single-source knowledge. MAC-RRG proposes iterative multi-agent collaboration that dynamically updates knowledge according to generation feedback, reducing hallucinations and improving interpretability.
We introduce MT-FGNN with Sample Relationship Learning to enhance Remaining Useful Life (RUL) prediction. The model leverages multi-term Fourier transforms over graphs to capture complex spatio-temporal dependencies. It reduces reliance on domain-specific pre-defined graphs by learning relationships directly from data, improving predictive performance.
An Accurate and Interpretable Hyper Graph Neural Network for GBM Survival Prediction argues for a hypergraph-based model that is both accurate and interpretable for glioblastoma patient survival. The model captures multi-modal imaging and clinical data as a hypergraph, enabling explainable reasoning about prognostic factors. This challenges the trade-off between accuracy and interpretability in medical GNNs.
TCMaster provides a confidence-aware querying and workload-guided physical design framework for multi-source Traditional Chinese Medicine knowledge graphs. It annotates edges with provenance-level confidence and rewrites Cypher queries to incorporate confidence predicates, enabling reliability-aware data access. The system integrates pharmacopoeias, prescriptions, molecular databases, and LLM-extracted micro-semantics at scale.