Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 32 papers for 2026-08-28
Nothing cleared the bar today. Papers read in full today: 10.
The load-bearing claim is that equivariance of the architecture is insufficient: the training coupling must also select aligned orbit representatives, and GW-based node alignment removes artificial edge/node ambiguity along flow-matching interpolations. The paper tests that claim unusually directly by holding the graph-transformer backbone fixed across Random, FLB, GW, and outer-OT variants: on the controlled SBM at five Euler steps, GW+outer lowers FGW-NNA from 0.796 to 0.568, and on ZINC250k inner GW improves five-step validity from 0.564 to 0.643 while reducing FCD from 18.45 to 15.02. The quotient-lift and diagonal-symmetrization results explain why the practical aligned-representative procedure is compatible with equivariant vector fields, rather than presenting GW as an unexplained preprocessing heuristic.
The strongest gains are specifically a low-integration-budget result on small/standard graph and molecular benchmarks, while the expensive nonconvex GW-to-Hungarian approximation is not shown to approximate Gromov-Monge reliably and the 500-step SOTA comparison confounds alignment with additional architectural changes.
The load-bearing diagnosis is specific and useful: token-wise LayerNorm/RMSNorm destroys vector magnitude, while ordinary dot-product attention preferentially rewards high-norm keys; graph multiset/cardinality information can reside precisely in that magnitude. AdaRMSN, simplified L2 attention, and a stronger RRWP stem jointly turn this diagnosis into competitive results: PPGT reaches 234/400 on BREC versus GRIT's 218, improves ZINC from GRIT's 0.059 to 0.0566 MAE, and is statistically tied with GRIT on Peptides-Function while improving some LRGB tasks. The ablation is directionally supportive, though it validates the package mostly on ZINC rather than isolating the mechanism across settings.
The headline is stronger than the evidence: PPGT remains a graph-specific RRWP/RPE Transformer with custom attention masking and a multi-part PE stem, while its largest practical gains are small, PCQM4Mv2 uses one seed, and the O(N^2) model needs subgraph sampling for truly large graphs.
GRAIN tackles brittleness of LLM based real world graph reasoning when node identifiers and task phrasing shift. It replaces multi agent latency with a single agent reinforcement learning framework that learns invariance to naming and narrative changes while preserving topological reasoning. The approach aims to bridge name and narrative shifts to improve robustness without relying on expensive coordination.
Decentralized Multitask Learning over Learned Task Graphs presents a two-phase approach: (i) estimate a generalized task graph by observing distributed SGD iterates and a generalized Laplacian, (ii) exploit the learned graph to diffusion-cooperate across tasks. It handles unknown task relationships in a decentralized setting.
Inductive Correlation Clustering with Graph Neural Networks offers an inductive GNN-based approach to correlation clustering, addressing scalability and inductive limitations of existing CC methods by learning cluster structures without predefining k.
Leakage-Free Evaluation and Distribution-Robust Spatio-Temporal Graph Learning for Inductive Kriging proposes a leakage-free 3x3 spatio-temporal partition to ensure honest evaluation in inductive kriging, enabling robust OOD assessment.
Cone Extended Rayleigh Quotients for Directed Graph Learning proposes a learning-oriented framework using cone-extended Rayleigh quotients for nonsymmetric propagation operators. It provides spectral certificates, sensitivity analysis, and adaptive control without symmetry or nonnegativity assumptions, giving computable bounds on eigen-structure.
Rethinking Message Passing as Retrieval reframes GNNs as retrieval-augmented models where each layer applies an MLP to node features along with a permutation-invariant summary of retrieved graph context. This view explains why neighborhood information helps and offers potential efficiency benefits.
ProRetrieval: Learning to Orchestrate Hybrid Search via Executable Program Synthesis
Relational Over-Regularization: Graph-Based AI-Generated Text Detection via Sentence Transition Deviation identifies a sentence-pair level structural signal where LLM-generated text shows inflated inter-sentence transition variance; formalizes as Relational Over-Regularization for detection.
Chemical reactions are transformations in electron space not just graph edits. MAELLE introduces discrete flow matching on electron occupation vectors to map reactants to products, capturing mechanistic electron rearrangements rather than topology edits. This yields a principled flow based framework for mechanistic reaction prediction.
The authors introduce Subgraph Filtering for Fair Graph Neural Networks (SF-GNN), a lightweight method that controls how biased signals propagate during aggregation by filtering subgraphs. By explicitly shaping local structural pathways, the approach improves fairness without requiring global constraints on representations or predictions.
PACIFIER: Pacing Opinion Depolarization via a Unified Graph Learning Framework studies moderation in FJ model-based polarization; frames the problem as autoregressive sequential planning and proposes a unified graph-learning method to pace and depolarize opinions.
Beyond Linearization: Attributed Table Graphs for Table Reasoning
LitEm is a neural regression model that adds embeddings to enable transductive KG embedding methods to predict numerical attributes in knowledge graphs. It demonstrates improved accuracy on numerical attribute prediction in KGs.
This paper proposes SLM-Conditioned Hierarchical Relation Routing, which integrates a small language model into graph message passing for labeled property graphs. The model learns to condition which semantic evidence influences propagation, enabling dynamic, target-specific routing of relations in a hierarchical GNN.
Graph-Based Modeling of Financial Volatility Dynamics introduces FA-GSTN, a Finance-Aware Graph Spatio-Temporal Network that treats realized volatility forecasting as modeling the evolution of a structured financial object via a spatio-temporal graph built from market data including implied volatility surfaces.
Refine-POI: Reinforcement Fine-Tuned Large Language Models for Next Point-of-Interest Recommendation describes reinforcement-fine-tuning techniques to improve POI recommendation, preserving semantic continuity in topology and offering diverse outputs beyond top-1.
SymbolLKG: Towards Verifiable Logical Reasoning via Logical Knowledge Graph and Symbolic Solvers proposes a neuro-symbolic architecture integrating a Logical Knowledge Graph with symbolic solvers to enable verifiable multi-step logical reasoning in LLMs.
C-Unseen: Weak Signal Detection in Dynamic Temporal Knowledge Graphs via LLM Reasoning defines weak signals as rare, semantically coherent subgraphs that proliferate over time in dynamic KG; uses LLM reasoning to detect them, enabling early warning.
CPGRec+: A Balance-oriented Framework for Personalized Video Game Recommendations
Let Them Steal: Trapping Large Language Model Extraction Attacks with Knowledge Honeypot
REPREC: Representation Driven Parameter-Efficient Recommendation System
Empowering Compact LLMs with Fusion of Layer-wise Exits for Recommendation
Graph-Based Pseudo-multimodal Contrastive Learning for 12-Lead ECG Representations builds a graph-structured representation across ECG leads and uses pseudo-multimodal contrastive learning to capture inter-lead dependencies and global waveform patterns, improving representations for cardiovascular diagnosis.
BekchiAI: Measuring, Observing, and Controlling LLM Agents in One Click introduces a benchmark and platform for measuring agentic skills of LLM agents and enabling live monitoring and control of tool-using agents.
pro-team at LLMs4OL 2026 Tasks Flagship and Reuse: Retrieval-Augmented Generation and Vocabulary-Constrained Filtering for Ontology Learning describes an offline RAG pipeline with demonstration retrieval and vocabulary filtering for ontology learning tasks, using Qwen2.5-14B-Instruct and all-MiniLM-L6-v2 for demonstration retrieval.
DRL: A Deterministic Relational Middleware Layer for Transaction-Safe Enterprise NL2SQL Under Schema-Graph Scaling
MedFG-VQA: Low-Frequency Memory and Graph Attention for Lightweight Medical VQA
Scaling Graph Neural Networks for Friend Recommendation details a production-ready GNN ranking system that uses multi-hash user embeddings and temporal neighbor sampling to scale to large social graphs while maintaining recommendation quality.
From SQL to Knowledge Graphs: An LLM Driven Multi-Agent Approach with Data Schema Improvement
Conversational Recommendation over Live E-Commerce Catalogues with Self-Refreshing Retrieval