Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 11 papers for 2026-02-25
The paper introduces the Very Big Video Reasoning (VBVR) Dataset, an unprecedentedly large dataset designed to study how video models scale in reasoning beyond perceptual quality. It emphasizes grounding intelligence in spatiotemporal environments to reason about continuity, interactions, and causality, and provides benchmarks to analyze scaling behavior as data grows.
SkillOrchestra introduces a framework for skill-aware orchestration among multiple agents. Instead of end-to-end routing, it leverages skill transfer to route tasks to capable tools, addressing coarse input-level decisions and the instability of RL-trained orchestrators in multi-turn settings.
This work surveys Vision-Language-Action (VLA) models and argues for a unified design space and evaluation. By addressing inconsistencies across models, it reexamines VLA design under a common framework and, starting from a simple VLA baseline, distills recipes for building stronger VLA systems.
This paper reports an exploratory red-teaming study of autonomous language-model powered agents operating in a live laboratory with persistent memory and tools. Over two weeks, twenty researchers interacted with the agents under benign and adversarial conditions, documenting eleven representative failure cases from LM autonomy, tool use, and multi-party communication, including unauthorized compliance and disclosure of sensitive information.
ManCAR proposes Manifold-Constrained Adaptive Reasoning for sequential recommendation. It argues that latent multi-step reasoning drifts when driven by target-focused objectives and reframes reasoning as navigation on a collaborative manifold with adaptive test-time computation, aiming to produce more plausible, efficient reasoning trajectories.
We present a defense against model extraction attacks on GNNs by introducing a decision boundary-aware signature (CITED). The signature makes it harder for attackers to approximate the decision boundary when querying the service. Experiments on MLaaS-style setups demonstrate improved robustness of GNNs under various extraction strategies.
The work benchmarks GNN models on molecular regression tasks, using SMILES-to-graph representations and fixed-size fingerprints. They employ CKA-based representation analysis to compare how different GNNs encode molecular information, linking representation similarity to predictive performance.
We investigate how graph topology shapes GNN activation patterns. By analyzing curvature and using Massive Activations in Graph Transformers, they reveal how topological bottlenecks influence which edges get activated and how this interacts with learned preferences.
PhyGHT introduces a physics-guided hypergraph transformer for signal purification at the HL-LHC, addressing extreme pileup noise. The method models physics priors within a hypergraph framework to separate rare signal components from overwhelming backgrounds, improving reconstruction observables.
SymGraph is a symbolic framework for expressive and interpretable graph learning, aiming to transcend 1-WL expressivity and improve interpretability beyond standard message-passing backbones. By integrating symbolic reasoning, it seeks to enhance both accuracy and transparency in graph models.
The paper develops a spectral framework for graph neural operators using graphons as limits of graph sequences, proving convergence of neural operators under various graphon regularity assumptions and discussing transferability.
MAGNET proposes Modality-Guided Mixture of Graph Experts with Entropy-Triggered Routing for multimodal recommendation. It uses modality-specific experts and routing decisions based on information entropy to handle heterogeneous modalities and reduce entanglement, improving performance in sparse and long-tail scenarios.
HELP advances GraphRAG by HyperNode Expansion and Logical Path-Guided Evidence Localization, addressing hallucinations and semantic noise in LLM-augmented retrieval. It expands graph nodes and uses logical paths to localize trustworthy evidence, balancing accuracy and efficiency.
The Initial Exploration Problem in Knowledge Graph Exploration identifies the challenge novices face when navigating unfamiliar KGs: not knowing what questions are possible, how data is structured, or how to start. The paper theorizes IEP and discusses strategies to reduce it.
E-MMKGR presents a unified multimodal knowledge graph framework for e-commerce, constructing an e-commerce-specific Multimodal KG (E-MMKG) and learning unified item representations via GNN propagation and KG-oriented optimization, enabling cross-task applicability.
The paper proposes Probability-Invariant Random Walk Learning on gyral folding-based cortical similarity networks to diagnose Alzheimer's and Lewy body dementia. It addresses inter-individual variability in cortical folding that leads to irregular graphs, aiming for robust classification.