Showing 13 papers for 2026-04-03
CRIT proposes a graph-based automatic data synthesis framework to enhance cross-modal multi-hop reasoning. By generating challenging data that requires combining textual and visual cues across multiple hops, CRIT addresses weak multi-hop reasoning and reduces hallucinations in vision-language models.
VIRSO is a graph-based neural operator for sparse-to-dense, real-time virtual sensing on irregular grids. It enables edge deployability and portability by reconstructing dense fields from sparse measurements with low latency in resource-constrained environments.
LEO is a spatio-temporal Graph Attention Network that fuses multi-modal sensor tracks for extended object fusion and tracking in autonomous driving. It combines the robustness of classical Bayesian methods with the adaptability of learned fusion by training on production-grade sensor data.
We show that the diffusion denoising objective, via denoising score matching, smooths gradients and enables faster, more stable causal structure learning on high-dimensional data. We also propose an adaptive k-hop mechanism to refine the learned DAGs.
What do temporal graph learning models learn? We examine reliability of benchmarks and the signals models rely on, showing that common datasets may encourage simple heuristics rather than true temporal dynamics. We propose diagnostics to uncover what structural and temporal cues drive predictions.
From Sublinear Graph Algorithms to LLM Test-Time Methods demonstrates how pre-existing knowledge affects test-time augmentation like RAG and tool use. It formalizes multi-step reasoning as an s-t connectivity problem and shows that greater pretraining knowledge reduces the amount of augmentation required, linking theory to practical LLM reasoning.
BRIDGE presents an end-to-end architecture that couples a sequence model with a graph module to jointly learn from sequential and relational data. It addresses tasks where entities generate event sequences and simultaneously relate to each other, achieving unified learning that outperforms modality-specific baselines.
We propose Cross-attentive Cohesive Subgraph Embedding to mitigate oversquashing in GNNs. By enriching node embeddings with cross-attentive information from cohesive subgraphs, the framework preserves long-range dependencies and improves performance in dense and heterophilic graphs.
Semantic Refinement with LLMs for Graph Representations argues that graph learning should adapt to structure–semantics heterogeneity across domains. It proposes using large language models to refine semantic information in graphs, enabling domain-aware representations and improved generalization.
We present a frequency-aware epileptic seizure detection framework using ictal-phase EEG analysis and Graph Convolutional Networks. By analyzing signals across separate frequency bands with a GCN, the approach improves detection performance and aligns with neurophysiological relevance.
We examine trust and reliance on AI in education, focusing on AI literacy and need for cognition as moderators of appropriate reliance during programming problem solving. Results show that trust shapes students' use of AI outputs and that learner traits modulate this relationship.
Lumos introduces a principled framework for specifying and certifying Language Model System (LMS) behaviors. It provides an imperative probabilistic programming DSL over graphs to generateIID prompts by subgraph sampling, and supports certification of arbitrary prompt distributions via integration with statistical certifiers, offering a hybrid of operational and formal guarantees.
Multi-Agent Video Recommenders surveys the evolution, patterns, and open challenges in agent-based video recommendation systems. It discusses architectures and coordination strategies where specialized agents handle video understanding, user modeling, and ranking, and outlines remaining challenges and future directions.