Showing 14 papers for 2026-07-08
Proposes a parameter-efficient Tri-Branch Modular Fusion Neural Network that integrates three modalities for molecular property prediction: 3D geometry via SchNet, SMILES embeddings via ChemBERTa, and explicit macroscopic descriptors via the Deep & Cross Network. The model aims to better capture long-range topological dependencies and physicochemical heuristics by fusing orthogonal information sources, moving beyond single-modality representations.
Introduces Canopy, a heterogeneous graph foundation model that unifies ten public and proprietary data sources into a single knowledge graph for metabolic engineering. By leveraging relational structure in the KG, it addresses limitations of constraint-based models and tabular ML that ignore biological relationships, enabling data-driven design and prediction of microbial production.
EntroPath introduces a manifold learning method that recovers geodesic geometry from data graphs using ensembles of diffusion paths based on maximum-entropy random walks. This approach avoids biases of locally normalized diffusion and shortest-path distances by leveraging the full path ensemble to compute robust dissimilarities.
Graph Convolutional Attention analyzes graph denoising and diffusion from a spectral perspective. It shows that under a denoising objective, linear attention is suboptimal and tends to learn only an average spectral denoising filter, highlighting limitations of standard attention for graph denoising and motivating alternative designs.
Develops a fast binary-splitting non-adaptive learning method for Erdős–Rényi graphs G(n,q) using group queries on node subsets. The approach uses binary splitting to efficiently identify edges with non-adaptive queries, achieving improved query complexity relative to general non-adaptive learning bounds for sparse graphs.
TACTIC-KG proposes using small agent teams to build cybersecurity knowledge graphs from heterogeneous, noisy CTI reports, avoiding end-to-end reliance on large language models. It emphasizes cost control, controllability, and stability while coordinating multiple agents to extract and integrate adversarial entities, actions, and relations.
Proposes integrating knowledge graphs with multilingual scholarly corpora to adapt LLMs for SSH research workflows. It addresses multilingual access and evaluation challenges, presenting an ongoing use case within the LLMs4EU project and ALT-EDIC infrastructure to tailor foundation models to SSH domains.
Danus introduces an orchestration system for mathematical reasoning using a shared fact graph as global memory. It coordinates multiple reasoning agents, enabling parallel proof search while organizing intermediate claims to improve reliability.
VisTCP provides a visualization framework to construct knowledge-graph-based representations for Traditional Chinese Painting, addressing semantic gaps and domain-specific objects and events. It supports archaeology and art history research by enabling structured semantic understanding of TCPs.
i-EXAM is a planning-powered tool that helps security teams model networks and perform what-if analyses to identify attack paths, evaluate security metrics, and generate diverse hardening strategies, with explanations in natural language via large language models.
LongCrafter presents a structured synthesis framework for creating diverse long-context supervision data using an evidence-grounded approach and a hierarchical task taxonomy, addressing limitations in task coverage, difficulty, and faithfulness in long-context SFT.
RSF-GLLM decouples differentiable graph reasoning from answer generation to bridge semantic gaps in multi-hop KG QA. It introduces Recurrent Soft-Flow to propagate continuous relevance scores with a GRU-guided updater and dynamic gating, enabling end-to-end trainable graph reasoning.
Trust-Aware Citation Cartel Ranking in Scholarly Knowledge Graphs combines graph structure with semantic citation intent to rank potentially cartel-like papers, enabling detection of suspicious density patterns while accounting for legitimate dense communities.
GORIO introduces GPU-centered remote I/O for graph ANNS over NVMe-oF, shifting data movement and computation toward GPUs to alleviate CPU bottlenecks. It enables efficient remote graph ANN on large vector indexes by exploiting NVMe-oF storage with GPU-driven I/O.