Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 17 papers for 2026-02-27
HyTRec introduces a Hybrid Temporal-Aware Attention architecture for long behavior sequential recommendation. It explicitly decouples long-term stable preferences from short-term intent spikes by using a dual-branch attention design that routes massive historical sequences through a linear attention path while reserving a separate path for short-term cues, balancing efficiency and retrieval precision. The approach aims to improve recommendation quality on very long user behavior sequences while maintaining scalable computation.
DualPath presents an inference system that overcomes KV-Cache storage I/O bottlenecks in agentic LLMs by introducing dual-path KV-Cache loading. Beyond the traditional storage-to-prefill path, it enables parallel KV-Cache loading for prefill and decoding stages, reducing bandwidth contention and boosting overall throughput. This design alleviates the storage bandwidth bottleneck that constrains multi-turn agentic LLM inference.
MolHIT introduces a Hierarchical Discrete Diffusion Model for molecular graph generation. By generalizing diffusion to a hierarchical discrete framework, it addresses the discrete nature of 2D molecular graphs and improves chemical validity and property alignment compared to prior graph diffusion methods. The approach advances reliable AI-driven molecular design with graph-structured data.
SkyReels-V4 is a unified multi-modal video foundation model for joint video and audio generation, inpainting, and editing. It uses a dual-stream Multimodal Diffusion Transformer where one branch generates video and the other produces temporally aligned audio, both guided by a shared text encoder from Multimodal LLMs. The model accepts rich multi-modal instructions (text, images, video clips, masks, and audio references) and leverages in-context learning to follow complex prompts.
DreamID-Omni proposes a unified framework for controllable human-centric audio-video generation. It aims to consolidate reference-based audio-visual generation, video editing, and audio-driven video animation into a single model, with precise, disentangled control over multiple character identities and voice timbres. The framework seeks to provide flexible, unified control for complex cross-modal human-centric generation tasks.
We developed a passive surveillance system for early stroke risk detection in high-risk individuals with diabetes, using patient-reported symptoms. A symptom taxonomy grounded in patients' own language and a dual ML pipeline combining heterogeneous GNNs and EN/LASSO identified patterns predictive of stroke, leading to a hybrid risk screening tool.
We propose a self-supervised heterogeneous graph neural network to improve spatial allocation for energy system coupling with mismatched spatial resolutions. It models high-resolution geographic units as graph nodes and fuses multiple geographical features to produce physically meaningful aggregation weights across scales.
ECHO introduces high-order operators to encode communities in attributed networks, offering a scalable, self-supervised solution that overcomes the semantic walls of feature over-smoothing in dense or heterophilic networks and the systems wall from O(N^2) memory. It reframes community detection to scale with data size.
DyGnROLE proposes a transformer-based dynamic graph model that explicitly separates source and destination representations to capture asymmetry. It uses distinct embedding vocabularies and role-aware semantics to model temporal dynamics.
We develop deep ensemble graph neural nets to reconstruct cosmic-ray arrival direction and energy from voltage traces on ground-based radio detectors. Representing triggered antennas as a graph, the GNN leverages physical knowledge to improve accuracy and reduce training data requirements.
This work presents a physics-inspired neural framework for large-scale graph coloring, combining graph neural networks with statistical-mechanics ideas and planting-based supervision to address the algorithmic phase-transition challenges.
Atlas-free Brain Network Transformer proposes an atlas-free approach to constructing brain networks, avoiding fixed anatomical atlases that cause misalignment and biases. It introduces a transformer-based method to learn brain connectivity directly from data.
GYWI combines author knowledge graphs with retrieval-augmented generation to provide controllable context and traceable inspiration paths for LLM-based scientific idea generation. The system centers on building an author-centric knowledge graph to guide generation.
The paper analyzes retriever-reranker pipelines for RAG over knowledge graphs in e-commerce, comparing how to scale retrieval across connected graphs and preserve graph structure. It discusses challenges and benchmarks for KG-based RAG in practical applications.
Contextual Memory Virtualisation (CMV) treats accumulated LLM understanding as version-controlled state via a DAG-based memory model, enabling structured state management and lossless trimming of history for long-running agent sessions.
This work bridges granularity mismatch between LLMs and knowledge graphs by addressing token-level vs entity-level representations, proposing methods to better align semantic text with graph structure for knowledge graph completion.
TCM-DiffRAG develops a personalized syndrome differentiation reasoning method for Traditional Chinese Medicine using knowledge graphs and chain-of-thought prompting to tackle diverse diagnostic patterns and individual differences.
G-reasoner presents foundation models for unified reasoning over graph-structured knowledge, addressing fragmented information in RAG by integrating graph-structured data into reasoning with LLMs.
MAGNET proposes Modality-Guided Mixture of Adaptive Graph Experts with entropy-triggered routing to fuse heterogeneous multimodal signals for recommendation, addressing modality imbalance and entangled representations.
PoSh introduces a detailed image description metric that uses scene graphs as structured rubrics to guide LLMs-as-a-judge, enabling fine-grained attribution and relation-aware evaluation.
AlayaLaser tackles I/O-bound misperception in on-disk graph-based vector search by proposing an optimized index layout and search strategy that focuses on compute efficiency for high-dimensional vectors.
VeloANN improves SSD-resident graph indexing for high-throughput vector search by introducing a locality-aware data layout and coroutine-based asynchronous processing to reduce storage stalls and improve CPU utilization.