Showing 9 papers for 2026-04-25
GS-Quant introduces a semantic-aware quantization framework that bridges continuous graph embeddings and discrete LLM tokens for knowledge graph completion. By generating semantically coherent codes that respect hierarchical reasoning, it aims to preserve structure and meaning during quantization, improving KGC performance.
AtomicRAG proposes atom-entity graphs to compose retrieval-augmented generation with more flexible, atomic facts rather than fixed text blocks. It moves beyond treating chunks as indivisible units and reduces reliance on brittle triple-based entity linking, mitigating errors in relation extraction and enabling more robust reasoning.
KGiRAG presents an iterative, feedback-driven GraphRAG architecture to answer sensemaking queries that surpass LLMs' static knowledge and struggle with large contexts. By incorporating response quality assessments and iterative retrieval rounds, it grounds answers more accurately and reduces hallucinations.
TravelFraudBench (TFG) is a configurable benchmark for evaluating GNNs on fraud ring detection in travel platform graphs. It addresses gaps in existing benchmarks by simulating three travel-specific ring topologies—ticketing fraud with star-like device/IP clusters, ghost hotel schemes with reviewer-hotel bipartite cliques, and account takeover rings—allowing systematic evaluation of robustness across structures.
EngramaBench provides a benchmark for evaluating long-term conversational memory with structured graph retrieval. It comprises five personas, 100 multi-session conversations, and 150 queries across categories like factual recall, cross-space integration, temporal reasoning, adversarial abstention, and emergent synthesis. The paper evaluates Engrama, a graph-structured memory system, against GPT-4o prompting and Mem0.
PLOTTER introduces graph-based narrative planning, using an event-graph and character graph to maintain global coherence and character development. Through Evaluate-Plan-Revise the cycle, it iteratively refines plans before generation to improve narrative quality.
We propose a novel graph neural network architecture that combines residual Graph Isomorphism Networks (RGIN) with attention mechanisms to capture complex drug–drug and drug–target interactions. The approach aims to improve accuracy and efficiency in predicting synergistic drug combinations for effective therapy design.
Proposes a multimodal approach that leverages both textual content and graph structures to enable open-domain event extraction from documents. It aims to generalize to unseen event types and harnessing LLMs to enhance extraction, normalization, and reasoning.
Knowledge Capsules introduce structured nonparametric memory units to augment LLMs beyond parametric weights and pure context expansion. They aim to provide stable, directly accessible knowledge for long-horizon reasoning and multi-hop tasks, improving robustness of retrieval-augmented generation.