Showing 12 papers for 2026-03-13
We propose Graph Tokenization for bridging graphs and transformers: a framework that tokenizes graphs into sequential representations by combining reversible graph serialization with Byte Pair Encoding, preserving graph information for Transformer-based models. This approach aims to better capture structural information in graphs when processed by LLMs.
KEPo studies Knowledge Evolution Poison on Graph-based Retrieval-Augmented Generation. It shows that attackers can inject poisoned texts into external databases used by GraphRAG to manipulate LLM outputs for attacker-chosen queries. The work analyzes threat surfaces and discusses robustness strategies tailored to GraphRAG beyond conventional RAG approaches.
We propose Effective Resistance Rewiring (ERR), a simple topology correction that uses global effective resistance as a signal to rewire graphs and alleviate over-squashing in Graph Neural Networks. Unlike local criteria, ERR targets global connectivity constraints to improve long-range information flow with minimal changes.
DNS-GT introduces a graph-based Transformer to learn embeddings of domain names from DNS queries for network intrusion detection. By modeling relationships among domains as a graph and applying Transformer-based representations, it aims to improve detection performance while reducing reliance on labeled data.
AGMARL-DKS proposes Adaptive Graph-Enhanced Multi-Agent Reinforcement Learning for dynamic Kubernetes scheduling. It tackles scalability and heterogeneity challenges of centralized RL schedulers, introducing graph-based multi-agent coordination to balance system stability, resource utilization, and costs in dynamic clusters.
MDER-DR introduces Map-Disambiguate-Enrich-Reduce, a KG-based indexing approach for multi-hop QA that preserves contextual nuance lost when text is reduced to triples. It covers both indexing and retrieval/inference phases in a domain-agnostic manner to improve multi-hop question answering.
HELM proposes Hierarchical and Explicit Label Modeling for multi-label image classification. It uses hierarchy-specific class tokens within a Vision Transformer to capture label interactions and employs graph convolutional networks to model label dependencies, enabling effective use of unlabeled data and handling complex hierarchies.
From Entity-Centric to Goal-Oriented Graphs (GoGs) for Minecraft introduces a GoG framework where each node represents a goal and edges encode logical dependencies, enabling LLM-based step-by-step knowledge retrieval and planning in Minecraft.
EST (Entity State Tuning) harmonizes structure and sequence for temporal knowledge graph forecasting by maintaining persistent entity representations across timestamps, addressing episodic amnesia. It is encoder-agnostic and enables long-term dependencies to be captured by forecasters.
Enhancing Heterogeneous Multi-Agent Cooperation in Decentralized MARL via GNN-driven Intrinsic Rewards proposes using graph neural network derived intrinsic rewards to improve cooperation among heterogeneous agents under partial observability and sparse rewards.
ARK presents Answer-Centric Retriever Tuning via KG-augmented Curriculum Learning, a fine-tuning framework that optimizes the retriever for answer alignment by identifying high-quality evidence and training with a knowledge-graph–augmented curriculum, bridging retriever objectives and downstream answer quality.
OMNIA proposes a two-stage approach to knowledge graph completion that closes the loop by leveraging LLMs to bridge structural and semantic reasoning for KGC, reconciling noisy/ incomplete outputs with graph structures.