Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 3 papers for 2026-08-15
The paper introduces a new unified, multi-dimensional benchmark for complex graph reasoning in large language models, designed to be programmatically generated, structurally controllable, and scalable to long inputs. It identifies shortcomings of existing benchmarks—limited data complexity coverage, heavy manual construction, and lack of unified evaluation across text-based and code-based reasoning—and proposes a five-stage semi-automatic framework to address these gaps, with an emphasis on cross-modal evaluation and comparability.
The work presents EGRL, an edge generation-guided, relation-aware learning framework for RNA–protein interaction prediction (RPIP). It mitigates data sparsity and cold-start issues by augmenting graphs with edge generation and learning rich, relation-aware representations, moving beyond homogeneous graphs or fixed meta-paths. The approach aims to improve RPIP accuracy and generalization in challenging, low-resource settings.
The paper scrutinizes GraphRAG evaluation practices, arguing that current assessments suffer from unrelated questions and various biases that can skew conclusions about performance. It proposes an unbiased evaluation framework to measure GraphRAG quality more fairly, including improved data selection, evaluation protocols, and metric design. The goal is to enable reliable, apples-to-apples comparisons of retrieval-augmented generation systems that use graphs.