Graph Traversal Context Crucial for Citation Faithfulness in Agentic GraphRAG
A recent paper on arXiv (2605.15109) examines the concept of citation faithfulness in Agentic GraphRAG, where an agent navigates a knowledge graph prior to generating responses and citations. The research presents faithfulness as an issue at the trajectory level, emphasizing that final citations should reflect the graph's traversal, its structure, and entities that were visited but not cited. Through controlled ablation studies, it is revealed that while cited evidence is often essential—its absence significantly alters answers and diminishes accuracy—it is not solely adequate, as precise answers can also rely on the context of uncited traversals and the surrounding graph structure. Thus, the study indicates that citation assessment should encompass the full traversal context.
Key facts
- Paper arXiv:2605.15109 examines citation faithfulness in Agentic GraphRAG.
- Agentic GraphRAG involves an agent exploring a knowledge graph before producing answers and citations.
- Citation faithfulness is framed as a trajectory-level problem.
- Final citations should support the answer and account for graph traversal, structure, and visited-but-uncited entities.
- Controlled ablation experiments compared isolating, removing, and masking cited and uncited graph entities.
- Removing cited evidence substantially changes answers and reduces accuracy.
- Accurate answers can depend on uncited traversal context and surrounding graph structure.
- Citation evaluation should consider the entire traversal context.
Entities
Institutions
- arXiv