GraphEcho: Structural Redundancy and Evidence Provenance in LLM Graph Agents
The paper warns that more graph paths do not always mean more evidence. AI systems that use tools to follow graph links can count repeated routes as support. Provenance-aware training reduced repeated walks, but it could also cover fewer sources and lose accuracy on scientific claims. Independence has to be measured, not assumed from route count.
What the source reports
The GraphEcho paper studies AI systems that follow graph links while answering questions. It reports that those systems can mistake repeated routes for separate support. A training method that tracks provenance reduced repeated walks. Provenance means the source path behind a claim. The same change could cover fewer distinct sources. It could also lower accuracy on scientific claims. The limit is clear: less repetition did not mean better results on every measure.
Original source
- Title
- GraphEcho: Structural Redundancy and Evidence Provenance in LLM Graph Agents
- Author
- Sikun Wang, Yixi Zhou, Lei Fan, Fan Zhang
- Publication
- arXiv
- Date
- Thursday, September 17, 2026