CARA is a training‑free, causally guided retrieval agent designed for retrieval‑augmented generation over text‑rich knowledge graphs. It evaluates retrieved evidence with two causal relevance criteria: consequentiality, which measures whether adding evidence improves answerability, and essentiality, which measures whether removing evidence weakens answerability.
These assessments are combined with path consistency to steer multi‑hop graph exploration. To merge evidence from multiple agents, CARA employs reciprocal rank fusion.
We augment three STaRK benchmarks with ground‑truth answers and conduct downstream QA evaluation. Experiments show that CARA consistently outperforms existing retrievers in retrieval accuracy, answer quality, and faithfulness, while examining fewer graph nodes and requiring fewer retrieval operations. The results demonstrate that causal relevance assessment makes graph retrieval more focused and effective, boosting overall retrieval and QA performance.
Review