Aiming at the problem that evidence selection in retrieval enhancement generation focuses on semantic relevance, and it is difficult to ensure causal necessity and minimum sufficiency, a retrieval enhancement generation method based on verifiable causal sufficient evidence set was proposed. Method A counterfactual delete-replace intervention is used to construct an evidence-level supervision signal. A pre-trained language model causal re-ranking, R-GCN evidence complementary modeling, quality proxy function calibration, and minimum sufficient set search are combined to jointly optimize the answer quality, evidence size, and reasoning cost. Experiments on HotpotQA, HOVER, QASPER and ALCE datasets show that the proposed method achieves F1 of 67.1 on HotpotQA and 82.3 on HOVER, with an average evidence size reduction of 18%-22%. The context overhead is reduced by up to 60%, and it maintains better robustness under the condition of conflict evidence and near-repeated interference. The results show that the proposed method can effectively improve the verifiability, compactness and stability of the retrieval enhancement generation system.
