Abstract: How can we automatically assess the quality of machine translation output in the absence of human reference translations? A solution to Quality Estimation of Machine Translation involves employing the translations from heterogeneous translation systems as the pseudo reference. We propose a pseudo reference generation method incorporating a reflection mechanism that takes the machine translation results under evaluation as a critical input to the pseudo reference generation process. The reflection agent infers the pseudo reference with four key steps: back translation generation from the machine translation result, reflection by comparing the source language sentence and the back translation, correction suggestions for the machine translations, and candidate pseudo reference generating. Experimental results on the WMT’23 sentence-level Quality Estimation of Machine Translation benchmark dataset demonstrate that the proposed method significantly enhances Quality Estimation of Machine Translation performance.