Air pollutant sources and regional transport are important issues in air quality control. The GRAPES-CUACE (Global-Regional Assimilation and Prediction System coupled with CMA Unified Atmospheric Chemistry Environment) aerosol adjoint model was applied in detecting the emission sources of a haze episode in Beijing during Nov. 19–21, 2012. Two high PM<sub>2.5</sub> peaks were measured respectively at 05:00 and 23:00 Beijing Time on Nov. 21<sup>st</sup> 2012, which were set as the objective functions for the aerosol adjoint model. The sensitive emission regions of the first PM<sub>2.5</sub> peak were tracked to the west and south to Beijing with 2–3 day cumulative transport of air pollutants to Beijing, while the sensitive emission regions of the second peak were mainly located at the south to Beijing, where southeasterly moist air transport lead to the hygroscopic growth of particles and pollutant convergence in front of the Tai-hang Mountains in the daytime of Nov. 21<sup>st</sup>. The temporal variations of the sensitivity coefficients for the two PM<sub>2.5</sub> concentration peaks in Beijing revealed that the response time of Beijing haze pollution to the local emissions was about 1–2 hours and to the surrounding emissions was about 7–8 hours. The contribution of surrounding emissions accounted for 67.3 % and 61.0 % for the first and second PM<sub>2.5</sub> pollution peaks respectively. The adjoint results were compared with the forward sensitivity simulations set of the Models-3/CMAQ system. Two modeling approches were well comparable in the assessments of atmospheric pollution control schemes for critical emission regions. The adjoint methods had higher computational efficiency than the forward sensitivity simulations.