Due to wide spread usage of the satellite positioning techniques especially GPS, we need to precisely determine geoid model in order to use GPS measurements for height determination, as an alternative of traditional leveling techniques in geodetic applications. Precise local geoid modelling using GPS/Leveling data, apart from the existing models such as geopotential models and gravimetric geoid models could be an interesting investigation topic. An important question is, ‘What accuracy level can be achieved using this approach?’ However precession of this modelling could be influenced by some issues such as data quality or modelling techniques. In this paper, we attempt to assess the implementation of modern learning-based computing techniques including artificial neural networks and adaptive network-based fuzzy inference systems compared with multivariate polynomial regression equations in GPS/Leveling Geoid modeling. This assessment carried out in a small and dense network of GPS/Leveling benchmarks in contrast with previous studies, located in shahin-shahr, Isfahan. And these high quality data make it possible to achieve an accuracy of better than 1 cm. The results show a few millimeter superiority of ANN and ANFIS derived geoid models in terms of root mean square error, as well as in terms of coefficient of determination. And RMSE=8cm, R2=0.9949 and RMSE=7cm, R2=0.9964 achieved for this models respectively. Therefore ANFIS derived geoid model provide the most accurate geoid heights in the study area.