This study explored the processing of multi-dimensional stimuli during mental comparisons by means of behavioral and event-related potential (ERP) measures. As multidimensional stimuli served compound digits with separable physical and semantic dimensions, each graduated in size. Fourteen adult volunteers were presented with the digits zero through nine, which were constructed by suitably arranged small digits of the same number set. The subject's task was to decide whether the physically larger digit was numerically larger or smaller than the physically smaller one. Large and small numbers as well as the numerical differences between them appeared, equiprobably, in randomized order. Reaction time (RT), frequency of errors, and ERPs registered from Fz, Cz and Pz (time constant 5.00 s; 30 Hz frequency cut off) were selectively averaged according to three dimensions (i) the numerical differences between simultaneously presented large and small digits, (ii) the 10 physically large digits, and (iii) the 10 physically small digits. As a reflection of the symbolic distance effect, the mean RTs and errors decreased and the P300 amplitude increased with increasing differences between the numbers compared. The P300 amplitude changed in a U-shaped function of the numerical value of the physically large digits. This result is in accordance with the adaptation-level hypothesis and indicates that prior to numerical comparisons the physically large numbers were primarily encoded and arranged into the internal number scale. The fact that the U-shaped amplitude trend was found for the physically large but not for the physically small digits is interpreted as an expression of the advantage of global information. The approach of separation task-relevant stimulus dimensions on the basis of P300 changes is proposed as an informative method for the study of multi-dimensional-stimulus processing in multi-task environments.