The comprehension processes of adult students solving two-step problems of comparison were studied using eye-movement experiments based on the assumption that eye fixations are synchronous with internal cognitive processes. Twenty university students each solved 24 two-step mathematical word problems of three sentences each with consistent language (CL) or inconsistent language (IL) structures. In CL problems, the unknown variable was the subject of the second sentence; in IL, the unknown was the object of the relational sentence. Raw eye movement data were transformed into consecutive fixations. Overall results were generally in line with the model of A. Lewis and R. Mayer (1987) (LM model) suggesting the increased difficulty of IL problems as indicated by increased time required to process the sentences and the number of errors of reversal of correct mathematical operation. Apparently, just as it is the case for the performance variable error rate, the effect of inconsistent language on aspects of the solution process (such as total solution time, initial reading time, and fixation time on the relational clause) is exhibited only when the compare problems have to be processed and solved under rather heavy cognitive demands. Some theoretical conceptions are developed as alternatives to parts of the LM model. One table and one bar graph supplement the discussion. (SLD)