There have been several problems in experimental studies about the mediational processes of discrimination shift learning. Two of these problems are that a non-reversal shift (NRS) learning was treated as a secondary event to be comparred with results in a reversal shift (RS) learning, and that the abilities required in the shifts were only indirectly measured by the amount of Subjects' (Ss') verbal knowledge about the relevant dimensions or values. We must devise an effective method to get more direct information about the abilities required in the shifts. The present study was designed to solve the above-mentioned problems. Discrimination shift learning has been dealt with for a long time mainly from the point of view of behaviorism, but in this paper I would like to re-examine the verbalmediational hypothesis of Kendler, H. H. & Kendler, T. S.(1962), in terms of Piaget's views of the development of logical manipUlation by classification. The tasks comparing the sizes of intradimensional value-classes (intra-d. task) and comparing the sizes of interdimensional value-classes (inter-d. task) were devised in order to measure the Ss' levels of logical manipulations prior to the first learning. All Ss (5 and 6 year-old) got correct answers in the intra-d. task, while they made various responses in the inter-d. task. So they were divided into two groups (higher group: H. G. and lower group: L. G.) by the scores of this inter-d. task. And the inter-d. task is analogous to Piaget's “class inclusion task”, hence H. G. are judged as children shifting into the concrete operational stage in which cognition by language gradually becomes dominant, on the other had, L. G. are judged as children remaining at the intuitive thinking stage in which cognition by perception is still dominant. Therefore, H. G. can be considered to correspond to the older children and L. G. to the younger ones or the children staying at a transitional period in the Kendlers' hypothesis. Results were as follows: (1) There was no difference between H. G. and L. G. in a RS learning.(2) H. G. obtained better results than L. G. in a NRS learnig.(3) H. G. attained a RS learning more easily than a NRS learning.(4) L. G. also learned a RS learning more easily than a NRS learning and the difference between the RS and NRS learning was greater in L. G. than in H. G. The last result was quite contradictory to that deduced from the Kendlers' hypothesis, and it cannot be explained from their views. Here in this paper a new hypothesis is proposed whose construction is based on “the development of logical manipulation by classification”. In this hypothesis the term of “mediational learning” is used separately in each learning of RS and NRS. What are suggested from this new hypothesis are as follows: (1) Children who can make mediational responses in RS learning can't always make mediational ones in NRS, learning.(2) The verbal factor (or Piaget's logical operation) influences a NRS learning rather than a RS learning. Accordingly, it is at the time when a child shifts to the concrete operational stage where he can simultaneously consider the same stimuli from the two different dimensions, that he starts making mediational responses in NRS learning.(3) If an experimenter uses the familiar stimuli for Ss, they may be able to make mediational responses in RS learning, at a rather early period when cognitiOn by language is not yet dominant. Therefore, the existence of “non-verbal mediational events” may be expected.