Abstraction May be Postponed Until the Transfer Phase Throughout this discussion, we have tacitly assumed that the abstract knowledge revealed transfer tasks is elaborated during the study phase. Alternatively, however, subjects may simply encode specific items during the study phase, store them memory, and perform abstractive operations on their reminded representations the presence of new transfer items. For example, subjects asked for a grammaticality judg- ment about DJDMJJ the transfer phase may retrieve study items such as CVCPVV, SCPTVV, and so on, and then abstract a common feature: Repetitions of letters occur at the end of the items all cases. This alternative bears on the distinction between early and late computation models of categorization (Estes, 1986). Although early computation models, which full analysis of study exemplars takes place the encoding phase, are implicit traditional frameworks, there is a growing body of evidence for the late, or in line (Smith, 1989), computation model, which at least some processing occurs while subjects are dealing with new stimuli (e.g., Brooks, 1990; Medin & Ross, 1990). This alternative is obviously of crucial importance to the issue at hand. If researchers aim at eliciting implicit processing the study phase (although unsatisfactorily, as stressed above), they have never controlled the nature of the processing occurring the test phase. Typically, subjects at the beginning of the test phase are informed that the study strings were generated by a complex set of rules and that they should now assess the well-formedness of new items with regard to these rules. These instructions inevitably shift subjects to a rule discovery mental set. Mathews et al. (1989) did not mention the rule-based structure of items, but their test procedure remained explicit nature insofar as their subjects were instructed to make direct comparison between the study and test items. 2 Thus, clear evidence for a model which abstrac- tion is performed during the transfer task would be highly damaging to the claim that performance artificial grammar settings testifies to implicit abstraction. The recent Mathews et al. (1989) experiments were not designed to tackle this issue but nevertheless provided data suggesting that at least some rules are abstracted the transfer phase. For instance, above-chance grammaticality judgments for items made up of new letters did not appear immediately after letter change but rather after some feedback trials. This result is consonant with the idea that subjects do not generate ready-to-use abstract rules study phase but rather build them when given a problem that prompts rule elaboration. In the same vein, abstract verbalization was apparentlymthe Mathews et al. (1989) article is not entirely clear on this point--produced only on request the instructions.