A novel variation of the probed recall, negative recency paradigm was employed to test the general hypothesis that the amount of learning from an initial recall test is a function of the level of processing used during initial retrieval, and to test specific predictions concerning the role of initial tests in the production and elimination of the negative recency effect. On each of 32 trials, 32 subjects each saw four words followed by a searchmode signal and a probe word. The level of initial retrieval processing (acoustic or semantic) was controlled by the search-mode signal. Initial test response latencies were recorded, and the learning derived from initial tests was assessed by an unexpected final free recall test. Final recall of words initially retrieved in the semantic-search condition showed no negative recency, while final recall of words initially retrieved in the acoustic-search condition did show negative recency. Conditional analyses showed that memory was incremented more by semantic-level initial retrieval than by acoustic-level initial retrieval. Implications of these results for an explanation of negative recency effects and for other learning-from-tests effects are discussed.The paradoxical result known as the negative recency effect is simply stated: In a list that is serially presented for free recall or probed recall, the last items presented are the most likely to be recalled on an initial test and are the least likely to be recalled on a subsequent test. Logically, the negative recency effect could result from inferior initial storage of recency items, from greater