We use a novel paradigm to test models of long-term recognition memory: After studying lists, tests are made with two items, both OLD, both NEW, or one of each. Some tests used Two-Alternative Forced Choice (2AFC) in which Ss were asked to choose the item more likely OLD (Experiment 2 asked Ss to choose the item more likely NEW); other tests used four-way classification (4WC) in which Ss were asked to classify the two items as 1) both old, 2) both new, 3) left old, right new, or 4) left new, right old. Each S studied lists containing 12 words, 24 words, 12 pictures, 24 pictures, or lists of 12 words randomly mixed with 12 pictures (so tests were both words, both pictures or one each). All the choice probabilities were predicted well by the Retrieving Effectively from Memory model (REM) of Shiffrin and Steyvers (1997) using mostly the three 1997 parameter values and the REM decision threshold of odds of 1.0. Signal-detection modeling (unequal variance Gaussian strength distributions) predicted the choice probabilities with different parameters for different conditions. Initial analysis and modeling of Response times suggested that REM may be well suited to predict differing accuracy and response time results for judgments of OLD and NEW.