In a recent study, we tested perceptual learning in adults with autism spectrum disorder (ASD) (Harris et al., 2015), employing the standard and well-established texture-learning paradigm [TDT; (Karni and Sagi, 1991; Sagi, 1995; Harris et al., 2012)]. In this paradigm, observers learn to discriminate an oriented texture target embedded at a fixed location in a background of elements having a different orientation. Performance is measured as a function of the time-interval between the onset of the target and a mask (stimulus onset asynchrony, SOA), with threshold defined as the minimal time (SOA) to reach a predefined criterion level of performance. Typical observers improve their performance (show reduced thresholds) with training across 3–4 days, but need to relearn the task when the target is moved to a different location in the visual field, showing specificity. We (Harris et al., 2015) reported similar results with observers with ASD, but unlike the typical observers who showed faster learning at the second location (Sagi, 2011), ASD observers showed difficulty in relearning the task at the second location, suggesting that the training with the target at the first location might have interfered with the training at the new, second location. We termed this anomalous poor learning “over-specificity” (OS) to reflect the narrowness of the learning and the failure to generalize, and quantified OS as the average threshold difference between the second and the first learning curves (for generalization OS 0). A modified learning paradigm, where standard target trials were interleaved with no-target trials (“dummy” trials) during training, showed generalization of learning (OS 1) has never been observed before.