Self-monitoring of errors in picture naming has been shown to correlate positively with aphasia treatment outcomes (Marshall et al., 1994). To explore potential explanations, we took advantage of aphasics’ item-level inconsistency in naming accuracy, e.g., in the baseline phase of treatment studies. We looked for a “monitoring benefit” such that items erroneously named on baseline 1 would be more likely to be named correctly on baseline 2 if accompanied by spontaneous monitoring, relative to unmonitored errors. Such a monitoring benefit might reflect learning (e.g., in response to the self-generated error signal) and/or the differential strength (e.g., proximity to threshold) of monitored items compared to those that are not monitored. To implicate learning requires evidence that the monitoring benefit is directionally asymmetric, i.e., the monitoring-related change in accuracy from baseline 1 to 2 (forward direction) is greater than the monitoring-related change from baseline 2 to trial 1 (backward). Twelve participants with chronic stroke aphasia participated. All were mild-to-moderate in aphasia severity and naming impairment. Each participant named the same 615 pictures, without feedback, on two occasions (2 “baselines”) in separate weeks. On each trial, we scored the accuracy of the first attempt and the type of error; additionally, each error was assigned a monitoring code: DetNoCorr (detected without correction; Ex. 1, 3); DetCorr (detected with correction (Ex., 2, 4); or NoDet (not detected). Ex. 1. (T = squirrel) “chipmunk, no”: Semantic error; DetNoCorr Ex. 2. (T = squirrel) “chipmunk, no, squirrel”: Semantic error; DetCorr Ex. 3. (T = umbrella) “umbelella, that’s not right”: Phonological error; DetNoCorr Ex. 4. (T = umbrella) “umbelella, umbrella”: Phonological error; DetCorr We used mixed effects logistic regression to assess whether the log odds of changing from error to correct was predicted by monitoring status of the error (DetCorr vs. NoDet; DetNoCorr vs. NoDet); whether the monitoring benefit interacted with direction of change (forward, backward); and whether effects varied by error type. Figure 1 (top) shows that the proportion accuracy change was higher for DetCorr, relative to NoDet, consistent with a monitoring benefit. The difference in log odds was significant for semantic errors in both directions (forward: coeff. = -1.73; z= -7.78; p < .001; backward: coeff = -0.92; z= -3.60; p < .001), and for phonological errors in both directions (forward: coeff. = -0.74; z= -2.73; p=.006; backward : coeff. = -.76; z = -2.73; p = .006). The difference between DetNoCorr and NoDet was not significant in any condition. Figure 1 (bottom) shows that for Semantic errors, there was a directional asymmetry favoring the Forward condition (interaction: coeff. = .79; z = 2.32; p = .02). Phonological errors, in contrast, produced comparable effects in Forward and Backward direction. The results demonstrated a benefit for errors that were detected and corrected. This monitoring benefit was present in both the forward and backward direction, supporting the Strength hypothesis. Of greatest interest, the monitoring benefit for Semantic errors was greater in the forward than backward direction, indicating a role for learning.