The reliability of event-related potentials (ERPs) was studied in 10 healthy adults who were tested 8 times over 7-10 day intervals using a standard auditory oddball paradigm. The difference waveforms, obtained by subtracting the averaged waveforms for frequent trials from those obtained in rare trials, were designed to analyze the components of the ERPs, such as the P300, and to focus on the reliability of the probability effect on the ERPs. The between-session reliability (8 sessions) and the within-session reliability (order of blocks or of different visual procedures) were computed for the obtained difference waveforms. The between-session reliabilities, expressed as the intraclass correlations (r') for the P300 amplitude, area and latency, were 0.70, 0.61 and 0.65, respectively. The within-session reliability, presented as the Pearson correlation coefficients (r) for the three P300 measures were 0.43, 0.35 and 0.25 for different eyes. The values were 0.45, 0.39, 0.42 between the first and the second blocks (eyes-open) and 0.58, 0.47 and 0.29 (eyes-closed). These findings indicate that the P300 amplitude calculated from the difference waveforms may be the most stable marker for the between-session reliabilities. There were no significant differences in the P300 measures over the 8 sessions, suggesting that habituation may not occur with the difference waveform reflecting the probability effect on ERPs. The difference waveform may be useful in research on repetitive group ERPs.