Foreshocks are one of the few well-documented precursors to large earthquakes; therefore, understanding their nature is very important for earthquake prediction and hazard mitigation. However, the triggering role of foreshocks is not yet clear. It is possible that foreshocks are a self-triggering cascade of events that simply happen to trigger an unusually large aftershock; alternatively, foreshocks might originate from an external aseismic process that ultimately triggers the mainshock. In the former case, the foreshocks will have limited utility for forecasting. The latter case has been observed for several individual large earthquakes; however, it remains unclear how common it is and how to distinguish foreshock sequences from other seismicity clusters that do not lead to large earthquakes. Here we analyze foreshocks of three M>7 mainshocks in southern California. These foreshock sequences appear similar to earthquake swarms, in that they do not start with their largest events and they exhibit spatial migration of seismicity. Analysis of source spectra shows that all three foreshock sequences feature lower average stress drops and depletion of high-frequency energy compared with the aftershocks of their corresponding mainshocks. Using a longer-term stress-drop catalog, we find that the average stress drop of the Landers and Hector Mine foreshock sequences is comparable to nearby swarms. Our observations suggest that these foreshock sequences are manifestations of aseismic transients occurring close to the mainshock hypocenters, possibly related to localized fault zone complexity, which have promoted the occurrence of both the foreshocks and the eventual mainshock.