Note also that one must be extremely careful in specifying the predictors. There may be neural activity that varies with the duration of the intra-trial spacing (i.e., delay related activity). In this case, designs may suffer from multicollinearity, even with long null events.
Here are some relevant papers that I've been looking at for my own benefit:
Andrade, a, Paradis, a L., Rouquette, S., & Poline, J. B. (1999). Ambiguous results in functional neuroimaging data analysis due to covariate correlation. NeuroImage, 10(4), 483-6. doi:10.1006/nimg.1999.0479
Ollinger, J. M., Corbetta, M., & Shulman, G. L. (2001). Separating processes within a trial in event-related functional MRI. NeuroImage, 13(1), 218-29. doi:10.1006/nimg.2000.0711
Ruge, H., Brass, M., Lohmann, G., & von Cramon, D. Y. (2003). Event-related analysis for event types of fixed order and restricted spacing by temporal quantification of trial-averaged fMRI time courses. Journal of magnetic resonance imaging : JMRI, 18(5), 599-607. doi:10.1002/jmri.10397
Ruge, H., Goschke, T., & Braver, T. S. (2009). Separating event-related BOLD components within trials: the partial-trial design revisited. NeuroImage, 47(2), 501-13. Elsevier Inc. doi:10.1016/j.neuroimage.2009.04.075
Serences, J. T. (2004). A comparison of methods for characterizing the event-related BOLD timeseries in rapid fMRI. NeuroImage, 21, 1690 - 1700. doi:10.1016/j.neuroimage.2003.12.021
In particular, Ruge et al. (2009) have a very rigorous treatment of partial trials. I wonder if it's even possible to reduce multicollinearity substantially through sub-12 sec jittering alone, perhaps through an optimal search of the design space (e.g., Wager & Nichols, 2003, or Kao et al., 2009). I haven't run any simulations, but my expectation is that it's probably not possible.