See Wold, S., Sjöström, M., & Eriksson, L. (2001). PLS-regression: a basic tool of chemometrics. Chemometrics and intelligent laboratory systems, 58(2), 109-130. for answers to this question. High positive or negative coefficients mean that the predictor x has a large effect on the response variable y that you want to model. On the other hand, one x variable can explain a lot of the variability in other x variables. The VIP takes this into account. As a result, you can have a high VIP if x explains a large part of the variability of X, and at the same time have low coefficients for x if it has a small effect on the response variable y. Wold et al. 2001 use the VIP as well as the coefficients to prune the model to eliminate unimportant variables. They do this by eliminating the x variables with low VIP and low coefficients.