Hi Sonali
The 'WEIGHT MATRIX Wm' scale factor, to which I assume you're referring, is on a relative, not absolute, scale so is not comparable between different models, i.e. the results will be substantially different when you change the model keeping Wm fixed as you discovered. If you want the weight to be more easily comparable for different models (and also for data with different resolution limits) you need it to be on an absolute scale: this is what you get with 'WEIGHT AUTO Wa'. For a theoretically perfect model the optimal value of Wa would be 1, i.e. the geometrical and X-ray weights would be on the same absolute scale, though in practice the optimal value of Wa usually turns out to be a little higher than 1 (say between 1 and 4). The default Wa value (using just 'WEIGHT AUTO') is 10: I find this is suitable for refining MR solutions where you need the model to be less geometrically rigid so the X-ray contribution needs to be inflated relatively, but as the model improves Wa needs to be decreased towards the theoretical value of 1 (or certainly not much less than 1).
Note that this Wa is conveniently the same as the Wa used in X-PLOR, CNS & phenix.refine (so it can be transferred between programs), though in the latter case I believe it stands for weight(absolute), not weight(automatic). If I have this wrong, no doubt the X-PLOR/CNS/phenix.refine people will put me straight!
Other than that I agree with everything Robbie said & you should heed his advice.