Well, the thing is that your data are most likely not good enough to refine these extremely correlated parameters.

If you ever hope to refine B values you MUST have very good quality X-ray data at high 2Theta. Do not forget though that you are limited by the intensity fall off at higher T. Much better chances are with good quality neutron diffraction data. If none of these options are possible (measurement T-is too high or you have no access to neutrons) then:

1) you may try to find literature values and trust them

 2) do expansivity measurements from a several K to room T and refine B values as long as you can. At some point you will have to linearly extrapolate from lower temperatures. Not the best solution but still better than using some meaningless numbers.

Thanks for your answer. I was actually refining a neutron data. So refining B parameters should work out but still the parameters are negative even if the fit is good.

OK, I see. The fit can be very good even with totally wrong B numbers. If your refinement gets stuck in some local minimum then minus B values may be the only way to increase the match. This is also why humans are still needed to do the analysts :). 

Was this constant wavelength (CW) or time of flight (TOF)?

For CW neutron data you should check the resolution curve of your diffractometer. Some monochromators like pyrolytic graphite are really lousy at higher 2-Theta. Ge and Cu are much better in this respect even if they give typicality far smaller neutron flux at the sample position.  Also here you need fairly high 2_Theta range to pin B values down at longer wavelengths. You may need to reposition the detector and integrate data to cover wider range.  

I have very little hands-on experience with TOF but as far as I know B values are not all that correlated and if your profile function and resolution in d-spacing are good you should have no problem.

Thanks for u suggestion. It is a constant neutron data and pyrolytic graphite was used as a monochromator. And my data is till 95 degrees. 

Ehhh, this may be the problem. For D20 powder diffractometer at ILL the peak resolution for the pyrolytic graphite is at ~45 deg (2.41 Lambda). I guess similar will be in your case. At higher angles there is a very strong loss of the resolution. You cannot reliably refine B values at higher temperatures in such case. I think I was able to refine these for CO2 D2O clathrate up to 120 K an only after grouping shifts for O, D and CO2 at two sites. At even higher T one could refine global isotropic displacement assuming their linear change (that might true but not necessarily is).