Thanks Aidan and Steve,
Time step was the key idea. I was using a time-step of 1 fs from the RDX example, which was evidently too large for CNTs simulated for longer simulation times. Adri van Duin (Slide 30 at link below) suggests using a time step below 0.025.
www.wag.caltech.edu/home/duin/FFgroup/Dirt.ppt
I re-ran my AI-REBO/ReaxFF comparison simulations with appropriate respective time-steps (same total # of time steps 2000 fs). Reaxff sustained a stable CNT structure with total energy conserved to 4+ decimal places. And there is also acceptable quantitative agreement between the two calculations- a ~10% difference in total energy at ~460K.
Maybe this post about the time step will help other users. In any case, thanks very much for your input!
Sincerely,
Nicholas J. Lee