I have a 3.05Å structure. I refine it to Rwork=25.2 Rfree=28.81. I use
Molprobity to check the structure. The clash score is 100th and Molprobity
score is 99th. I was suggested to run simulated anneal to help reduce
Rwork and Rfree. I tried Cartesian and Torsion and different seed. The
best result I got is: torsion anneal, start_temp=2500,final_temp=300,
random_seed=4250146. The Rwork:25.8, Rfree=31.2. I checked the new model
and found that some places in a domain were not correctly traced. The
density for that domain is not good and have several flexible loops. My
question is: does simulated anneal always work and reduce Rwork/Rfree?

No, especially when the structure is nearly finished.

At the final stage of refinement, does it necessary to use simulated anneal
again?

Definitely not necessary. Simulated annealing is a good way to improve a model that fits the data poorly, or to remove model bias, but at the end of refinement you should be much more conservative.

There are several other potential ways to reduce the R-free - such as using TLS, or optimizing the weights. But you already have excellent geometry and your R-free and R-free/R-work gap are both reasonable for that resolution, so it's possible that further refinement won't help.

-Nat