Simulated tempering was invented by Parisi and Marinari to study disordered
Ising spin models (spins in a random external magnetic field). These models
have a first order phase transition, so there are two coexisting states at
--- a high energy and low energy state.
The standard Metropolis algorithm
tends to get stuck in one of the these states.
Tempering does a Monte Carlo update of the temperature
--- i.e. try to change T,
and do a Metropolis accept/reject depending on 
.
Sweeping slightly above and below 
, moves the configuration
in and out of high and low E
states, thus providing a correct sampling of the configurations at 
.
The repeated heating and cooling is like tempering in metals.