Fixed-exit approach that trains models to work at specific predetermined depths, leaving performance poor at intermediate depths.