“Our approach achieves a 3.0% improvement in reducing instruction counts over the compiler, outperforming two state-of-the-art baselines that require thousands of compilations.”
A 7B-parameter model trained from scratch beats LLVM’s code-size passes by 3.0%. It does not optimize anything itself. It picks a list of compiler flags and LLVM does the work. Seven billion parameters to choose flags is a heavy way to save 3%.