Do you have data for other benchmarks? +7% for HLE isn't nothing but it'd be more compelling if you could show you're consistently doing better with your method across more domains (especially coding, which seems like the primary use-case these days).