Optimo's Mixture-of-Prompts Hits 3.97x Speedup on Human Code and 13.51x on LLM-Generated Code

arXiv 2607.23665·medium signalXBlueskyLinkedInCopy link

Optimo (arXiv 2607.23665) attacks a real gap in LLM code optimization: existing approaches fail to identify what is actually worth optimizing and do single-level rewrites. It uses differential profiling to locate time-critical structures, then routes them to specialized optimization strategies — MoE-style experts, but over prompts — operating across four abstraction levels from algorithmic improvements down to API-usage tweaks. On the COFFE and Effibench benchmarks it reaches up to 57.48% opt% (programs that stay correct and get at least 10% faster) with up to 3.97x speedup on human-written code, beating the strongest baseline by up to 96.51% in opt%; on LLM-generated code it reaches 42.42% opt% and up to 13.51x speedup, quantifying how much slack sits in agent-written code.