Optimo finds 13.51x more speed in LLM-generated code

Optimo finds 13.51x more speed in LLM-generated code than in human code, per a paper posted to arXiv. That's the slack sitting in code agents write: on LLM-generated code alone, Optimo hits a 42.42% optimization rate, fixes that run at least 10% faster.

The approach is differential profiling: it compares runs to isolate the specific structures burning time. Each one then gets routed to a matching strategy, a kind of mixture-of-experts built from prompts instead of trained weights. That routing runs across four abstraction levels, from algorithm choice down to API usage.

On the COFFE and Effibench benchmarks, the optimization rate reached 57.48%, correct code that ran at least 10% faster. That beat the strongest baseline tested against it by up to 96.51% on the same measure.

The real story is the 3.97x versus 13.51x split: LLM-generated code has more than three times the untapped speed of human-written code. That gap says more about what these agents are optimized to do, which is pass checks, than about the optimizer that found it. Worth watching whether Optimo, or something built the same way, becomes a standard pass after code generation instead of a one-off research benchmark.