Love the benchmarks. Is better to use single LLM for performance or would always advise to add a self reflection step