LLM evaluations are very sensitive to the details of the prompt's structure. This post shows how using structured generation reduces the results' variance and t

LLM evaluations are very sensitive to the details of the prompt's structure. This post shows how using structured generation reduces the results' variance and the ranking shifts.