Why not use traditional metrics like BLEU and ROUGE?

Traditional evaluation metrics like BLEU and ROUGE have some value, but they also have major limitations:

• They require a reference to compare against: While you may have such ground truth data in your development dataset, you will never have this in production.

• Traditional metrics will not offer any reasoning capabilities Most developers are now using LLMs for much more complex use cases than can be evaluated by traditional methods.

In contrast, LLM evaluators:

• Can perform complex and nuanced tasks that include reasoning capabilities
• Come a lot closer to human-in-the-loop level of accuracy

Intuitively, this makes sense. The best tool we have for handling reasoning tasks on large pieces of text are LLMs. So why would you use anything else for evals?