We are increasingly seeing LLMs being used to create challenging benchmarks that are then used for evaluating LLMs. Is this a valid approach to evaluation construction? Do we loose anything in this process?

We examine both the 𝑣𝑎𝑙𝑖𝑑𝑖𝑡𝑦 and 𝑑𝑖𝑓𝑓𝑖𝑐𝑢𝑙𝑡𝑦 of LLM-generated versions of two high-quality reading comprehension datasets: CondaQA & DROP. We find that validity is not an issue. We are able to get LLMs to generate instances that are highly valid.

We perform a human study and even find that LLM-generated data is preferred! We ask NLP researchers to act as dataset creators and gather preferences between synthetic and human-authored data.

But are these instances similarly difficult? We explore the difficulty of synthetic benchmarks by comparing performance on synthetic & human-written data across a suite of models. We find that performance is consistently higher on generated versions of the datasets.

Key takeaways: - While LLM generated evals may be 𝑣𝑎𝑙𝑖𝑑, as a whole they lose crucial aspects in complexity. - LLMs are promising where complexity is less critical, but human annotators are vital for benchmarks assessing real-world generalization & nuanced scenarios.

We hope that our work will inspire future research into: - Can further prompt review improve the difficulty of synthetic data? - What other axes (representativeness, diversity) are affected when using LLMs to generate benchmarks?

More results and analysis can be found in the paper. We welcome any discussion, thanks for reading!! (Full Link: https://t.co/35AshcUfT2)