Correlating and Predicting Human Evaluations of Language Models from Natural Language Processing Benchmarks

Rylan Schaeffer, Punit Singh Koura, Binh Tang, Ranjan Subramanian, Aaditya K Singh, Todor Mihaylov, Prajjwal Bhargava, Lovish Madaan, Niladri S. Chatterji, Vedanuj Goswami, Sergey Edunov, Dieuwke Hupkes, Sanmi Koyejo, Sharan Narang

Keywords: language models, evaluations, human evaluations, benchmarks, NLP benchmarks

TL;DR: We compare human evaluations and academic evaluations of language models against one another

Abstract: The field of natural language processing (NLP) historically evaluated language models using benchmarks with automated metrics. However, the recent advent of highly capable chat language models (LMs) has caused a tectonic shift from NLP benchmarks to human evaluations. The relationship between these two evaluation processes is unclear and underexplored for chat LMs. Broadly, to what extent are human evaluations and NLP benchmarks correlated with one another? How well can computationally inexpensive and automated benchmarks predict expensive and time-intensive human evaluations? Which benchmarks provide predictive signals for human preference for LMs? What role, if any, should benchmarks play in the era of chat LMs? To answer these questions, we conducted a large-scale study of the relationships between human evaluations and benchmarks. We show that benchmarks are broadly highly correlated with human evaluations, and we identify which benchmarks exhibit strong correlations with human evaluations and which do not. Having established that reliable correlations exist, we fit models to predict a language model’s human evaluation scores from its academic evaluation scores and provide evidence that such predictive models can generalize across LM scales.

Primary Area: datasets and benchmarks

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8665