100 instances is all you need: predicting LLM success by testing on a few instances

ICLR 2025 Conference Submission11903 Authors

Keywords: AI evaluation, Large Language Models, Benchmarking

Abstract: Predicting if LLMs will succeed on individual task instances (i.e., prompts) is essential to ensure their reliability in high-stakes applications. To do so, we can evaluate a LLM on a set of instances and train an "assessor" to predict its performance. However, this requires evaluating each new LLM on sufficiently many instances. In this work, we build a "generic assessor" predicting the performance of any LLM on an instance by using the LLM's performance on a small set of reference instances and the features of the considered instance. In practice, we make use of existing evaluation results to extract the representative instances and train the assessor. Thus, the performance of a new LLM can be predicted by only testing it on the reference instances, leveraging the information contained in other LLMs' evaluations. We conduct empirical studies on HELM-Lite and KindsOfReasoning, a new collection of existing reasoning datasets that we introduce, where we evaluate all instruction-fine-tuned OpenAI models until $\texttt{gpt4-0125-preview}$. We find that a few instances (around 100) are enough to achieve predictive power comparable to the LLM-specific assessors trained on the complete set of several thousand instances. Interestingly, randomly selecting the reference instances performs comparably to the advanced selection methods we tested. Finally, we identify a sharp drop in the predictive power of the generic and specific assessors in out-of-distribution scenarios, suggesting that the inherent predictability of LLMs is low.

Primary Area: foundation or frontier models, including LLMs

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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 11903