The Base-Rate Effect on LLM Benchmark Performance: Disambiguating Test-Taking Strategies from Benchmark Performance

TL;DR

This paper investigates how base-rate probabilities influence large language model benchmark performance, revealing that test-taking strategies can confound true task ability measurements, and proposes a new task to better distinguish these factors.

Contribution

The study identifies the impact of base-rate effects on LLM benchmark results and introduces the Nvr-X-MMLU task to separate test-taking strategies from genuine task performance.

Findings

Base-rate differences significantly affect LLM test performance.

Counterfactual prompting mitigates the base-rate effect.

The Nvr-X-MMLU task disambiguates test-taking ability from task performance.

Abstract

Cloze testing is a common method for measuring the behavior of large language models on a number of benchmark tasks. Using the MMLU dataset, we show that the base-rate probability (BRP) differences across answer tokens are significant and affect task performance ie. guess A if uncertain. We find that counterfactual prompting does sufficiently mitigate the BRP effect. The BRP effect is found to have a similar effect to test taking strategies employed by humans leading to the conflation of task performance and test-taking ability. We propose the Nvr-X-MMLU task, a variation of MMLU, which helps to disambiguate test-taking ability from task performance and reports the latter.