Active Evaluation Acquisition for Efficient LLM Benchmarking

TL;DR

This paper proposes an RL-based method for selecting a subset of prompts in LLM benchmarking, reducing evaluation costs while maintaining accuracy by modeling dependencies among test examples.

Contribution

It introduces a novel reinforcement learning policy that leverages dependencies among test examples to improve evaluation efficiency in large language model benchmarks.

Findings

Significantly reduces the number of prompts needed for evaluation.

Maintains accurate performance estimates with fewer evaluation costs.

Outperforms previous subset selection methods.

Abstract

As large language models (LLMs) become increasingly versatile, numerous large scale benchmarks have been developed to thoroughly assess their capabilities. These benchmarks typically consist of diverse datasets and prompts to evaluate different aspects of LLM performance. However, comprehensive evaluations on hundreds or thousands of prompts incur tremendous costs in terms of computation, money, and time. In this work, we investigate strategies to improve evaluation efficiency by selecting a subset of examples from each benchmark using a learned policy. Our approach models the dependencies across test examples, allowing accurate prediction of the evaluation outcomes for the remaining examples based on the outcomes of the selected ones. Consequently, we only need to acquire the actual evaluation outcomes for the selected subset. We rigorously explore various subset selection policies and introduce a novel RL-based policy that leverages the captured dependencies. Empirical results demonstrate that our approach significantly reduces the number of evaluation prompts required while maintaining accurate performance estimates compared to previous methods.