Optimization before Evaluation: Evaluation with Unoptimised Prompts Can be Misleading

TL;DR

This paper demonstrates that prompt optimization significantly influences large language model evaluation results, emphasizing the need for model-specific prompt tuning during assessments.

Contribution

It reveals the impact of prompt optimization on model rankings and advocates for per-model prompt tuning in evaluation frameworks.

Findings

Prompt optimization alters model rankings substantially.

Using unoptimized prompts can lead to misleading evaluation results.

Practitioners should perform prompt optimization for accurate model comparison.

Abstract

Current Large Language Model (LLM) evaluation frameworks utilize the same static prompt template across all models under evaluation. This differs from the common industry practice of using prompt optimization (PO) techniques to optimize the prompt for each model to maximize application performance. In this paper, we investigate the effect of PO towards LLM evaluations. Our results on public academic and internal industry benchmarks show that PO greatly affects the final ranking of models. This highlights the importance of practitioners performing PO per model when conducting evaluations to choose the best model for a given task.