How Reliable Are Automatic Evaluation Methods for Instruction-Tuned LLMs?

TL;DR

The paper questions the reliability of automatic evaluation metrics for instruction-tuned LLMs and conducts a comprehensive meta-evaluation using Pairwise Accuracy with Tie Calibration to compare Rouge-L, GPT-4 as a judge, and BertScore against human judgments. It finds that GPT-4 aligns closely with humans when gold reference answers are available, but exhibits an overly positive bias without references, especially in free-form tasks. Rouge-L provides a cost-effective alternative for short-answer English tasks, while BertScore shows strong performance for long-form outputs; however, cross-lingual evaluation (e.g., Swedish) reduces alignment for several metrics. The study provides task-sensitive guidelines for applying automatic evaluation in developing and assessing instruction-tuned LLMs and highlights the need for broader language coverage in future work.

Abstract

Work on instruction-tuned Large Language Models (LLMs) has used automatic methods based on text overlap and LLM judgments as cost-effective alternatives to human evaluation. In this paper, we perform a meta-evaluation of such methods and assess their reliability across a broad range of tasks. In evaluating how well automatic methods align with human evaluations, correlation metrics are the most commonly employed method despite their inherent limitations when dealing with ties and different scales. To address these shortcomings, we use Pairwise Accuracy as an alternative to standard correlation measures. We observe that while automatic evaluation methods can approximate human ratings under specific conditions, their validity is highly context-dependent. Specifically, the simple ROUGE-L metric correlates very well with human ratings for short-answer English tasks but is unreliable in free-form generation tasks and cross-lingual scenarios. The effectiveness of the more advanced method of using GPT-4 as a judge diminishes significantly if reference answers are not included in the prompt, which is the scenario where this method has the potential to provide the most value compared to other metrics. Our findings enhance the understanding of how automatic methods should be applied and interpreted when developing and evaluating instruction-tuned LLMs.

Figures (6)

• Figure 1: Human ratings and Rouge"-L scores per task and test language at the bottom and their $\text{PA}$ at the top. Human scores are normalized to a range between 0 and 1. Short-answer tasks are on the left and long-answer ones are on the right.
• Figure 2: Human and GPT-4-gold's ratings per task and test language on the bottom and their $\text{PA}$ on top. Ratings are normalized to a range of 0 to 1. Short-answer tasks are on left and long-answer ones on right.
• Figure 3: Pairwise accuracy between GPT-4-gold and BertScore for long-answer English tasks.
• Figure 4: Human and GPT-4-no-gold's ratings per task and test language on the bottom and their $\text{PA}$ on top. Ratings are normalized to a range of 0 to 1. Short-answer tasks are on left and long-answer ones on right.
• Figure A1: Human and BertScore ratings per task and test language on the bottom and their $\text{PA}$ on top. The human ratings are normalized to a range of 0 to 1. Short-answer tasks are on left and long-answer ones on right.