HREF: Human Response-Guided Evaluation of Instruction Following in Language Models

TL;DR

This paper introduces HREF, a new human response-guided evaluation benchmark for instruction-following in language models, improving reliability and reducing bias compared to traditional LLM-based assessments.

Contribution

It develops a novel evaluation benchmark using human-written responses, demonstrating enhanced agreement with human judgments and providing a comprehensive, bias-reduced assessment framework.

Findings

Human responses improve evaluation reliability by up to 3.2%

HREF covers 11 task categories with 4,258 samples

Evaluation setup is free from contamination and emphasizes individual task performance

Abstract

Evaluating the capability of Large Language Models (LLMs) in following instructions has heavily relied on a powerful LLM as the judge, introducing unresolved biases that deviate the judgments from human judges. In this work, we reevaluate various choices for automatic evaluation on a wide range of instruction-following tasks. We experiment with methods that leverage human-written responses and observe that they enhance the reliability of automatic evaluations across a wide range of tasks, resulting in up to a 3.2% improvement in agreement with human judges. We also discovered that human-written responses offer an orthogonal perspective to model-generated responses in following instructions and should be used as an additional context when comparing model responses. Based on these observations, we develop a new evaluation benchmark, Human Response-Guided Evaluation of Instruction Following (HREF), comprising 4,258 samples across 11 task categories with a composite evaluation setup, employing a composite evaluation setup that selects the most reliable method for each category. In addition to providing reliable evaluation, HREF emphasizes individual task performance and is free from contamination. Finally, we study the impact of key design choices in HREF, including the size of the evaluation set, the judge model, the baseline model, and the prompt template. We host a live leaderboard that evaluates LLMs on the private evaluation set of HREF.