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

TL;DR

This work reexamines automatic evaluation methods for instruction following in large language models by incorporating human written responses as references. It demonstrates that human references can boost agreement with human judgments across tasks and reveal that these references provide complementary information to model-generated outputs. The authors introduce HREF, a 4,258-prompt benchmark spanning 11 task categories with a composite evaluation setup that selects the best method per category, and they provide development and private evaluation sets to prevent leakage. They show that an open-weight judge like Llama-3.1-70B-Instruct achieves strong human agreement, and that a per-category composite approach yields robust, task-specific evaluation while preserving reproducibility and privacy. Overall, HREF advances reliable, task-centric evaluation of instruction following with an emphasis on human references and a private, contaminant-free testing regime that scales with model progress.

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.

Figures (15)

• Figure 1: An overview of our composite method leverage the human-written response to judge between two responses given an instruction. The example and the prompt shown in the figure are not exact. See details of these methods in Section \ref{['subsec:methods']}.
• Figure 2: The distribution of length difference between sampled model responses and the base line model responses. The distribution is symmetrical.
• Figure 3: Human Agreement Rate using model-generated v.s. human-written responses. Human response outperforms model response for LLM-based evaluation methods but underperforms for embedding-based evaluation methods.
• Figure 4: Judges Preference Between Model-Generated Responses v.s. Human-Written Responses. Model-generated responses are in great favor of all the judges.
• Figure 5: Task Categorical Distribution of the three subsets in HREF. Left: evaluation set; Middle: development set; Right: human agreement set.