One prompt is not enough: Instruction Sensitivity Undermines Embedding Model Evaluation

TL;DR

This paper demonstrates that evaluating instruction embedding models with a single prompt is unreliable due to high sensitivity to prompt phrasing, affecting performance scores and model rankings.

Contribution

The study provides an extensive empirical analysis of prompt sensitivity across multiple models, datasets, and prompts, highlighting the need for more robust evaluation methods.

Findings

Single-prompt evaluation misrepresents true performance.

Default prompts can unfairly favor or penalize models.

Leaderboard rankings are not stable across different prompts.

Abstract

Instruction embedding models have become common among state-of-the-art models, however are evaluated using a single prompt per task. The single-point evaluation ignores a main problem of the instruction-based approach namely: sensitivity to the phrasing of the instruction. We present an empirical study of prompt sensitivity across 6 embedding models, 11 datasets, and 15 task-specific prompts per dataset, a total of 990. We show that reported scores misrepresent the distribution of scores over plausible prompts. The default prompt can both systematically understate or overstate performance. Furthermore, we show that the leaderboard ranking is not robust to prompt selection: by choosing prompts favorably, any model in our study can be promoted to first place. Our findings suggest that single-prompt evaluation is insufficient for instruction-tuned embedding models and that benchmarks should incorporate prompt robustness, either by evaluating over multiple prompts or by reporting sensitivity alongside point estimates.