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

2026-05-21 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

An empirical study reveals that instruction embedding models, despite their prevalence, are inadequately evaluated using single-prompt methods. The research, spanning 6 embedding models, 11 datasets, and 15 task-specific prompts per dataset (totaling 990 evaluations), demonstrates significant sensitivity to instruction phrasing. Reported scores frequently misrepresent the true performance distribution, with default prompts either systematically understating or overstating model capabilities. Crucially, the study found that leaderboard rankings are not robust, as strategic prompt selection can elevate any tested model to first place. These findings underscore the insufficiency of single-prompt evaluation and advocate for benchmarks to integrate prompt robustness, either through multi-prompt evaluation or by reporting sensitivity alongside point estimates.

Key takeaway

For machine learning engineers evaluating or deploying instruction-tuned embedding models, you must recognize that single-prompt evaluations are insufficient and can lead to misleading performance assessments. Your model's reported scores and leaderboard positions may not reflect its true robustness across varied instructions. To ensure reliable performance, incorporate multi-prompt testing into your evaluation pipelines and consider the range of scores, not just a single point estimate, when making deployment decisions.

Key insights

Instruction embedding model evaluations are unreliable due to high sensitivity to prompt phrasing, invalidating single-prompt benchmarks.

Principles

• Instruction sensitivity significantly impacts model performance.
• Single-prompt evaluation misrepresents model capabilities.
• Leaderboard rankings are not robust to prompt selection.

In practice

• Evaluate models using multiple diverse prompts.
• Report prompt sensitivity alongside performance scores.

Topics

• Instruction Embedding Models
• Prompt Sensitivity
• Model Evaluation
• NLP Benchmarking
• Leaderboard Robustness

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Supervision versus Demonstration-Based In-Context Learning for Multiword Expression Classification — Prompt sensitivity and demonstration quality critically impact LLM performance in metalinguistic classification tasks.
• Embedding Model Selection: 10 Scenario-Based Questions & Solutions — The most reliable embedding model selection combines public benchmark filtering with specific business data experimentation.
• Solving the Embedding Step in NLP Projects — Effective embeddings are crucial for NLP model generalization and reliable downstream task performance.
• Temporal Stability and Few-Shot Prompting in Math Task Assessment — Prompt engineering offers more reliable AI performance improvements on specialized tasks than passive model version updates.
• Correcting Prompt Dependence in LLM Benchmarks: A Bayesian Hierarchical Model with Embedding-Space Clustering — Reliable LLM vulnerability evaluation requires a Bayesian hierarchical model with embedding-space clustering to quantify uncertainty and mitigate prompt bias.
• Spurious Prompts: Can Irrelevant Prompts Steer Large Language Models? — Large Language Models exhibit surprising sensitivity to semantically irrelevant "spurious prompts," which can systematically steer their behavior.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.