Exploring the Prompt Sensitivity in LLM: A Methodological Framework and Empirical Analysis

Anonymous

Abstract: This paper examines LLM sensitivity to natural language prompts and proposes methods to enhance robustness. Despite their versatility, LLMs show performance volatility with prompt changes. We introduce \lib, which featuring diverse, semantically consistent prompts mimicking human expression patterns for multiple LLM evaluations. Experiments with \lib confirm that model size or baseline metrics do not correlate with prompt sensitivity, and subtle perturbations can impact results. We find in-context examples and diverse training instructions improve LLM resilience against different question forms. We believe this work will serve as a helpful tool in studying LLM robustness under human-like expressions.

Paper Type: long

Research Area: Resources and Evaluation

Contribution Types: Model analysis & interpretability

Languages Studied: English