Prompt architecture induces methodological artifacts in large language models.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Linguistics

Background:

• Large language models (LLMs) are increasingly used for various tasks.
• The way prompts are structured, termed 'prompt architecture,' can influence LLM outputs.
• Methodological artifacts, a form of statistical bias, may arise from prompt design.

Purpose of the Study:

• To investigate how prompt architecture influences responses from LLMs (GPT-3, GPT-4, Llama 3.1).
• To identify specific prompt features (order, label, framing, justification) causing bias.
• To test strategies for mitigating these prompt-induced biases.

Main Methods:

• Conducted five large-scale, full-factorial experiments using standard zero-shot similarity evaluation tasks.
• Analyzed the impact of prompt order, label, framing, and justification on LLM responses.
• Tested mitigation strategies including randomization specifications and aggregation across prompts.

Main Results:

• Prompt architecture elements (order, label, framing, justification) robustly affect LLM responses across models.
• LLMs exhibit response-order bias and label bias, moderated by framing and justification.
• Aggregating results from a full factorial design effectively eliminates response-order and label bias.

Conclusions:

• Individual prompts are inherently fallible due to subtle interactions with embedded language data.
• Prompt design characteristics significantly introduce statistical bias in LLM outputs.
• Aggregating across diverse prompts is a key strategy to overcome prompt-induced biases in LLMs.