Year: 2025
Type: article
Abstract: This research investigates the impact of various prompt engineering techniques on the length, cost, complexity, and accuracy of responses from large language models (LLMs). By comparing direct prompting with zero-shot, few-shot, and chain-of-thought (CoT) methods on tasks like GSM8K and creative writing, I analyze the trade-offs between token usage and response quality. Results show that while zero-shot CoT prompting is highly effective and cost-efficient, other methods like Least-to-Most and Tree-of-Thought add significant length and complexity without proportional accuracy gains. Additionally, I discuss the financial implications, finding that GPT-4’s unique pricing structure narrows the cost difference between manual/few-shot and zero-shot methods. Complexity analysis reveals that more intricate prompts often lead to convoluted outputs, challenging human review and implementation. Our findings guide the selection of prompt engineering strategies to optimize both performance and resource utilization
Author: Milind Cherukuri -
Institution: Caris Life Sciences (United States)
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4410031540