SELF-ADAPTIVE PROMPT ENGINEERING FOR COST-EFFICIENT LANGUAGE MODELS: SWITCHING BETWEEN CHAIN-OF-THOUGHT AND DIRECT ANSWER

초록

Large language models (LLMs) have shown remarkable capabilities across various tasks, but their performance often depends heavily on how prompts are crafted. As prompt engineering becomes increasingly important, finding efficient and effective prompting strategies is crucial. We propose Self-Adaptive Prompt Engineering (SAPE), a novel prompting method that adaptively selects between direct answering and step-by-step reasoning based on task complexity. SAPE aims to reduce unnecessary verbosity while maintaining high accuracy. We compare SAPE with two standard prompting strategies, Direct Answer (DA) and Chain of Thought (CoT), across three instruction-tuned language models: Llama-3.2-3B-Instruct, Phi-4-mini-Instruct, and Mistral-7B-Instruct-v0.3. Experimental results show that SAPE achieves accuracy comparable to CoT while significantly reducing token usage. We also analyze how each model responds to different prompts, highlighting variations in instruction adherence and response length. © 2026, International. All rights reserved.

키워드