<?xml version="1.1" encoding="utf-8"?>
<article xsi:noNamespaceSchemaLocation="http://jats.nlm.nih.gov/publishing/1.1/xsd/JATS-journalpublishing1-mathml3.xsd" dtd-version="1.1" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"><front><journal-meta><journal-id journal-id-type="publisher-id">TACS</journal-id><journal-title-group><journal-title>Technology and Application of Computer Science</journal-title></journal-title-group><issn>2998-8926</issn><eissn>2998-8934</eissn><publisher><publisher-name>Art and Technology</publisher-name></publisher></journal-meta><article-meta><article-id pub-id-type="doi">10.61369/TACS.</article-id><article-categories><subj-group subj-group-type="heading"><subject>Article</subject></subj-group></article-categories><title>生成式人工智能的使用策略: 基于任务复杂度与提示策略的交互效应实验研究</title><url>https://artdesignp.com/journal/TACS/3/10/10.61369/TACS.</url><author>李泽阳,孙伯通,吴佳涛,余跃涛,唐佳成</author><pub-date pub-type="publication-year"><year>2026</year></pub-date><volume>3</volume><issue>10</issue><history><date date-type="pub"><published-time>2026-05-28</published-time></date></history><abstract>生成式人工智能（GenAI）在理工科领域的应用日益广泛，但用户交互策略对输出质量的影响机制尚缺乏系统的实证研究。本文以代码生成与数学推理为实验任务，考察任务复杂度、提示策略及二者交互效应对生成准确率的影响。实验采用2（任务复杂度：高/ 低）&amp;times;3（提示策略：零样本/ 少样本/ 思维链）被试内设计，共收集1200次生成结果。研究发现：任务复杂度主效应显著（F(1,39)=28.34，p&amp;lt;0.001，&amp;eta;2=0.42），提示策略主效应显著（F(2,78)=15.67，p&amp;lt;0.001，&amp;eta;2=0.29），二者交互效应显著（F(2,78)=9.23，p&amp;lt;0.001，&amp;eta;2=0.19）。简单效应分析表明，在复杂任务中思维链策略显著优于零样本策略（MD=0.31，p&amp;lt;0.001）和少样本策略（MD=0.18，p=0.003），效应量Cohen'sd=0.74；在简单任务中三种策略差异不显著（F(2,78)=1.34，p=0.27）。基于实验结果，提出&amp;ldquo;复杂度适配&amp;rdquo;的三级使用策略框架，为不同场景下生成式人工智能的高效调用提供实证依据与操作指南。</abstract><keywords>生成式人工智能；使用策略；提示工程；任务复杂度；交互效应</keywords></article-meta></front><body/><back><ref-list><ref id="B1" content-type="article"><label>1</label><element-citation publication-type="journal"><p>[1] Brown T, Mann B, Ryder N, et al. Language models are few-shot learners[J]. Advances in Neural Information Processing Systems, 2020, 33: 1877-1901.[2] Chen M, Tworek J, Jun H, et al. Evaluating large language models trained on code[J]. arXiv preprint arXiv:2107.03374, 2021.[3] Wei J, Wang X, Schuurmans D, et al. Chain-of-thought prompting elicits reasoning in large language models[J]. Advances in Neural Information Processing Systems,2022, 35: 24824-24837.[4] Wang X, Wei J, Schuurmans D, et al. Self-consistency improves chain of thought reasoning in language models[J]. arXiv preprint arXiv:2203.11171, 2022.[5] Yao S, Yu D, Zhao J, et al. Tree of thoughts: Deliberate problem solving with large language models[J]. Advances in Neural Information Processing Systems, 2023, 36:11809-11822.[6] Sweller J. Cognitive load theory, learning difficulty, and instructional design[J]. Learning and Instruction, 1994, 4(4): 295-312.[7] Hendrycks D, Burns C, Kadavath S, et al. Measuring mathematical problem solving with the MATH dataset[J]. NeurIPS 2021 Datasets and Benchmarks Track, 2021.[8] Cobbe K, Kosaraju V, Bavarian M, et al. Training verifiers to solve math word problems[J]. arXiv preprint arXiv:2110.14168, 2021.[9] Kahneman D. Thinking, fast and slow[M]. New York: Farrar, Straus and Giroux, 2011.[10] 罗会钧, 唐浩瑜. 生成式人工智能教育领域应用的风险及应对[J]. 中南大学学报( 社会科学版),2025,31(5):112-124.[11] 曹昌沐. 生成式人工智能在高校教师工作中的双刃剑效应[J]. 龙岩学院学报,2026,44(3):116-122.[12] 国家互联网信息办公室. 生成式人工智能服务管理暂行办法[S].2023.</p><pub-id pub-id-type="doi"/></element-citation></ref></ref-list></back></article>
