| 666 | 0 | 5 |
| 下载次数 | 被引频次 | 阅读次数 |
在计算资源不断增强的供给推动下，大语言模型(Large Language Models, LLMs)的参数规模持续扩大，其在自然语言处理中的任务表现也更加卓越。但在面临推理问题，尤其是在常识推理或数学问题上，仍然存在一定的局限性。思维链(Chain of Thought, CoT)技术通过引导模型生成推理步骤，显著提升了其在不同领域问题的解决能力。从训练方式的角度梳理了CoT的理论基础系统和技术演进，对如政务服务、企业数字化等应用场景做了进一步讨论。结合(Artificial Intelligence, AI)的发展趋势，从AI智能化程度的角度论述了CoT在LLMs走向更高认知水平中必不可少的作用，并指出其在当前面临的挑战与亟需解决的技术瓶颈。
Abstract：Driven by the ever-increasing supply of computational resources, the parameter size of Large Language Models(LLMs) continues to expand and their task performance in natural language processing has become more superior. However, there are still limitations when faced with reasoning problems, especially in common-sense reasoning or mathematical problems. Chain of Thought(CoT) significantly improves its ability to solve problems in different domains by guiding the model to generate reasoning steps. In this paper, we not only sort out the theoretical foundation system and technical evolution of CoT from the perspective of training method, but also further discuss application scenarios such as government service and enterprise digitalisation. Finally, in the light of the development trend of Artificial Intelligence(AI), the paper discusses the essential role of CoT in the development of LLMs towards a higher cognitive level from the perspective of the degree of AI, and points out the challenges and technical bottlenecks that need to be solved at the present time.
[1] BROWN T B,MANN B,RYDER N,et al.Language Models are Few-shot Learners[EB/OL].(2020-05-28)[2025-04-25].http://arxiv.org/abs/2005.14165.
[2] WEI J,WANG X Z,SCHUURMANS D,et al.Chain-of-thought Prompting Elicits Reasoning in Large Language Models[EB/OL].(2022-01-28) [2025-04-25].http://arxiv.org/abs/2201.11903.
[3] 舒文韬，李睿潇，孙天祥，等.大型语言模型：原理、实现与发展[J].计算机研究与发展，2024,61(2):351-361.
[4] YAO S Y,YU D,ZHAO J,et al.Tree of Thoughts:Deliberate Problem Solving with Large Language Models[EB/OL].(2023-05-17) [2025-04-25].http://arxiv.org/abs/2305.10601.
[5] ZHANG Z S,ZHANG A,LI M,et al.Multimodal Chain-of-thought Reasoning in Language Models[EB/OL].(2023-02-02) [2025-04-25].http://arxiv.org/abs/2302.00923.
[6] 袁毓林，卢达威.怎样利用语言知识资源进行语义理解和常识推理[J].中文信息学报，2018,32(12):11-23.
[7] WANG X Z,WEI J,SCHUURMANS D,et al.Self-consistency Improves Chain of Thought Reasoning in Language Models[EB/OL].(2022-03-21)[2025-04-25].http://arxiv.org/abs/2203.11171.
[8] WANG Z L,ZHANG H,LI C L,et al.Chain-of-table:Evolving Tables in the Reasoning Chain for Table Understanding[EB/OL].(2024-01-09)[2025-04-25].http://arxiv.org/abs/2401.04398.
[9] 罗焕坤，葛一烽，刘帅.大语言模型在数学推理中的研究进展[J].计算机工程，2024,50(9):1-17.
[10] 黄峻，林飞，杨静，等.生成式AI的大模型提示工程：方法、现状与展望[J].智能科学与技术学报，2024,6(2):115-133.
[11] COBBE K,KOSARAJU V,BAVARIAN M,et al.Training Verifiers to Solve Math Word Problems[EB/OL].(2021-10-27)[2025-04-25].http://arxiv.org/abs/2110.14168.
[12] 陈孟科，边赟，梁云浩，等.基于6W2H的大语言模型思维链提示框架WH-CoT[J].计算机应用，2024,44(增刊2):1-6.
[13] 桑晨扬，马廷淮，谢欣彤，等.基于大语言模型多阶段推理的情绪支持对话生成方法[J].计算机科学与探索，2024,18(11):2925-2939.
[14] LIU J X,WANG Y,DU J W,et al.MedCoT:Medical Chain of Thought via Hierarchical Expert[EB/OL].(2024-12-18)[2025-04-25].http://arxiv.org/abs/2412.13736.
[15] LEE S,OH S,PARK S,et al.FINALE:Finance Domain Instruction-tuning Dataset with High-quality Rationales via Chain-of-thought Prompting[EB/OL].(2025-02-16)[2025-04-25].https://www.semanticscholar.org/paper/FINALE-%3A-Finance-Domain-Instruction-Tuning-Dataset-Lee-Oh/942633d606bfb6ab8654b71b2ee07b05904efd96.
[16] 赵凯琳，靳小龙，王元卓.小样本学习研究综述[J].软件学报，2021,32(2):349-369.
[17] 祝钧桃，姚光乐，张葛祥.深度神经网络的小样本学习综述[J].计算机工程与应用，2021,57(7):22-33.
[18] VINYALS O,BLUNDELL C,LILLICRAP T,et al.Matching Networks for One Shot Learning[EB/OL].(2016-06-13)[2025-04-25].http://arxiv.org/abs/1606.04080.
[19] 颜玉松，周圆，王琮，等.基于预训练大模型的行动方案生成方法[J].计算机科学，2025,52(1):80-86.
[20] 李鑫尧，李晶晶，朱磊，等.资源受限的大模型高效迁移学习算法研究综述[J].计算机学报，2024,47(11):2491-2521.
[21] 张延.迁移学习在大模型中的挑战与突破[J].中国信息界，2024(9):186-189.
[22] 罗焕坤，葛一烽，刘帅.大语言模型在数学推理中的研究进展[J].计算机工程，2024,50(9):1-17.
[23] KOJIMA T,GU S S,REID M,et al.Large Language Models are Zero-shot Reasoners[EB/OL].(2022-05-24)[2025-04-25].http://arxiv.org/abs/2205.11916.
[24] LEE G G,LATIF E,WU X S,et al.Applying Large Language Models and Chain-of-thought for Automatic Scoring[EB/OL].(2023-12-30)[2025-04-25].http://arxiv.org/abs/2312.3748.
[25] WANG L,XU W Y,LAN Y H,et al.Plan-and-solve Prompting:Improving Zero-shot Chain-of-thought Reasoning by Large Language Models[EB/OL].(2023-05-16)[2025-04-25].http://arxiv.org/abs/2305.04091.
[26] WANG X Z,WEI J,SCHUURMANS D,et al.Rationale-augmented Ensembles in Language Models[EB/OL].(2022-07-02)[2025-04-25].http://arxiv.org/abs/2207.00747.
[27] WU Y H,JIANG A Q,LI W D,et al.Autoformalization with Large Language Models[EB/OL].(2022-05-25)[2025-04-25].http://arxiv.org/abs/2205.12615.
[28] ZELIKMAN E,WU Y H,MU J,et al.STaR:Bootstrapping Reasoning with Reasoning[EB/OL].(2022-03-28)[2025-04-26].http://arxiv.org/abs/2203.14465.
[29] GAO L Y,MADAAN A,ZHOU S Y,et al.PAL:Program-aided Language Models[EB/OL].(2022-11-18)[2025-04-26].http://arxiv.org/abs/2211.10435.
[30] ZHANG T H,GE J X,LUO H Y,et al.Natural Language Embedded Programs for Hybrid Language Symbolic Reasoning[EB/OL].(2023-19-19)[2025-04-26].http://arxiv.org/abs/2309.10814.
[31] CHEN W H,MA X G,WANG X Y,et al.Program of Thoughts Prompting:Disentangling Computation from Reasoning for Numerical Reasoning Tasks[EB/OL].(2022-11-22)[2025-04-26].http://arxiv.org/abs/2211.12588.
[32] WU Z Q,XU B D,CUI R C,et al.Rethinking Chain-of-thought from the Perspective of Self-training[EB/OL].(2024-12-14)[2025-04-26].http://arxiv.org/abs/2412.10827.
[33] 崔金满，李冬梅，田萱，等.提示学习研究综述[J].计算机工程与应用，2024,60(23):1-27.
[34] ZHAO R C,LI X X,JOTY S,et al.Verify-and-edit:A Knowledge-enhanced Chain-of-thought Framework[EB/OL].(2023-05-05)[2025-04-26].http://arxiv.org/abs/2305.03268.
[35] WANG K H,DUAN F Y,WANG S R,et al.Knowledge-driven CoT:Exploring Faithful Reasoning in LLMs for Knowledge-intensive Question Answering[EB/OL].(2023-08-25)[2025-04-26].http://arxiv.org/abs/2308.13259.
[36] 游璇钰，廖汉彬.惠州12345热线引入DeepSeek[N].惠州日报，2025-02-18(2).
[37] 刘琼萍，王云丽.湖南首个城市级应用CS-DeepSeek上线[N].长沙晚报，2025-02-15(1).
[38] 王铭俊，黄果，张延惠.新湖南客户端全线产品接入DeepSeek[N].湖南日报，2025-02-16(4).
[39] SAE Intermatinal.Taxonomy and Definitions for Terms Related to On-road Motor Vehicle Automated Driving System[P/OL].(2014-01-16)[2025-04-26].https://www.sae.org/standards/content/j3016_201401/.
[40] CHOWDHERY A,NARANG S,DEVLIN J,et al.PaLM:Scaling Language Modeling with Pathways[EB/OL].(2022-04-05)[2025-04-26].http://arxiv.org/abs/2204.02311.
[41] QIAN S B,SINDHUJAN A,KABRA M,et al.What do Large Language Models Need for Machine Translation Evaluation?[EB/OL].(2024-10-04)[2025-04-26].http://arxiv.org/abs/2410.03278.
[42] 刘安平，金昕，胡国强.人工智能大模型综述及金融应用展望[J].人工智能，2023(2):29-40.
[43] TURPIN M,MICHAEL J,PEREZ E,et al.Language Models Don’t Always Say What They Think:Unfaithful Explanations in Chain-of-thought Prompting[EB/OL].(2023-05-07)[2025-04-26].http://arxiv.org/abs/2305.04388.
[44] SAPAROV A,PANG R Y Z,PADMAKUMAR V,et al.Testing the General Deductive Reasoning Capacity of Large Language Models Using OOD Examples[EB/OL].(2023-05-24)[2025-04-26].http://arxiv.org/abs/2305.15269.
[45] LANHAM T,CHEN A N,RADHAKRISHNAN A,et al.Measuring Faithfulness in Chain-of-thought Reasoning[EB/OL].(2023-07-17)[2025-04-26].http://arxiv.org/abs/2307.13702.
基本信息:
DOI：
中图分类号:TP391.1;TP18
引用信息:
[1]杜家乐,陈曙东,叶亮等.大语言模型中的思维链技术综述[J].无线电通信技术,2025,51(05):877-887.
基金信息:
北京市科技计划项目(Z231100001323004)~~