面向闭源大语言模型的增强研究综述

doi:10.3778/j.issn.1673-9418.2407021

摘要/Abstract

摘要： 随着大语言模型在自然语言处理领域的快速发展，以GPT系列为代表的闭源大语言模型的性能增强成为一个挑战。由于无法访问模型内部的参数权重，传统的训练方法，如微调技术，难以应用于闭源大语言模型，这使得在这些模型上进一步优化变得困难。同时，闭源大语言模型已经广泛应用于下游实际任务，因此研究如何增强闭源大语言模型的性能具有重要意义。聚焦于闭源大语言模型的增强研究，对提示工程（prompt engineering）、检索增强生成（retrieval augmented generation）、智能体（agent）三种技术进行了分析，并针对不同方法的技术特性和模块架构进行了进一步细分，详细介绍了每种技术的核心思想、主要方法及其应用效果，研究了不同增强方法在推理能力、生成可信度、任务适应性等方面的优越性和局限性。讨论了这三种技术的组合应用方法，结合具体案例，强调了组合技术在增强闭源大语言模型性能方面的巨大潜力。总结了现有技术的研究现状和存在的问题，对未来闭源大语言模型增强技术的发展进行了展望。

关键词: 闭源模型, 大语言模型, 提示工程, 检索增强生成, 智能体

Abstract: With the rapid development of large language models in the field of natural language processing, performance enhancement of closed-source large language models represented by the GPT family has become a challenge. Due to the inaccessibility of parameter weights inside the models, traditional training methods, such as fine-tuning techniques, are difficult to be applied to closed-source large language models, which makes it difficult for further optimization on these models. Meanwhile, closed-source large language models have been widely used in downstream real-world tasks, and thus it is important to investigate how to enhance the performance of closed-source large language models. This paper focuses on the enhancement of closed-source large language models, analyzes three techniques, namely prompt engineering, retrieval augmented generation, and agent, and further subdivides them according to the technical characteristics and modular architectures of the different methods. The core idea, main method and application effect of each technology are introduced in detail, and the superiority and limitation of different augmentation methods in terms of reasoning ability, generation credibility and task adaptability are studied. In addition, this paper also discusses the combined application of these three techniques, combining with specific cases to emphasize the great potential of the combined techniques in enhancing model performance. Finally, this paper summarizes the research status and problems of the existing techniques, and looks forward to the future development of enhancement techniques for closed-source large language models.

Key words: closed-source model, large language model, prompt engineering, retrieval augmented generation, agent

刘华玲, 张子龙, 彭宏帅. 面向闭源大语言模型的增强研究综述[J]. 计算机科学与探索, 2025, 19(5): 1141-1156.

LIU Hualing, ZHANG Zilong, PENG Hongshuai. Review of Enhancement Research for Closed-Source Large Language Model[J]. Journal of Frontiers of Computer Science and Technology, 2025, 19(5): 1141-1156.

参考文献

    [1] BROWN T, MANN B, RYDER N, et al. Language models are few-shot learners[C]//Advances in Neural Information Processing Systems 33, 2020: 1877-1901.
    [2] VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[C]//Advances in Neural Information Processing Systems 30, 2017: 5998-6008.
    [3] GOYAL T, LI J J, DURRETT G. News summarization and evaluation in the era of GPT-3[EB/OL]. [2024-04-19]. https://arxiv.org/abs/2209.12356.
    [4] LE SCAO T, FAN A, AKIKI C, et al. BLOOM: a 176B-parameter open-access multilingual language model[EB/OL]. [2024-04-19]. https://arxiv.org/abs/2211.05100.
    [5] ACHIAM J, ADLER S, AGARWAL S, et al. GPT-4 technical report[EB/OL]. [2024-04-19]. https://arxiv.org/abs/2303.08774.
    [6] SUN Y, WANG S, FENG S, et al. Ernie 3.0: large-scale knowledge enhanced pre-training for language understanding and generation[EB/OL]. [2024-04-19]. https://arxiv.org/ abs/2107.02137.
    [7] COLIN R, NOAM S, ADAM R, et al. Exploring the limits of transfer learning with a unified text-to-text transformer[J]. The Journal of Machine Learning Research, 2020, 21(1): 5485-5551.
    [8] HUANG L, YU W J, MA W T, et al. A survey on hallucination in large language models: principles, taxonomy, challenges, and open questions[EB/OL]. [2024-04-19]. https://arxiv.org/abs/2311.05232.
    [9] SPIRLING A. Why open-source generative AI models are an ethical way forward for science[J]. Nature, 2023, 616(7957): 413.
[10] WEI J, BOSMA M, ZHAO V Y, et al. Finetuned language models are zero-shot learners[EB/OL]. [2024-04-19]. https://arxiv.org/abs/2109.01652.
[11] KAPLAN J, MCCANDLISH S, HENIGHAN T, et al. Scaling laws for neural language models[EB/OL]. [2024-04-19]. https://arxiv.org/abs/2001.08361.
[12] MIN S, LYU X X, HOLTZMAN A, et al. Rethinking the role of demonstrations: what makes in-context learning work?[EB/OL]. [2024-04-19]. https://arxiv.org/abs/2202.12837.
[13] ZHAO T, WALLACE E, FENG S, et al. Calibrate before use: improving few-shot performance of language models[C]//Proceedings of the 38th International Conference on Machine Learning, 2021: 12697-12706.
[14] WEBSON A, PAVLICK E. Do prompt-based models really understand the meaning of their prompts? [EB/OL]. [2024-04-19]. https://arxiv.org/abs/2109.01247.
[15] WEI J, WANG X Z, SCHUURMANS D, et al. Chain-of-thought prompting elicits reasoning in large language models[EB/OL]. [2024-04-19]. https://arxiv.org/abs/2201.11903.
[16] KOJIMA T, GU S S, REID M, et al. Large language models are zero-shot reasoners[C]//Advances in Neural Information Processing Systems 35, 2022: 22199-22213.
[17] WANG X Z, WEI J, SCHUURMANS D, et al. Self-consistency improves chain of thought reasoning in language models[EB/OL]. [2024-05-16]. https://arxiv.org/abs/2203. 11171.
[18] ZHOU D, SCH?RLI N, HOU L, et al. Least-to-most prompting enables complex reasoning in large language models[EB/OL]. [2024-05-16]. https://arxiv.org/abs/2205.10625.
[19] ZHANG Z S, ZHANG A, LI M, et al. Automatic chain of thought prompting in large language models[EB/OL]. [2024-05-16]. https://arxiv.org/abs/2210.03493.
[20] LIU H M, TENG Z Y, CUI L Y, et al. LogiCoT: logical chain-of-thought instruction tuning[C]//Findings of the Association for Computational Linguistics: EMNLP 2023. Stroudsburg: ACL, 2023.
[21] HU H, LU H, ZHANG H, et al. Chain-of-symbol prompting for spatial reasoning in large language models[C]//Proceedings of the 1st Conference on Language Modeling, 2024.
[22] ZHOU Y, GENG X, SHEN T, et al. Thread of thought unraveling chaotic contexts[EB/OL]. [2024-04-19]. https://arxiv. org/abs/2311.08734.
[23] LU X C, LIU Z M, GUO S, et al. DRPT: disentangled and recurrent prompt tuning for compositional zero-shot learning[EB/OL]. [2024-05-16]. https://arxiv.org/abs/2305.01239.
[24] ZHAO X Y, MA C Q. Prompt recursive search: a living framework with adaptive growth in LLM auto-prompting[EB/OL]. [2024-09-23]. https://arxiv.org/abs/2408.01423.
[25] DENG Y H, ZHANG W T, CHEN Z X, et al. Rephrase and respond: let large language models ask better questions for themselves[EB/OL]. [2024-05-16]. https://arxiv. org/abs/2311.04205.
[26] YAO S, YU D, ZHAO J, et al. Tree of thoughts: deliberate problem solving with large language models[C]//Advances in Neural Information Processing Systems 36, 2023.
[27] LIU P F, YUAN W Z, FU J L, et al. Pre-train, prompt, and predict: a systematic survey of prompting methods in natural language processing[J]. ACM Computing Surveys, 2023, 55(9): 1-35.
[28] LONG J Y. Large language model guided tree-of-thought[EB/OL]. [2024-05-16]. https://arxiv.org/abs/2305.08291.
[29] CHIA Y K, CHEN G Z, TUAN L A, et al. Contrastive chain-of-thought prompting[EB/OL]. [2024-05-16]. https://arxiv.org/abs/2311.09277.
[30] ZHENG W, SHARAN S P, JAISWAL A K, et al. Outline, then details: syntactically guided coarse-to-fine code generation[C]//Proceedings of the 40th International Conference on Machine Learning, 2023: 42403-42419.
[31] HUANG Q, ZOU Z, XING Z, et al. AI chain on large language model for unsupervised control flow graph generation for statically-typed partial code[EB/OL]. [2024-05-16]. https://arxiv.org/abs/2306.00757.
[32] BESTA M, BLACH N, KUBICEK A, et al. Graph of thoughts: solving elaborate problems with large language models[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2024, 38(16): 17682-17690.
[33] WEI A, HAGHTALAB N, STEINHARDT J. Jailbroken: how does LLM safety training fail?[EB/OL]. [2024-05-16]. https://arxiv.org/abs/2307.02483.
[34] LIU Y, DENG G L, XU Z Z, et al. Jailbreaking ChatGPT via prompt engineering: an empirical study[EB/OL]. [2024-05-16]. https://arxiv.org/abs/2305.13860.
[35] LEWIS P, PEREZ E, PIKTUS A, et al. Retrieval-augmented generation for knowledge-intensive NLP tasks[C]//Advances in Neural Information Processing Systems 33, 2020: 9459-9474.
[36] CHASE H. LangChain[EB/OL]. [2024-12-11]. https://github.com/langchain-ai/langchain.
[37] KHATTAB O, SANTHANAM K, LI X L, et al. Demonstrate-search-predict: composing retrieval and language models for knowledge-intensive NLP[EB/OL]. [2024-05-24]. https://arxiv.org/abs/2212.14024.
[38] PIETSCH M, M?LLER T, KOSTIC B, et al. Haystack: the end-to-end NLP framework for pragmatic builders[EB/OL]. [2024-12-11]. https://github.com/deepset-ai/haystack.
[39] YANG Z L, QI P, ZHANG S Z, et al. HotpotQA: a dataset for diverse, explainable multi-hop question answering[EB/OL]. [2024-05-24]. https://arxiv.org/abs/1809.09600.
[40] ZHANG X, WU J, HE Z Y, et al. Medical exam question answering with large-scale reading comprehension[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2018, 32(1).
[41] ZHA L Y, ZHOU J L, LI L Y, et al. TableGPT: towards unifying tables, nature language and commands into one GPT[EB/OL]. [2024-05-24]. https://arxiv.org/abs/2307.08674.
[42] HE X X, TIAN Y J, SUN Y F, et al. G-retriever: retrieval-augmented generation for textual graph understanding and question answering[EB/OL]. [2024-05-24]. https://arxiv. org/abs/2402.07630.
[43] WANG X T, YANG Q W, QIU Y T, et al. KnowledGPT: enhancing large language models with retrieval and storage access on knowledge bases[EB/OL]. [2024-05-24]. https://arxiv.org/abs/2308.11761.
[44] ZHONG Z J, LIU H W, CUI X Y, et al. Mix-of-granularity: optimize the chunking granularity for retrieval-augmented generation[EB/OL]. [2024-08-20]. https://arxiv.org/abs/2406. 00456.
[45] CHEN T, WANG H W, CHEN S H, et al. Dense X retrieval: what retrieval granularity should we use?[EB/OL]. [2024-05-24]. https://arxiv.org/abs/2312.06648.
[46] LI D, LI Z, YANG Y, et al. Knowledge graph-enhanced large language model for domain-specific question answering systems[EB/OL]. [2024-10-22]. https://doi.org/10.36227/techrxiv.172963127.77889256/v1.
[47] LIU J. LlamaIndex[EB/OL]. [2024-05-24]. https://github. com/run-llama/llama_index.
[48] DAI Z Y, ZHAO V Y, MA J, et al. Promptagator: few-shot dense retrieval from 8 examples[EB/OL]. [2024-05-24]. https://arxiv.org/abs/2209.11755.
[49] ZHANG P T, XIAO S T, LIU Z, et al. Retrieve anything to augment large language models[EB/OL]. [2024-05-24]. https://arxiv.org/abs/2310.07554.
[50] WANG L, YANG N, WEI F R. Query2doc: query expansion with large language models[EB/OL]. [2024-05-24]. https://arxiv.org/abs/2303.07678.
[51] SHAO Z H, GONG Y Y, SHEN Y L, et al. Enhancing retrieval-augmented large language models with iterative retrieval-generation synergy[C]//Findings of the Association for Computational Linguistics: EMNLP 2023. Stroudsburg: ACL, 2023: 9248-9274.
[52] GAO L Y, MA X G, LIN J, et al. Precise zero-shot dense retrieval without relevance labels[C]//Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics. Stroudsburg: ACL, 2023: 1762-1777.
[53] MA X B, GONG Y Y, HE P C, et al. Query rewriting in retrieval-augmented large language models[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Stroudsburg: ACL, 2023: 5303-5315.
[54] LIU J N, JIN J J, WANG Z H, et al. RETA-LLM: a retrieval-augmented large language model toolkit[EB/OL]. [2024-05-24]. https://arxiv.org/abs/2306.05212.
[55] YU Z C, XIONG C Y, YU S, et al. Augmentation-adapted retriever improves generalization of language models as generic plug-in[C]//Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics. Stroudsburg: ACL, 2023: 2421-2436.
[56] SHI W J, MIN S, YASUNAGA M, et al. REPLUG: retrieval-augmented black-box language models[EB/OL]. [2024-05-24]. https://arxiv.org/abs/2301.12652.
[57] YANG H Y, LI Z T, ZHANG Y, et al. PRCA: fitting black-box large language models for retrieval question answering via pluggable reward-driven contextual adapter[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Stroudsburg: ACL, 2023.
[58] XU F Y, SHI W J, CHOI E. RECOMP: improving retrieval-augmented LMs with compression and selective augmentation[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2310. 04408.
[59] LUO Z Y, XU C, ZHAO P, et al. Augmented large language models with parametric knowledge guiding[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2305.04757.
[60] GAO Y F, XIONG Y, GAO X Y, et al. Retrieval-augmented generation for large language models: a survey[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2312.10997.
[61] XU P, PING W, WU X C, et al. Retrieval meets long context large language models[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2310.03025.
[62] ZHU Y, YUAN H, WANG S, et al. Large language models for information retrieval: a survey[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2308.07107.
[63] MA Y B, CAO Y X, HONG Y, et al. Large language model is not a good few-shot information extractor, but a good reranker for hard samples![EB/OL]. [2024-06-20]. https://arxiv.org/abs/2303.08559.
[64] ZHUANG S Y, LIU B, KOOPMAN B, et al. Open-source large language models are strong zero-shot query likelihood models for document ranking[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2310.13243.
[65] CHENG X, LUO D, CHEN X, et al. Lift yourself up: retrieval-augmented text generation with self-memory[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2305.02437v3.
[66] ZOU A, WANG Z F, CARLINI N, et al. Universal and transferable adversarial attacks on aligned language models[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2307.15043.
[67] KANG M, KWAK J M, BAEK J, et al. Knowledge graph-augmented language models for knowledge-grounded dialogue generation[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2305.18846.
[68] LI X Z, LIU Z H, XIONG C Y, et al. Structure-aware language model pretraining improves dense retrieval on structured data[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2305. 19912.
[69] ZHAO P H, ZHANG H L, YU Q H, et al. Retrieval-augmented generation for AI-generated content: a survey[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2402.19473.
[70] HAN C, WANG Q F, PENG H, et al. LM-infinite: zero-shot extreme length generalization for large language models[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2308.16137.
[71] CUCONASU F, TRAPPOLINI G, SICILIANO F, et al. The power of noise: redefining retrieval for RAG systems[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2401.14887.
[72] DU Z X, QIAN Y J, LIU X, et al. GLM: general language model pretraining with autoregressive blank infilling[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2103.10360.
[73] MADAAN A, TANDON N, GUPTA P, et al. Self-refine: iterative refinement with self-feedback[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2303.17651.
[74] SHINN N, CASSANO F, GOPINATH A, et al. Reflexion: language agents with verbal reinforcement learning[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2303.11366v4.
[75] ZHANG W Q, TANG K, WU H, et al. Agent-Pro: learning to evolve via policy-level reflection and optimization[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2402.17574.
[76] TOY J, MACADAM J, TABOR P. Metacognition is all you need? Using introspection in generative agents to improve goal-directed behavior[EB/OL]. [2024-06-20]. https://arxiv. org/abs/2401.10910.
[77] ZHOU P, MADAAN A, POTHARAJU S P, et al. How FaR are large language models from agents with theory-of-mind?[EB/OL]. [2024-06-20]. https://arxiv.org/abs/2310. 03051.
[78] ZHAO A, HUANG D, XU Q, et al. ExpeL: LLM agents are experiential learners[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2024, 38(17): 19632-19642.
[79] LIU B, JIANG Y Q, ZHANG X H, et al. LLM+P: empowering large language models with optimal planning proficiency[EB/OL]. [2024-06-27]. https://arxiv.org/abs/2304. 11477.
[80] LI Y C, WEN H, WANG W J, et al. Personal LLM agents: insights and survey about the capability, efficiency and security[EB/OL]. [2024-06-27]. https://arxiv.org/abs/2401.05459.
[81] MAI J J, CHEN J, LI B, et al. LLM as a robotic brain: unifying egocentric memory and control[EB/OL]. [2024-06-27]. https://arxiv.org/abs/2304.09349.
[82] SONG C H, SADLER B M, WU J M, et al. LLM-planner: few-shot grounded planning for embodied agents with large language models[C]//Proceedings of the 2023 IEEE/CVF International Conference on Computer Vision. Piscataway: IEEE, 2023: 2986-2997.
[83] VALMEEKAM K, MARQUEZ M, SREEDHARAN S, et al. On the planning abilities of large language models - a critical investigation[C]//Advances in Neural Information Processing Systems 36, 2023: 75993-76005.
[84] HONG S R, ZHUGE M C, CHEN J Q, et al. MetaGPT: meta programming for a multi-agent collaborative framework[EB/OL]. [2024-06-27]. https://arxiv.org/abs/2308.00352.
[85] ZHANG H X, DU W H, SHAN J M, et al. Building cooperative embodied agents modularly with large language models[EB/OL]. [2024-06-27]. https://arxiv.org/abs/2307. 02485.
[86] TALEBIRAD Y, NADIRI A. Multi-agent collaboration: harnessing the power of intelligent LLM agents[EB/OL]. [2024-06-27]. https://arxiv.org/abs/2306.03314.
[87] ABDELNABI S, GOMAA A, SIVAPRASAD S, et al. LLM-deliberation: evaluating LLMs with interactive multi-agent negotiation games[EB/OL]. [2024-06-27]. https://arxiv.org/abs/2309.17234.
[88] AGASHE S, FAN Y, REYNA A, et al. LLM-coordination: evaluating and analyzing multi-agent coordination abilities in large language models[EB/OL]. [2024-06-27]. https:// arxiv.org/abs/2310.03903.
[89] LIANG T, HE Z W, JIAO W X, et al. Encouraging divergent thinking in large language models through multi-agent debate[EB/OL]. [2024-06-27]. https://arxiv.org/abs/2305.19118.
[90] ZHANG C Y, YANG K J, HU S Y, et al. ProAgent: building proactive cooperative agents with large language models[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2024, 38(16): 17591-17599.
[91] OZCAN F, SUBRAHMANIAN V S, DIX J. Improving performance of heavily loaded agents[EB/OL]. [2024-06-27]. https://arxiv.org/abs/cs/0012004.
[92] TIAN Y, YANG X, ZHANG J Y, et al. Evil geniuses: delving into the safety of LLM-based agents[EB/OL]. [2024-06-27]. https://arxiv.org/abs/2311.11855.
[93] KANG B, KIM J, YUN T R, et al. Prompt-RAG: pioneering vector embedding-free retrieval-augmented generation in niche domains, exemplified by Korean medicine[EB/OL]. [2024-06-27]. https://arxiv.org/abs/2401.11246.
[94] TAN H Z, LUO Q, JIANG L, et al. Prompt-based code completion via multi-retrieval augmented generation[EB/OL]. [2024-06-27]. https://arxiv.org/abs/2405.07530.
[95] HU Z B, WANG C, SHU Y F, et al. Prompt perturbation in retrieval-augmented generation based large language models[C]//Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. New York: ACM, 2024: 1119-1130.
[96] JUNPRUNG E. Exploring the intersection of large language models and agent-based modeling via prompt engineering[EB/OL]. [2024-06-27]. https://arxiv.org/abs/2308. 07411.
[97] LIU Z H, ZENG R N, WANG D X, et al. Agents4PLC: automating closed-loop PLC code generation and verification in industrial control systems using LLM-based agents[EB/OL]. [2024-12-11]. https://arxiv.org/abs/2410.14209.
[98] HUANG X W, RUAN W J, HUANG W, et al. A survey of safety and trustworthiness of large language models through the lens of verification and validation[J]. Artificial Intelligence Review, 2024, 57(7): 175.
[99] GALLEGOS I O, ROSSI R A, BARROW J, et al. Bias and fairness in large language models: a survey[J]. Computational Linguistics, 2024, 50(3): 1097-1179.
[100] TANNERU S H, AGARWAL C, LAKKARAJU H. Quantifying uncertainty in natural language explanations of large language models[C]//Proceedings of the 2024 International Conference on Artificial Intelligence and Statistics, 2024: 1072-1080.