pith. sign in

arxiv: 2606.01385 · v1 · pith:I4R3A6EPnew · submitted 2026-05-31 · 💻 cs.SE · cs.AI

Bridging Requirements and Architecture: Multi-Agent Orchestration with External Knowledge and Hierarchical Memory

Ruiyin Li , Yiran Zhang , Xiyu Zhou , Yangxiao Cai , Peng Liang , Weisong Sun , Jifeng Xuan , Zhi Jin
show 1 more author
Yang Liu
This is my paper

Pith reviewed 2026-06-28 16:29 UTC · model grok-4.3

classification 💻 cs.SE cs.AI
keywords multi-agent orchestrationsoftware architecturerequirements to architectureLLM-based agentsRAGhierarchical memoryquality evaluation
0
0 comments X

The pith

A multi-agent framework converts requirements into complete, modular software architectures using external knowledge and memory.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents MAAD as a system with four agents that work together to design software architectures from given requirements. By incorporating recognized standards through retrieval and maintaining design history in memory, the agents refine their outputs over iterations. Experiments show these architectures score higher in completeness, modularity, and traceability, while a dedicated agent creates evaluation reports to lessen the need for manual reviews. Performance varies with the underlying language model used.

Core claim

MAAD orchestrates Analyst, Modeler, Designer, and Evaluator agents to autonomously transform requirements specifications into comprehensive, multi-view architectural blueprints with quality attribute assessments. It incorporates RAG to inject recognized architectural standards and patterns and leverages hierarchical memory for iterative refinement. Evaluation across case studies demonstrates superior generation of complete, modular, and traceable architectures compared to baselines, with the Evaluator producing structured reports that reduce manual validation efforts. The quality depends on the LLM's reasoning capacity.

What carries the argument

Four specialized agents (Analyst, Modeler, Designer, Evaluator) coordinated with RAG-injected knowledge and hierarchical memory for iterative design refinement.

If this is right

  • Architectures generated are more complete, modular, and traceable.
  • The Evaluator agent autonomously produces structured quality evaluation reports.
  • This reduces manual validation efforts significantly.
  • Quality of outputs heavily depends on the reasoning capacity of the underlying language model.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Applying similar agent orchestration to other design tasks could automate more of the software development lifecycle.
  • The memory mechanism might help in maintaining consistency across larger projects with evolving requirements.
  • Testing the framework with different knowledge bases could reveal how much external information is necessary for good results.

Load-bearing premise

The specialized agents can collaborate effectively and produce valid architectures solely through RAG knowledge and hierarchical memory without requiring human corrections during the process.

What would settle it

Observing whether the agents produce incoherent or invalid architectures in a new set of requirements without any human input during iterations.

Figures

Figures reproduced from arXiv: 2606.01385 by Jifeng Xuan, Peng Liang, Ruiyin Li, Weisong Sun, Xiyu Zhou, Yang Liu, Yangxiao Cai, Yiran Zhang, Zhi Jin.

Figure 1
Figure 1. Figure 1: Overview of the MAAD Framework external knowledge infusion. The Evaluator agent performs iterative assessments of intermediate architectural artifacts through inter-agent interactions, examining their consistency with SRS requirements, architectural completeness, design consistency between architectural decisions, quality attributes (QAs), and stated design rationale, and adherence to architecture design p… view at source ↗
Figure 2
Figure 2. Figure 2: Process of the Study Design [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of the Class Diagrams Generated by MAAD and MetaGPT [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of Sequence Diagrams Generated by MAAD and MetaGPT [PITH_FULL_IMAGE:figures/full_fig_p024_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The Component Diagram of SFS Generated with Reference Knowledge [PITH_FULL_IMAGE:figures/full_fig_p028_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The Component Diagram of SFS Generated without Reference Knowledge [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of class diagrams generated by GPT-5.2, Qwen3.5, DeepSeek-R1, and Llama3.3 [PITH_FULL_IMAGE:figures/full_fig_p030_7.png] view at source ↗
read the original abstract

Software architecture design is a critical yet inherently complex and knowledge-intensive phase that requires balancing competing quality attributes and adapting to evolving requirements. Traditionally, this process has been time-consuming, labor-intensive, and heavily reliant on architects, often resulting in limited exploration of alternative architectural decompositions and styles, especially under the pressures of agile development. While LLM-based agents have shown promising performance across various software engineering tasks, their application to architecture design remains relatively scarce and requires systematic exploration. To address these challenges, we proposed MAAD (Multi-Agent Architecture Design), a knowledge-driven framework that orchestrates four specialized agents (i.e., Analyst, Modeler, Designer and Evaluator) to autonomously and collaboratively transform requirements specifications into comprehensive, multi-view architectural blueprints with quality attribute assessments. MAAD incorporates RAG to inject recognized architectural standards and patterns into the workflow and leverages a hierarchical memory mechanism that captures design history for iterative refinement. We evaluated MAAD through comparative experiments against MetaGPT, using quantitative architecture-level metrics across 10 case studies and qualitative feedback from industry architects on 10 real-world specifications. Results show that MAAD generates more complete, modular, and traceable architectures than the baseline, and its dedicated Evaluator agent autonomously produces structured quality evaluation reports that significantly reduce manual validation efforts. Furthermore, we found that the quality of the generated architecture heavily depends on the underlying LLM's reasoning capacity, with GPT-5.2 and Qwen3.5 outperforming other models across most evaluation settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes MAAD, a multi-agent framework orchestrating four specialized agents (Analyst, Modeler, Designer, Evaluator) that uses RAG to inject architectural standards and hierarchical memory for iterative refinement, to autonomously generate multi-view architectures and quality assessments from requirements. It reports comparative results on 10 case studies against MetaGPT showing superior completeness, modularity, and traceability via quantitative architecture-level metrics, plus qualitative feedback from industry architects on 10 real-world specs; the Evaluator is claimed to produce structured reports that significantly reduce manual validation, with performance depending on the underlying LLM (e.g., GPT-5.2, Qwen3.5).

Significance. If the reported gains hold under rigorous metric definitions, this would be a useful empirical contribution to multi-agent LLM systems for knowledge-intensive SE tasks, extending prior agent frameworks like MetaGPT with explicit RAG and memory mechanisms for architecture design. The work provides an initial demonstration of autonomous quality evaluation reports, which could reduce architect workload if validated.

major comments (3)
  1. [Evaluation / comparative experiments] Evaluation section (comparative experiments on 10 case studies): the headline claim of more complete, modular, and traceable architectures rests on 'quantitative architecture-level metrics' whose definitions, scoring rubrics, formulas, or inter-rater validation procedures are not supplied; without these it is impossible to determine whether reported deltas reflect genuine improvement or differences in metric application.
  2. [Abstract and results] Abstract and results discussion: the claim that the Evaluator agent 'significantly reduce[s] manual validation efforts' is unsupported by any timed measurements, count of validation steps saved, or controlled comparison; this is load-bearing for the practical-utility argument.
  3. [MAAD framework / workflow] Workflow description: the framework assumes the four agents reliably collaborate and produce valid outputs using only RAG and hierarchical memory without human correction during iteration, yet no failure rates, oversight logs, or robustness statistics are reported to support this assumption.
minor comments (2)
  1. [Results] The paper mentions GPT-5.2; clarify whether this is a hypothetical or specific model version and provide exact model identifiers used in experiments.
  2. [Figures and tables] Figure captions and table legends could more explicitly link visual outputs to the quantitative metrics discussed in the text.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important gaps in transparency and evidence. We address each major point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Evaluation / comparative experiments] Evaluation section (comparative experiments on 10 case studies): the headline claim of more complete, modular, and traceable architectures rests on 'quantitative architecture-level metrics' whose definitions, scoring rubrics, formulas, or inter-rater validation procedures are not supplied; without these it is impossible to determine whether reported deltas reflect genuine improvement or differences in metric application.

    Authors: We agree that the absence of explicit metric definitions, rubrics, formulas, and validation procedures limits the interpretability of the results. In the revised manuscript, we will add a dedicated subsection in the Evaluation section providing precise definitions, scoring criteria, formulas, and details on any inter-rater procedures used. revision: yes

  2. Referee: [Abstract and results] Abstract and results discussion: the claim that the Evaluator agent 'significantly reduce[s] manual validation efforts' is unsupported by any timed measurements, count of validation steps saved, or controlled comparison; this is load-bearing for the practical-utility argument.

    Authors: The claim rests on qualitative observations and architect feedback rather than quantitative time or step measurements. We will revise the abstract and results sections to qualify the language (e.g., 'produces structured reports that can reduce manual validation efforts') and explicitly note the lack of timed comparisons as a limitation requiring future work. revision: partial

  3. Referee: [MAAD framework / workflow] Workflow description: the framework assumes the four agents reliably collaborate and produce valid outputs using only RAG and hierarchical memory without human correction during iteration, yet no failure rates, oversight logs, or robustness statistics are reported to support this assumption.

    Authors: Our experiments emphasized successful outcomes on the 10 case studies and did not systematically log failure rates or interventions. We will add a discussion of workflow assumptions and any observed interaction patterns, while acknowledging the absence of robustness statistics as a limitation and suggesting it as future work. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical framework with external baseline comparison

full rationale

The paper proposes MAAD as an empirical multi-agent framework evaluated on 10 case studies against the external baseline MetaGPT, using quantitative architecture-level metrics and qualitative architect feedback. No equations, derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the abstract or described workflow. Central claims rest on comparative results and RAG/memory mechanisms rather than any self-referential reduction; the evaluation is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that LLM agents can autonomously handle complex, knowledge-intensive architecture design when augmented with retrieval and memory; no free parameters or invented physical entities are introduced.

axioms (1)
  • domain assumption LLM-based agents can effectively collaborate on architecture design when supplied with external standards via RAG and design history via hierarchical memory
    Invoked in the description of the MAAD workflow and its claimed autonomy.
invented entities (1)
  • MAAD four-agent orchestration (Analyst, Modeler, Designer, Evaluator) no independent evidence
    purpose: To autonomously transform requirements into architectural blueprints
    New system proposed in the paper; no independent evidence outside the reported experiments is provided.

pith-pipeline@v0.9.1-grok · 5818 in / 1384 out tokens · 39105 ms · 2026-06-28T16:29:06.077057+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 5 canonical work pages · 3 internal anchors

  1. [1]

    ISO/IEC/IEEE 42010. 2022. ISO/IEC/IEEE 42010:2022 - Software, systems and enterprise - Architecture description

    2022

  2. [2]

    ISO/IEC/IEEE 42020. 2019. ISO/IEC/IEEE 42020:2019 - Software, systems and enterprise - Architecture processes

    2019

  3. [3]

    Aakash Ahmad, Muhammad Waseem, Peng Liang, Mahdi Fahmideh, Mst Shamima Aktar, and Tommi Mikkonen

  4. [4]

    InProceedings of the 27th International Conference on Evaluation and Assessment in Software Engineering (EASE)

    Towards human-bot collaborative software architecting with ChatGPT. InProceedings of the 27th International Conference on Evaluation and Assessment in Software Engineering (EASE). ACM, 279–285

  5. [5]

    2021.Software Architecture in Practice(4th ed.)

    Len Bass, Paul Clements, and Rick Kazman. 2021.Software Architecture in Practice(4th ed.). Addison-Wesley Professional

    2021

  6. [6]

    Egon Börger, Alessandra Cavarra, and Elvinia Riccobene. 2000. Modeling the dynamics of UML state machines. In Proceedings of the 7th International Workshop on Abstract State Machines (ASM). Springer, 223–241

    2000

  7. [7]

    Chidamber and Chris F

    Shyam R. Chidamber and Chris F. Kemerer. 1994. A metrics suite for object oriented design.IEEE Transactions on Software Engineering20, 6 (1994), 476–493

    1994

  8. [8]

    Ngo Cong-Lem, Ali Soyoof, and Diki Tsering. 2025. A systematic review of the limitations and associated opportunities of ChatGPT.International Journal of Human–Computer Interaction41, 7 (2025), 3851–3866

    2025

  9. [9]

    Rudra Dhar, Karthik Vaidhyanathan, and Vasudeva Varma. 2024. Can LLMs Generate Architectural Design Decisions? - An Exploratory Empirical study. InProceedings of the 21st IEEE International Conference on Software Architecture (ICSA). IEEE, 79–89

    2024

  10. [10]

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. 2024. Improving factuality and reasoning in language models through multiagent debate. InProceedings of the 41st International Conference on Machine Learning (ICML). OpenReview.net, 1–9

    2024

  11. [11]

    Abhimanyu Dubey et al. 2024. The Llama 3 Herd of Models.arXiv preprint arXiv:2407.21783(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  12. [12]

    Tobias Eisenreich, Sandro Speth, and Stefan Wagner. 2024. From Requirements to Architecture: An AI-Based Journey to Semi-Automatically Generate Software Architectures. InProceedings of the 1st International Workshop on Designing Software (Designing). ACM, 52–55

    2024

  13. [13]

    Matteo Esposito, Xiaozhou Li, Sergio Moreschini, Noman Ahmad, Tomas Cerny, Karthik Vaidhyanathan, Valentina Lenarduzzi, and Davide Taibi. 2026. Generative AI for Software Architecture. Applications, Challenges, and Future Directions.Journal of Systems and Software231 (2026), 112607

    2026

  14. [14]

    2004.Domain-Driven Design: Tackling Complexity in the Heart of Software

    Eric Evans. 2004.Domain-Driven Design: Tackling Complexity in the Heart of Software. Addison-Wesley Professional

    2004

  15. [15]

    2006.The Architecture Analysis & Design Language (AADL): An Introduction

    Peter H Feiler, David P Gluch, and John J Hudak. 2006.The Architecture Analysis & Design Language (AADL): An Introduction. Technical Report CMU/SEI-2006-TN-011

    2006

  16. [16]

    Alessio Ferrari, Giorgio Oronzo Spagnolo, and Stefania Gnesi. 2017. Pure: A dataset of public requirements documents. InProceedings of the 25th IEEE International Requirements Engineering Conference (RE). IEEE, 502–505

    2017

  17. [17]

    David Garlan, Robert Allen, and John Ockerbloom. 2009. Architectural mismatch: Why reuse is still so hard.IEEE Software26, 4 (2009), 66–69

    2009

  18. [18]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. 2025. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.arXiv preprint arXiv:2501.12948(2025). ACM Trans. Softw. Eng. Methodol., Vol. 0, No. 0, Article 0. Publication date: 2026. 0:38 Li et al

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  19. [19]

    Junda He, Christoph Treude, and David Lo. 2025. LLM-Based Multi-Agent Systems for Software Engineering: Literature Review, Vision and the Road Ahead.ACM Transactions on Software Engineering and Methodology34, 5 (2025), 1–30

    2025

  20. [20]

    Christine Hofmeister, Philippe Kruchten, Robert L Nord, Henk Obbink, Alexander Ran, and Pierre America. 2007. A general model of software architecture design derived from five industrial approaches.Journal of Systems and Software 80, 1 (2007), 106–126

    2007

  21. [21]

    Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, and Liyang Zhou. 2024. MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework. In Proceedings of the 12th International Conference on Learning Representations (ICLR). OpenReview.net, 1–29

    2024

  22. [22]

    Dongming Jin, Shengxin Zhao, Zhi Jin, Xiaohong Chen, Chunhui Wang, Zheng Fang, and Hongbin Xiao. 2024. An evaluation of requirements modeling for cyber-physical systems via LLMs.arXiv preprint arXiv:2408.02450(2024)

    work page arXiv 2024

  23. [23]

    2000.ATAM: Method for Architecture Evaluation

    Rick Kazman, Mark Klein, and Paul Clements. 2000.ATAM: Method for Architecture Evaluation. Technical Report CMU/SEI-2000-TR-004

    2000

  24. [24]

    Philippe B Kruchten. 1995. The 4 + 1 view model of architecture.IEEE Software12, 6 (1995), 42–50

    1995

  25. [25]

    Usha Kumari and Shuchita Upadhyaya. 2011. An interface complexity measure for component-based software systems. International Journal of Computer Applications36, 1 (2011), 46–52

    2011

  26. [26]

    1996.Large-Scale C++ Software Design

    John Lakos. 1996.Large-Scale C++ Software Design. Addison-Wesley Professional

    1996

  27. [27]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, and Tim Rocktäschel. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. InProceedings of the 34th Annual Conference on Neural Information Processing Systems (NeurIPS). Curran Associates, 9459–9474

    2020

  28. [28]

    Jialong Li, Mingyue Zhang, Nianyu Li, Danny Weyns, Zhi Jin, and Kenji Tei. 2024. Generative AI for self-adaptive systems: State of the art and research roadmap.ACM Transactions on Autonomous and Adaptive Systems19, 3 (2024), 1–60

    2024

  29. [29]

    Ruiyin Li, Peng Liang, Yifei Wang, Yangxiao Cai, Weisong Sun, and Zengyang Li. 2026. Unveiling the Role of ChatGPT in Software Development: Insights from Developer-ChatGPT Interactions on GitHub.ACM Transactions on Software Engineering and Methodology(2026)

    2026

  30. [30]

    Ruiyin Li, Yiran Zhang, Xiyu Zhou, Yangxiao Cai, Peng Liang, Weisong Sun, Jifeng Xuan, Zhi Jin, and Yang Liu. 2026. Replication Package for the Paper: Bridging Requirements and Architecture: Multi-Agent Orchestration with External Knowledge and Hierarchical Memory. https://github.com/RuiyinL/MAAD

    2026

  31. [31]

    Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts.Transactions of the Association for Computational Linguistics12 (2024), 157–173

    2024

  32. [32]

    2017.Clean Architecture: A Craftsman’s Guide to Software Structure and Design

    Robert C Martin. 2017.Clean Architecture: A Craftsman’s Guide to Software Structure and Design. Prentice Hall Press

    2017

  33. [33]

    Thomas J. McCabe. 1976. A complexity measure.IEEE Transactions on Software EngineeringSE-2, 4 (1976), 308–320

    1976

  34. [34]

    Nenad Medvidovic and Richard N Taylor. 2000. A classification and comparison framework for software architecture description languages.IEEE Transactions on Software Engineering26, 1 (2000), 70–93

    2000

  35. [35]

    2004.MDA Distilled: Principles of Model-Driven Architecture

    Stephen J Mellor, Kendall Scott, Axel Uhl, and Dirk Weise. 2004.MDA Distilled: Principles of Model-Driven Architecture. Addison-Wesley Professional

    2004

  36. [36]

    2025.Update to GPT-5 System Card: GPT-5.2

    OpenAI. 2025.Update to GPT-5 System Card: GPT-5.2. Technical Report

    2025

  37. [37]

    2020.Fundamentals of Software Architecture: An Engineering Approach

    Mark Richards and Neal Ford. 2020.Fundamentals of Software Architecture: An Engineering Approach. O’Reilly Media

    2020

  38. [38]

    Larissa Schmid, Tobias Hey, Martin Armbruster, Sophie Corallo, Dominik Fuchß, Jan Keim, Haoyu Liu, and Anne Koziolek. 2025. Software Architecture Meets LLMs: A Systematic Literature Review.arXiv preprint arXiv:2505.16697 (2025)

    work page arXiv 2025

  39. [39]

    1996.Software Architecture: Perspectives on An Emerging Discipline

    Mary Shaw and David Garlan. 1996.Software Architecture: Perspectives on An Emerging Discipline. Vol. 1. Prentice Hall

    1996

  40. [40]

    Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L

    Theodore R. Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L. Griffiths. 2024. Cognitive Architectures for Language Agents.Transactions on Machine Learning Research(2024), 1–32

    2024

  41. [41]

    Xiaoxi Sun, Jinpeng Li, Yan Zhong, Dongyan Zhao, and Rui Yan. 2025. Towards detecting LLMs hallucination via markov chain-based multi-agent debate framework. InProceedings of the 50th IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 1–5

    2025

  42. [42]

    Yashar Talebirad and Amirhossein Nadiri. 2023. Multi-agent collaboration: Harnessing the power of intelligent LLM agents.arXiv preprint arXiv:2306.03314(2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  43. [43]

    Qwen Team. 2026. Qwen3.5: Towards Native Multimodal Agents. https://qwen.ai/blog?id=qwen3.5

    2026

  44. [44]

    Dan Tofan. 2010. Tacit architectural knowledge. InProceedings of the Fourth European Conference on Software Architecture (ECSA) Companion. ACM, 9–11

    2010

  45. [45]

    Zhiyuan Wan, Yun Zhang, Xin Xia, Yi Jiang, and David Lo. 2023. Software Architecture in Practice: Challenges and Opportunities. InProceedings of the 31st ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering (ESEC/FSE). ACM, 1457–1469. ACM Trans. Softw. Eng. Methodol., Vol. 0, No. 0, Article 0. Publica...

    2023

  46. [46]

    2012.Experimentation in Software Engineering

    Claes Wohlin, Per Runeson, Martin Höst, Magnus C Ohlsson, Björn Regnell, and Anders Wesslén. 2012.Experimentation in Software Engineering. Springer Science & Business Media

    2012

  47. [47]

    Mengfei Yang, Bin Gu, Zhenhua Duan, Zhi Jin, Naijun Zhan, Yunwei Dong, Cong Tian, Ge Li, Xiaogang Dong, and Xiaofeng Li. 2022. Intelligent program synthesis framework and key scientific problems for embedded software. Chinese Space Science and Technology42, 4 (2022), 1–7

    2022

  48. [48]

    Yiran Zhang, Ruiyin Li, Peng Liang, Sun Weisong, and Liu Yang. 2025. Knowledge-Based Multi-Agent Framework for Automated Software Architecture Design. InProceedings of the 33rd ACM International Conference on the Foundations of Software Engineering (FSE) IVR Track. ACM, 530–534

    2025

  49. [49]

    Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, Longyue Wang, Ahn Tuan Luu, Wei Bi, Freda Shi, and Shuming Shi. 2025. Siren’s Song in the AI Ocean: A Survey on Hallucination in Large Language Models.Computational Linguistics51, 4 (2025), 1–45

    2025

  50. [50]

    Zeyu Zhang, Quanyu Dai, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. 2025. A survey on the memory mechanism of large language model-based agents.ACM Transactions on Information Systems 43, 6 (2025), 1–47. ACM Trans. Softw. Eng. Methodol., Vol. 0, No. 0, Article 0. Publication date: 2026

    2025