pith. sign in

arxiv: 2506.10484 · v2 · submitted 2025-06-12 · 💻 cs.SE

EXPEREPAIR: Dual-Memory Enhanced LLM-based Repository-Level Program Repair

Fangwen Mu , Junjie Wang , Lin Shi , Song Wang , Shoubin Li , Qing Wang This is my paper

Pith reviewed 2026-05-19 10:05 UTC · model grok-4.3

classification 💻 cs.SE
keywords LLM-based program repairrepository-level repairdual-memory systemsepisodic memorysemantic memorydynamic prompt compositionSWE-Bench
0
0 comments X

The pith

ExpeRepair improves LLM repository-level repair by storing past fixes in episodic and semantic memories to build dynamic prompts.

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

The paper presents ExpeRepair as a way for large language models to handle software bugs in entire code repositories more effectively than before. It draws from human cognition by maintaining an episodic memory of specific past repair examples and a semantic memory of broader insights drawn from those examples. When a new issue arises the system retrieves relevant items from both memories and assembles a custom prompt on the fly instead of relying on a fixed template. This experience-driven approach yields the highest pass@1 scores among tested open-source methods on standard benchmarks. A reader would care because it points toward AI repair tools that can learn and improve across successive tasks rather than starting fresh each time.

Core claim

ExpeRepair organizes prior repair knowledge into episodic memory that stores concrete repair demonstrations and semantic memory that encodes abstract reflective insights; at inference it retrieves relevant demonstrations and insights from both memories and integrates them via dynamic prompt composition to replace static prompts with context-aware ones, producing pass@1 scores of 60.3 percent on SWE-Bench Lite and 74.6 percent on SWE-Bench Verified, the best among evaluated open-source methods.

What carries the argument

Dual-memory knowledge accumulation with episodic memory for concrete demonstrations, semantic memory for abstract insights, retrieval activation of both, and dynamic prompt composition that adapts prompts from accumulated experience.

Load-bearing premise

Retrieval from the episodic and semantic memories plus dynamic prompt composition will reliably surface useful prior repair knowledge for previously unseen issues rather than introducing noise or irrelevant context.

What would settle it

Disabling the memory retrieval and dynamic composition steps and measuring whether pass@1 scores on SWE-Bench Lite and Verified fall to the level of a static-prompt baseline.

Figures

Figures reproduced from arXiv: 2506.10484 by Fangwen Mu, Junjie Wang, Lin Shi, Qing Wang, Shoubin Li, Song Wang.

Figure 1
Figure 1. Figure 1: The overview of EXPEREPAIR. 3.1 Workflow of EXPEREPAIR Conventional learning approaches [33, 14, 27, 13] typically rely on computationally expensive fine￾tuning to improve the capabilities of LLM-based agents. In contrast, EXPEREPAIR equips LLM agents with the dual-memory mechanism, enabling them to continuously acquire knowledge from their own repair experiences without requiring parametric updates. Howev… view at source ↗
Figure 2
Figure 2. Figure 2: Intersection analysis between EXPERE￾PAIR and leading baselines. Since PatchPilot does not disclose per-instance resolution results, we sub￾stitute Moatless for PatchPilot in this analysis. In addition, we conduct an intersection analysis comparing the issues resolved by EXPEREPAIR against four leading baselines. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Results of EXPEREPAIR on different models To evaluate the adaptability of EXPEREPAIR across various large language models (LLMs), we conduct an experiment by integrating our framework with four popular LLMs: o1-mini [24], DeepSeek-R1 [8], Claude 3.5 [2], and Claude 3.7 [3]. This setup aims to isolate the influence of different underlying models while keeping all other components of EXPEREPAIR unchanged. Fo… view at source ↗
Figure 4
Figure 4. Figure 4: This prompt was used to generate the reproduction script during the test generation phase. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: This prompt was used to generate the validation tests during the patch validation phase. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: This prompt was used to generate the candidate patches during the patch generation phase. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: This prompt was used to refine the candidate patches during the patch validation phase. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: This prompt was used to summarize and update the insights during the memory storage [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
read the original abstract

Automatically repairing software issues remains a fundamental challenge at the intersection of software engineering and AI. Although recent advances in Large Language Models (LLMs) have demonstrated potential for repository-level repair tasks, current methods exhibit two notable limitations: (1) they often address issues in isolation, neglecting to incorporate insights from previously resolved issues, and (2) they rely on static, rigid prompting strategies that constrain their ability to generalize across diverse and evolving contexts. We propose ExpeRepair, a novel LLM-based program repair framework inspired by the dual-memory systems of human cognition, where episodic and semantic memory synergistically support learning and decision-making. Unlike existing methods, ExpeRepair continuously learns from historical repair experiences via dual-channel knowledge accumulation, enabling it to adaptively reuse past knowledge during inference. Specifically, ExpeRepair organizes prior repair knowledge into two complementary memories: an episodic memory that stores concrete repair demonstrations, and a semantic memory that encodes abstract, reflective insights. At inference time, ExpeRepair activates both memory systems by retrieving relevant demonstrations from episodic memory and recalling high-level repair insights from semantic memory. It further enhances adaptability through dynamic prompt composition, integrating both memory types to replace static prompts with context-aware, experience-driven prompts. We evaluate ExpeRepair on two benchmarks: SWE-Bench Lite and SWE-Bench Verified. Experimental results show that ExpeRepair achieves pass@1 scores of 60.3% and 74.6% on the two benchmarks, respectively, achieving the best performance among the evaluated open-source methods. We have open-sourced ExpeRepair at https://github.com/ExpeRepair/ExpeRepair.

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 / 3 minor

Summary. The paper proposes ExpeRepair, a dual-memory LLM framework for repository-level program repair. It maintains an episodic memory of concrete repair demonstrations and a semantic memory of abstract insights, retrieved at inference time and combined via dynamic prompt composition to replace static prompts. Evaluated on SWE-Bench Lite and SWE-Bench Verified, it reports pass@1 scores of 60.3% and 74.6% respectively, claiming the best results among evaluated open-source methods, with the code open-sourced.

Significance. If the performance gains are shown to stem from genuine out-of-distribution knowledge reuse rather than retrieval artifacts or leakage, the dual-memory design offers a cognitively motivated alternative to static prompting in LLM-based repair. The open-sourcing and concrete benchmark numbers are positive for reproducibility; the work could influence follow-on research on experience-driven adaptation in software engineering tasks.

major comments (3)
  1. [§3.2] §3.2 (Memory Construction): The manuscript does not specify the exact corpus, repositories, or historical issues used to seed the episodic and semantic memories. Without explicit confirmation that SWE-Bench test issues and same-repository bugs were excluded from memory population, the reported gains cannot be confidently attributed to generalization rather than data leakage or prompt inflation.
  2. [§4.2–4.3] §4.2–4.3 (Experimental Results and Ablations): The evaluation lacks detailed ablation controls isolating the contribution of episodic retrieval, semantic insights, and dynamic prompt composition. The headline pass@1 numbers are therefore difficult to interpret as evidence for the dual-memory claim versus other factors such as base LLM choice or prompt engineering.
  3. [§3.3] §3.3 (Retrieval and Prompt Composition): The similarity function, retrieval threshold, and memory-update rules are not described. This omission prevents verification that retrieved content is relevant and non-noisy for previously unseen issues, directly bearing on the central claim that dual-memory retrieval improves repair over prior open-source baselines.
minor comments (3)
  1. [Abstract] Abstract: The statement that ExpeRepair achieves the 'best performance among the evaluated open-source methods' should be supported by an explicit list or table of compared systems and their scores.
  2. [Related Work] Related Work section: Additional citations to prior LLM repository repair systems (e.g., those using retrieval-augmented generation) would better situate the dual-memory contribution.
  3. [Figure 2] Figure 2 (architecture diagram): Labels distinguishing episodic vs. semantic memory flows and the dynamic composition step should be enlarged for clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment point by point below, indicating where revisions will be made to improve clarity and strengthen the presentation of our results.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (Memory Construction): The manuscript does not specify the exact corpus, repositories, or historical issues used to seed the episodic and semantic memories. Without explicit confirmation that SWE-Bench test issues and same-repository bugs were excluded from memory population, the reported gains cannot be confidently attributed to generalization rather than data leakage or prompt inflation.

    Authors: We agree that additional specificity on memory construction is needed to fully address concerns about potential leakage. In the revised manuscript we will expand §3.2 to explicitly describe the corpus: episodic memory is populated with concrete repair demonstrations drawn from historical GitHub issues and pull requests across a broad set of open-source repositories, while semantic memory is constructed via abstraction and reflection over those same experiences. We will add a clear statement confirming that the memory population process excluded all SWE-Bench test issues as well as any bugs from the same repositories appearing in the evaluation sets. This clarification will allow readers to attribute performance gains to out-of-distribution knowledge reuse. revision: yes

  2. Referee: [§4.2–4.3] §4.2–4.3 (Experimental Results and Ablations): The evaluation lacks detailed ablation controls isolating the contribution of episodic retrieval, semantic insights, and dynamic prompt composition. The headline pass@1 numbers are therefore difficult to interpret as evidence for the dual-memory claim versus other factors such as base LLM choice or prompt engineering.

    Authors: We acknowledge that more granular ablations would help isolate the contributions of each component. We will revise §§4.2–4.3 to include additional ablation experiments that separately disable episodic retrieval, semantic memory, and dynamic prompt composition while keeping the base LLM and other factors fixed. The updated results will be presented in a new table and accompanying analysis to demonstrate the incremental benefit of the full dual-memory design over partial configurations and static prompting baselines. revision: yes

  3. Referee: [§3.3] §3.3 (Retrieval and Prompt Composition): The similarity function, retrieval threshold, and memory-update rules are not described. This omission prevents verification that retrieved content is relevant and non-noisy for previously unseen issues, directly bearing on the central claim that dual-memory retrieval improves repair over prior open-source baselines.

    Authors: We thank the referee for noting this omission. In the revised §3.3 we will provide a complete technical description of the retrieval process, including the embedding model and cosine similarity function used for both memory types, the specific similarity threshold applied to select relevant items, and the memory-update rules that govern how new experiences are incorporated and how episodic content is periodically distilled into semantic insights. These details will enable independent verification that retrieved memories remain relevant and low-noise for unseen issues. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical results on external benchmarks with design choices independent of fitted inputs

full rationale

The paper presents ExpeRepair as an engineering framework that organizes prior repair knowledge into episodic and semantic memories and uses dynamic prompt composition at inference time. The central results are pass@1 scores measured directly on the external SWE-Bench Lite and Verified benchmarks rather than any internally derived predictions or reductions. No equations, fitted parameters renamed as predictions, self-definitional loops, or load-bearing self-citations appear in the derivation chain; the dual-memory design is explicitly described as an inspirational choice from human cognition, not a mathematical necessity derived from prior results by construction. The evaluation therefore remains self-contained against public benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework rests on the engineering assumption that historical repair traces can be partitioned into episodic demonstrations and semantic insights that remain useful at inference time; no new physical constants or mathematical axioms are introduced.

axioms (1)
  • domain assumption Prior repair experiences contain transferable knowledge that can be stored and retrieved to improve future repairs.
    Stated in the motivation section of the abstract as the basis for continuous learning from historical issues.

pith-pipeline@v0.9.0 · 5836 in / 1219 out tokens · 46362 ms · 2026-05-19T10:05:27.126462+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. ARISE: A Repository-level Graph Representation and Toolset for Agentic Fault Localization and Program Repair

    cs.SE 2026-05 unverdicted novelty 7.0

    ARISE adds a data-flow-augmented repository graph and three-tier tool API to LLM agents, raising Function Recall@1 by 17 points, Line Recall@1 by 15 points, and Pass@1 repair rate to 22% on SWE-bench Lite.

  2. Dynamic analysis enhances issue resolution

    cs.SE 2026-03 conditional novelty 7.0

    DAIRA integrates dynamic tracing into LLM agents to achieve 79.4% resolution rate on SWE-bench Verified for code defect repair.

  3. Beyond Fixed Tests: Repository-Level Issue Resolution as Coevolution of Code and Behavioral Constraints

    cs.SE 2026-04 unverdicted novelty 6.0

    Agent-CoEvo is a multi-agent LLM framework that coevolves code patches and test patches to resolve repository-level issues, outperforming fixed-test baselines on SWE-bench Lite and SWT-bench Lite.

  4. Rethinking the Value of Agent-Generated Tests for LLM-Based Software Engineering Agents

    cs.SE 2026-02 unverdicted novelty 5.0

    Agent-generated tests mainly act as observational feedback channels and do not meaningfully improve issue resolution success in current LLM software engineering agents.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages · cited by 4 Pith papers · 7 internal anchors

  1. [1]

    DARS: dynamic action re-sampling to enhance coding agent performance by adaptive tree traversal

    Vaibhav Aggarwal, Ojasv Kamal, Abhinav Japesh, Zhijing Jin, and Bernhard Schölkopf. DARS: dynamic action re-sampling to enhance coding agent performance by adaptive tree traversal. CoRR, abs/2503.14269, 2025

    work page arXiv 2025

  2. [2]

    Claude 3.5 sonnet v2

    anthropic. Claude 3.5 sonnet v2. https://www.anthropic.com/news/ claude-3-5-sonnet , 2025

    work page 2025

  3. [3]

    Claude 3.7 sonnet

    anthropic. Claude 3.7 sonnet. https://www.anthropic.com/claude/sonnet, 2025

    work page 2025

  4. [4]

    RepairAgent: An Autonomous, LLM-Based Agent for Program Repair

    Islem Bouzenia, Premkumar Devanbu, and Michael Pradel. Repairagent: An autonomous, llm-based agent for program repair. arXiv preprint arXiv:2403.17134, 2024

    work page internal anchor Pith review arXiv 2024

  5. [5]

    Yuxiao Chen, Jingzheng Wu, Xiang Ling, Changjiang Li, Zhiqing Rui, Tianyue Luo, and Yanjun Wu. When large language models confront repository-level automatic program repair: How well they done? In Proceedings of the 2024 IEEE/ACM 46th International Conference on Software Engineering: Companion Proceedings, pages 459–471, 2024

    work page 2024

  6. [6]

    Yuxiao Chen, Jingzheng Wu, Xiang Ling, Changjiang Li, Zhiqing Rui, Tianyue Luo, and Yanjun Wu. When large language models confront repository-level automatic program repair: How well they done? In Proceedings of the 2024 IEEE/ACM 46th International Conference on Software Engineering: Companion Proceedings, ICSE Companion 2024, Lisbon, Portugal, April 14-2...

    work page 2024

  7. [7]

    DARS. Dars. https://github.com/vaibhavagg303/DARS-Agent?tab= readme-ov-file, Feb 2025

    work page 2025

  8. [8]

    Deepseek-r1

    DeepSeek. Deepseek-r1. https://github.com/deepseek-ai/DeepSeek-R1, 2025

    work page 2025

  9. [9]

    Dual-process theories of higher cognition: Advancing the debate

    Jonathan St BT Evans and Keith E Stanovich. Dual-process theories of higher cognition: Advancing the debate. Perspectives on psychological science, 8(3):223–241, 2013

    work page 2013

  10. [10]

    Experepair project

    ExpeRepair. Experepair project. https://github.com/ExpeRepair/ExpeRepair, 2025

    work page 2025

  11. [11]

    Dual-process and dual-system theories of reasoning

    Keith Frankish. Dual-process and dual-system theories of reasoning. Philosophy Compass, 5(10):914–926, 2010

    work page 2010

  12. [12]

    Automatic software repair: A survey

    Luca Gazzola, Daniela Micucci, and Leonardo Mariani. Automatic software repair: A survey. In Proceedings of the 40th International Conference on Software Engineering, pages 1219–1219, 2018

    work page 2018

  13. [13]

    Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y . Wu, Y . K. Li, Fuli Luo, Yingfei Xiong, and Wenfeng Liang. Deepseek-coder: When the large language model meets programming - the rise of code intelligence. CoRR, abs/2401.14196, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  14. [14]

    Large Language Model based Multi-Agents: A Survey of Progress and Challenges

    Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  15. [15]

    Hjaltason and Hanan Samet

    Gísli R. Hjaltason and Hanan Samet. Properties of embedding methods for similarity searching in metric spaces. IEEE Trans. Pattern Anal. Mach. Intell., 25(5):530–549, 2003

    work page 2003

  16. [16]

    Self-evolving multi-agent collaboration networks for software development

    Yue Hu, Yuzhu Cai, Yaxin Du, Xinyu Zhu, Xiangrui Liu, Zijie Yu, Yuchen Hou, Shuo Tang, and Siheng Chen. Self-evolving multi-agent collaboration networks for software development. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net, 2025

    work page 2025

  17. [17]

    SWE-bench: Can Language Models Resolve Real-World GitHub Issues?

    Carlos E Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. Swe-bench: Can language models resolve real-world github issues? arXiv preprint arXiv:2310.06770, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  18. [18]

    Automated program repair

    Claire Le Goues, Michael Pradel, and Abhik Roychoudhury. Automated program repair. Communications of the ACM, 62(12):56–65, 2019. 10

    work page 2019

  19. [19]

    Patchpilot: A stable and cost-efficient agentic patching framework

    Hongwei Li, Yuheng Tang, Shiqi Wang, and Wenbo Guo. Patchpilot: A stable and cost-efficient agentic patching framework. arXiv preprint arXiv:2502.02747, 2025

    work page arXiv 2025

  20. [20]

    Llms as continuous learners: Improving the reproduction of defective code in software issues

    Yalan Lin, Yingwei Ma, Rongyu Cao, Binhua Li, Fei Huang, Xiaodong Gu, and Yongbin Li. Llms as continuous learners: Improving the reproduction of defective code in software issues. CoRR, abs/2411.13941, 2024

    work page arXiv 2024

  21. [21]

    Dynamine: finding common error patterns by mining software revision histories

    Benjamin Livshits and Thomas Zimmermann. Dynamine: finding common error patterns by mining software revision histories. ACM SIGSOFT Software Engineering Notes, 30(5):296–305, 2005

    work page 2005

  22. [22]

    Alibaba lingmaagent: Improving automated issue resolution via comprehensive repository exploration

    Yingwei Ma, Qingping Yang, Rongyu Cao, Binhua Li, Fei Huang, and Yongbin Li. Alibaba lingmaagent: Improving automated issue resolution via comprehensive repository exploration. arXiv preprint arXiv:2406.01422, 2024

    work page arXiv 2024

  23. [23]

    Ragfix: Enhancing llm code repair using rag and stack overflow posts

    Elijah Mansur, Johnson Chen, Muhammad Anas Raza, and Mohammad Wardat. Ragfix: Enhancing llm code repair using rag and stack overflow posts. In 2024 IEEE International Conference on Big Data (BigData), pages 7491–7496. IEEE, 2024

    work page 2024

  24. [24]

    Openai o1-mini

    OpenAI. Openai o1-mini. https://openai.com/index/ openai-o1-mini-advancing-cost-efficient-reasoning/ , 2025

    work page 2025

  25. [25]

    An analysis of patch plausibility and correctness for generate-and-validate patch generation systems

    Zichao Qi, Fan Long, Sara Achour, and Martin Rinard. An analysis of patch plausibility and correctness for generate-and-validate patch generation systems. In Proceedings of the 2015 international symposium on software testing and analysis, pages 24–36, 2015

    work page 2015

  26. [26]

    Robertson and Hugo Zaragoza

    Stephen E. Robertson and Hugo Zaragoza. The probabilistic relevance framework: BM25 and beyond. Found. Trends Inf. Retr., 3(4):333–389, 2009

    work page 2009

  27. [27]

    Code Llama: Open Foundation Models for Code

    Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton-Ferrer, Aaron Grattafiori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thoma...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  28. [28]

    A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications

    Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A systematic survey of prompt engineering in large language models: Techniques and applications. CoRR, abs/2402.07927, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  29. [29]

    Logan IV , Eric Wallace, and Sameer Singh

    Taylor Shin, Yasaman Razeghi, Robert L. Logan IV , Eric Wallace, and Sameer Singh. Au- toprompt: Eliciting knowledge from language models with automatically generated prompts. In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu, editors, Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November ...

    work page 2020

  30. [30]

    Repairllama: Efficient representations and fine-tuned adapters for program repair

    André Silva, Sen Fang, and Martin Monperrus. Repairllama: Efficient representations and fine-tuned adapters for program repair. arXiv preprint arXiv:2312.15698, 2023

    work page arXiv 2023

  31. [31]

    swebench

    swebench. swebench. https://www.swebench.com/, Feb 2025

    work page 2025

  32. [32]

    Openhands: An open platform for ai software developers as generalist agents

    Xingyao Wang, Boxuan Li, Yufan Song, Frank F Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, et al. Openhands: An open platform for ai software developers as generalist agents. In The Thirteenth International Conference on Learning Representations, 2024

    work page 2024

  33. [33]

    Sahraoui

    Martin Weyssow, Xin Zhou, Kisub Kim, David Lo, and Houari A. Sahraoui. Exploring parameter-efficient fine-tuning techniques for code generation with large language models. CoRR, abs/2308.10462, 2023

    work page arXiv 2023

  34. [34]

    Strago: Harnessing strategic guidance for prompt optimization

    Yurong Wu, Yan Gao, Bin Zhu, Zineng Zhou, Xiaodi Sun, Sheng Yang, Jian-Guang Lou, Zhiming Ding, and Linjun Yang. Strago: Harnessing strategic guidance for prompt optimization. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors,Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-16, 2024, pages...

    work page 2024

  35. [35]

    Agentless: Demystifying LLM-based Software Engineering Agents

    Chunqiu Steven Xia, Yinlin Deng, Soren Dunn, and Lingming Zhang. Agentless: Demystifying llm-based software engineering agents. arXiv preprint arXiv:2407.01489, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  36. [36]

    Swe-agent: Agent-computer interfaces enable automated software engineering

    John Yang, Carlos Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. Swe-agent: Agent-computer interfaces enable automated software engineering. Advances in Neural Information Processing Systems, 37:50528–50652, 2024

    work page 2024

  37. [37]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), 2023

    work page 2023

  38. [38]

    Autocoderover: Au- tonomous program improvement, 2024

    Yuntong Zhang, Haifeng Ruan, Zhiyu Fan, and Abhik Roychoudhury. Autocoderover: Au- tonomous program improvement, 2024

    work page 2024

  39. [39]

    Expel: LLM agents are experiential learners

    Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. Expel: LLM agents are experiential learners. In Michael J. Wooldridge, Jennifer G. Dy, and Sriraam Natarajan, editors, Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty- Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 20...

    work page 2024

  40. [40]

    Moatless tools

    Albert Örwall. Moatless tools. https://github.com/aorwall/moatless-tools, June 2024. A Appendix A.1 Implementation Test and Patch Generation. For each issue instance, EXPE REPAIR first uses the test agent to generate a reproduction script capable of triggering the issue. We set an iteration limit of 3 for script generation, producing at most one reproduct...

    work page 2024