Multi-agent Collaboration with State Management

Mengyang Liu; Taozhi Chen; Xue Jiang; Yihong Dong; Zhenhua Xu

arxiv: 2605.20563 · v1 · pith:H5YVKXMPnew · submitted 2026-05-19 · 💻 cs.MA · cs.AI· cs.CL· cs.LG· cs.SE

Multi-agent Collaboration with State Management

Mengyang Liu , Taozhi Chen , Zhenhua Xu , Xue Jiang , Yihong Dong This is my paper

Pith reviewed 2026-05-21 05:56 UTC · model grok-4.3

classification 💻 cs.MA cs.AIcs.CLcs.LGcs.SE

keywords multi-agent collaborationstate managementconflict resolutionshared workspaceLLM agentsconcurrent editingbenchmark evaluation

0 comments

The pith

Explicit state management outperforms workspace isolation for multi-agent collaboration

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

The paper proposes that managing the states of multiple agents directly, by controlling their access to a shared workspace, prevents silent conflicts that arise when agents edit code concurrently. Instead of isolating each agent in its own workspace and merging later, STORM detects and fixes conflicts immediately when an agent tries to write. This leads to better results on benchmarks for code commits and paper writing tasks, with gains of over 18 points on one and small gains on the other. A sympathetic reader would care because many AI systems now use teams of agents to solve big tasks, and fixing integration problems early could make those teams much more reliable.

Core claim

STORM manages agent states by mediating their interactions with the shared workspace, ensuring that each agent operates on a consistent view of the codebase and that conflicting edits are detected and resolved at write time, which outperforms the git-worktree-based baseline by +18.7 on Commit0-Lite and +1.4 on PaperBench.

What carries the argument

The state-oriented mediator in STORM that tracks agent interactions with the shared workspace to enforce consistent views and resolve conflicts at the point of writing.

If this is right

Conflicts are resolved during the write operation rather than through expensive post-hoc merges after agents complete their work.
STORM integrates into any existing multi-agent system without requiring changes to the agents' internal logic.
Higher overall task success rates are achieved on concurrent editing benchmarks when state management replaces workspace isolation.
Combining the multi-agent state-managed runs with single-agent executions produces the highest scores on the evaluated tasks.

Where Pith is reading between the lines

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

The same mediation approach could extend to non-code shared resources, such as multi-agent updates to a common knowledge base or simulation state.
Testing on larger real-world repositories would show whether the early conflict resolution scales without introducing new bottlenecks.
The reported cost efficiency opens the possibility of using STORM in resource-constrained deployments where repeated merges would otherwise waste compute.

Load-bearing premise

The benchmarks Commit0 and PaperBench, along with the specific implementation of conflict detection and resolution, are representative of real multi-agent collaboration scenarios and that performance gains are attributable to the state management mechanism.

What would settle it

A direct comparison on a new codebase with deliberately introduced concurrent edit conflicts, measuring whether integration failure rates drop under state mediation versus post-hoc workspace merging.

Figures

Figures reproduced from arXiv: 2605.20563 by Mengyang Liu, Taozhi Chen, Xue Jiang, Yihong Dong, Zhenhua Xu.

**Figure 2.** Figure 2: Scaling engineers with Sonnet 4.6 on Commit0-Lite. (a) Both test-based and repo-based [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Analysis summary on Sonnet 4.6. (a) STORM surfaces conflicts pre-commit, while [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Paired run timelines on jinja; the legend in (a) applies to both panels. (a) GitWorktree exposes the utils.py coupling only at merge review: the red diamond marks the rejected review, the shaded band the rework window, and Focus R2 the retry that is eventually accepted. (b) STORM detects the same coupling at decomposition time and co-assigns or sequences the dependent tasks (gold-outlined manager instructi… view at source ↗

**Figure 5.** Figure 5: Failure analysis for STORM runs at k=4 and k=8. Left: average failed-test symptom mix per run, decomposed into assertion/semantic, missing API/symbol, type/contract, not-implemented, and other runtime failures. Right: share of failed runs in which each run-level cause proxy fires, covering incomplete API, scope drift, budget/runtime, and accepted same-file overlap [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗

read the original abstract

Recent advances in multi-agent systems have shown great potential for solving complex tasks. However, when multiple agents edit a shared codebase concurrently, their changes can silently conflict and inconsistent views lead to integration failures. Existing multi-agent systems address this through workspace isolation (e.g., one git worktree per agent), but this defers conflict resolution to a post-hoc merge step where recovery is expensive. In this paper, we propose STORM, i.e., STate-ORiented Management for multi-agent collaboration. Specifically, STORM manages agent states by mediating their interactions with the shared workspace, ensuring that each agent operates on a consistent view of the codebase and that conflicting edits are detected and resolved at write time. We evaluate STORM on Commit0 and PaperBench across multiple LLMs. STORM outperforms the git-worktree-based multi-agent baseline by +18.7 on Commit0-Lite and +1.4 on PaperBench, while achieving comparable or better cost efficiency. Combined with single-agent runs, STORM reaches highest scores of 87.6 and 78.2 on the two benchmarks respectively, suggesting that explicit state management is a more effective foundation for multi-agent collaboration than workspace isolation. STORM can also be plugged into any multi-agent system seamlessly.

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.

Desk Editor's Note private letter to a colleague

STORM reports benchmark gains for multi-agent code tasks via write-time state mediation instead of isolation-plus-merge, but the causal role of the state abstraction itself is not clearly isolated from implementation details.

read the letter

STORM's core move is to insert a mediation layer that keeps agents on a consistent shared view and catches edit conflicts at the moment of writing rather than after the fact. That is the concrete difference from the git-worktree baseline described in the abstract, and the paper shows it can be dropped into existing multi-agent setups without rewriting the agents themselves. On Commit0-Lite and PaperBench the numbers are +18.7 and +1.4 over the isolation baseline, with comparable or better cost, which is the kind of result that could matter for people actually wiring these systems together for code or document work. Combined with single-agent runs it reaches the highest scores reported, 87.6 and 78.2. That is the practical upside worth noting. The experiments use multiple LLMs, so the gains are not tied to one model family. The plug-in claim is also useful if it holds up in the full implementation. The main soft spot is the missing ablation that would hold conflict detection and resolution logic fixed while switching only between state mediation and workspace isolation. Without that, the observed deltas could come from how STORM happens to implement merging or locking rather than from the architectural choice itself. The abstract also gives no error bars, run counts, or exclusion rules, so it is hard to judge how stable the improvements are. The benchmarks are narrow, so generalization to other collaboration settings remains open. This is the sort of paper that belongs in a reading group for people working on agent coordination in software engineering. A reader who wants a concrete alternative to post-hoc merging will get something usable from the results and the integration story. It deserves peer review because the problem is real, the baseline comparison is direct, and the numbers are large enough to be worth checking in detail even if the causal story needs tightening.

Referee Report

2 major / 1 minor

Summary. The paper introduces STORM (STate-ORiented Management), a framework for multi-agent collaboration on shared codebases. Unlike workspace isolation approaches such as per-agent git worktrees, STORM mediates agent interactions with a shared workspace to maintain consistent views and detect/resolve conflicts at write time. Evaluated on Commit0 and PaperBench across multiple LLMs, STORM outperforms a git-worktree baseline by +18.7 points on Commit0-Lite and +1.4 on PaperBench, reaching combined highest scores of 87.6 and 78.2. The central claim is that explicit state management provides a more effective foundation for multi-agent collaboration than workspace isolation, and that STORM can be plugged into existing multi-agent systems.

Significance. If the performance deltas are causally due to the state-management abstraction, the work could shift design patterns in multi-agent coding systems toward shared consistent state rather than deferred merges. The plug-in compatibility is a practical advantage for adoption. The empirical nature of the contribution means significance hinges on verification that gains are not artifacts of benchmark choice or unablated implementation details.

major comments (2)

Abstract: the central performance claim (+18.7 on Commit0-Lite, +1.4 on PaperBench) is presented without any mention of experimental controls, number of trials, statistical tests, error bars, or data exclusion rules. This directly undermines verification of whether the reported gains support the claim that state management outperforms workspace isolation.
Evaluation / baseline comparison: the manuscript contrasts STORM against a git-worktree baseline but provides no ablation that holds conflict detection, resolution, merging, and locking logic fixed while toggling only the isolation mechanism (shared consistent state vs. per-agent workspaces). Without this control, the observed deltas cannot be confidently attributed to the state-oriented architecture rather than differences in how STORM implements write-time mediation.

minor comments (1)

Abstract: the statement that STORM 'can also be plugged into any multi-agent system seamlessly' would benefit from a brief description or pseudocode of the integration interface in the main text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments. We address each major comment below, indicating where revisions will be made to strengthen the manuscript.

read point-by-point responses

Referee: Abstract: the central performance claim (+18.7 on Commit0-Lite, +1.4 on PaperBench) is presented without any mention of experimental controls, number of trials, statistical tests, error bars, or data exclusion rules. This directly undermines verification of whether the reported gains support the claim that state management outperforms workspace isolation.

Authors: We agree that the abstract would benefit from additional context on the experimental protocol. In the revised manuscript we will update the abstract to note that results are reported as averages over multiple independent trials with standard deviations, following standard statistical practices for the benchmarks. Full details on the number of runs, controls, error bars, and any data exclusion criteria already appear in the Evaluation section; the abstract revision will provide sufficient high-level information to support the reported deltas without requiring readers to consult the body for basic verification. revision: yes
Referee: Evaluation / baseline comparison: the manuscript contrasts STORM against a git-worktree baseline but provides no ablation that holds conflict detection, resolution, merging, and locking logic fixed while toggling only the isolation mechanism (shared consistent state vs. per-agent workspaces). Without this control, the observed deltas cannot be confidently attributed to the state-oriented architecture rather than differences in how STORM implements write-time mediation.

Authors: We acknowledge the value of a more tightly controlled ablation. The git-worktree baseline implements the standard per-agent isolation approach used in prior multi-agent coding systems, without STORM's shared-state mediation. To address the referee's concern, we will add a new ablation experiment in the revision that attempts to hold conflict detection, resolution, and locking logic as constant as possible while varying only the workspace isolation mechanism. We note that complete decoupling may introduce implementation artifacts, but the added study will help clarify the contribution of the shared consistent state. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical results are direct measurements

full rationale

The paper proposes the STORM system for explicit state management in multi-agent collaboration and reports performance deltas (+18.7 on Commit0-Lite, +1.4 on PaperBench) from direct experimental comparison against a git-worktree baseline. No derivation chain, fitted parameters renamed as predictions, or self-citation load-bearing steps exist; the central claim follows from benchmark outcomes on external tasks without reducing to inputs by construction. The evaluation is self-contained against the stated benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; the approach is described at the level of system design and benchmark outcomes.

pith-pipeline@v0.9.0 · 5767 in / 1115 out tokens · 39368 ms · 2026-05-21T05:56:53.281646+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

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

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

A write (ai, f, c′) is valid if and only if the agent’s local state is still consistent with the current workspace: ∀(g, v_obs_g)∈S_i : v_obs_g = v_cur_g.
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean LogicNat recovery theorem unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

STORM mediates each agent’s file reads and writes... inspired by optimistic concurrency control (Kung and Robinson, 1981).

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.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages

[1]

Chen Qian and Wei Liu and Hongzhang Liu and Nuo Chen and Yufan Dang and Jiahao Li and Cheng Yang and Weize Chen and Yusheng Su and Xin Cong and Juyuan Xu and Dahai Li and Zhiyuan Liu and Maosong Sun , title =

work page
[2]

MetaGPT: Meta Programming for

Sirui Hong and Mingchen Zhuge and Jonathan Chen and Xiawu Zheng and Yuheng Cheng and Jinlin Wang and Ceyao Zhang and Zili Wang and Steven Ka Shing Yau and Zijuan Lin and Liyang Zhou and Chenyu Ran and Lingfeng Xiao and Chenglin Wu and J. MetaGPT: Meta Programming for

work page
[3]

Giulio Starace and Oliver Jaffe and Dane Sherburn and James Aung and Jun Shern Chan and Leon Maksin and Rachel Dias and Evan Mays and Benjamin Kinsella and Wyatt Thompson and Johannes Heidecke and Amelia Glaese and Tejal Patwardhan , title =

work page
[4]

CoRR , volume =

Qingyun Wu and Gagan Bansal and Jieyu Zhang and Yiran Wu and Shaokun Zhang and Erkang Zhu and Beibin Li and Li Jiang and Xiaoyun Zhang and Chi Wang , title =. CoRR , volume =

work page
[5]

NeurIPS , year =

Wei Tao and Yucheng Zhou and Yanlin Wang and Wenqiang Zhang and Hongyu Zhang and Yu Cheng , title =. NeurIPS , year =

work page
[6]

CoRR , volume =

Han Li and Yuling Shi and Shaoxin Lin and Xiaodong Gu and Heng Lian and Xin Wang and Yantao Jia and Tao Huang and Qianxiang Wang , title =. CoRR , volume =

work page
[7]

2026 , eprint=

AgentForge: Execution-Grounded Multi-Agent LLM Framework for Autonomous Software Engineering , author=. 2026 , eprint=

work page 2026
[8]

2026 , eprint=

Effective Strategies for Asynchronous Software Engineering Agents , author=. 2026 , eprint=

work page 2026
[9]

2026 , eprint=

CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery , author=. 2026 , eprint=

work page 2026
[10]

2026 , eprint=

StatsClaw: An AI-Collaborative Workflow for Statistical Software Development , author=. 2026 , eprint=

work page 2026
[11]

Yihong Dong and Xue Jiang and Zhi Jin and Ge Li , title =

work page
[12]

2025 , eprint=

A Survey on Code Generation with LLM-based Agents , author=. 2025 , eprint=

work page 2025
[13]

Xue Jiang and Yihong Dong and Lecheng Wang and Zheng Fang and Qiwei Shang and Ge Li and Zhi Jin and Wenpin Jiao , title =

work page
[14]

NeurIPS , year =

Noah Shinn and Federico Cassano and Ashwin Gopinath and Karthik Narasimhan and Shunyu Yao , title =. NeurIPS , year =

work page
[15]

Patil and Kevin Lin and Sarah Wooders and Joseph E

Charles Packer and Vivian Fang and Shishir G. Patil and Kevin Lin and Sarah Wooders and Joseph E. Gonzalez , title =. CoRR , volume =

work page
[16]

Xu and Xiangru Tang and Mingchen Zhuge and Jiayi Pan and Yueqi Song and Bowen Li and Jaskirat Singh and Hoang H

Xingyao Wang and Boxuan Li and Yufan Song and Frank F. Xu and Xiangru Tang and Mingchen Zhuge and Jiayi Pan and Yueqi Song and Bowen Li and Jaskirat Singh and Hoang H. Tran and Fuqiang Li and Ren Ma and Mingzhang Zheng and Bill Qian and Yanjun Shao and Niklas Muennighoff and Yizhe Zhang and Binyuan Hui and Junyang Lin and et al. , title =

work page
[17]

RepoCoder: Repository-Level Code Completion Through Iterative Retrieval and Generation , booktitle =

Fengji Zhang and Bei Chen and Yue Zhang and Jacky Keung and Jin Liu and Daoguang Zan and Yi Mao and Jian. RepoCoder: Repository-Level Code Completion Through Iterative Retrieval and Generation , booktitle =

work page
[18]

CoRR , volume =

Disha Shrivastava and Denis Kocetkov and Harm de Vries and Dzmitry Bahdanau and Torsten Scholak , title =. CoRR , volume =

work page
[19]

NeurIPS , year =

Yangruibo Ding and Zijian Wang and Wasi Uddin Ahmad and Hantian Ding and Ming Tan and Nihal Jain and Murali Krishna Ramanathan and Ramesh Nallapati and Parminder Bhatia and Dan Roth and Bing Xiang , title =. NeurIPS , year =

work page
[20]

Yingwei Ma and Yongbin Li and Yihong Dong and Xue Jiang and Yanhao Li and Yue Liu and Rongyu Cao and Jue Chen and Fei Huang and Binhua Li , title =

work page
[21]

Jia Li and Ge Li and Yunfei Zhao and Yongmin Li and Huanyu Liu and Hao Zhu and Lecheng Wang and Kaibo Liu and Zheng Fang and Lanshen Wang and Jiazheng Ding and Xuanming Zhang and Yuqi Zhu and Yihong Dong and Zhi Jin and Binhua Li and Fei Huang and Yongbin Li and Bin Gu and Mengfei Yang , title =

work page
[22]

Yihong Dong and Jiazheng Ding and Xue Jiang and Ge Li and Zhuo Li and Zhi Jin , title =

work page
[23]

Guanzhi Wang and Yuqi Xie and Yunfan Jiang and Ajay Mandlekar and Chaowei Xiao and Yuke Zhu and Linxi Fan and Anima Anandkumar , title =. Trans. Mach. Learn. Res. , volume =

work page
[24]

Mengkang Hu and Tianxing Chen and Qiguang Chen and Yao Mu and Wenqi Shao and Ping Luo , title =

work page
[25]

O'Brien and Carrie Jun Cai and Meredith Ringel Morris and Percy Liang and Michael S

Joon Sung Park and Joseph C. O'Brien and Carrie Jun Cai and Meredith Ringel Morris and Percy Liang and Michael S. Bernstein , title =

work page
[26]

Chiu and Claire Cardie and Matthias Gall

Wenting Zhao and Nan Jiang and Celine Lee and Justin T. Chiu and Claire Cardie and Matthias Gall. Commit0: Library Generation from Scratch , booktitle =

work page
[27]

H. T. Kung and John T. Robinson , title =

work page
[28]

2026 , eprint=

MEMCoder: Multi-dimensional Evolving Memory for Private-Library-Oriented Code Generation , author=. 2026 , eprint=

work page 2026
[29]

2025 , eprint=

CodeCRDT: Observation-Driven Coordination for Multi-Agent LLM Code Generation , author=. 2025 , eprint=

work page 2025
[30]

2026 , month=jan, howpublished=

Wilson Lin , title=. 2026 , month=jan, howpublished=

work page 2026
[31]

Xue Jiang and Yihong Dong and Yongding Tao and Huanyu Liu and Zhi Jin and Ge Li , title =

work page
[32]

CoRR , volume =

Xue Jiang and Tianyu Zhang and Ge Li and Mengyang Liu and Taozhi Chen and Zhenhua Xu and Binhua Li and Wenpin Jiao and Zhi Jin and Yongbin Li and Yihong Dong , title =. CoRR , volume =

work page
[33]

2026 , eprint=

From I/O to Code with Discovery Agent , author=. 2026 , eprint=

work page 2026
[34]

Mission Control for AI Agents , howpublished =

work page

[1] [1]

Chen Qian and Wei Liu and Hongzhang Liu and Nuo Chen and Yufan Dang and Jiahao Li and Cheng Yang and Weize Chen and Yusheng Su and Xin Cong and Juyuan Xu and Dahai Li and Zhiyuan Liu and Maosong Sun , title =

work page

[2] [2]

MetaGPT: Meta Programming for

Sirui Hong and Mingchen Zhuge and Jonathan Chen and Xiawu Zheng and Yuheng Cheng and Jinlin Wang and Ceyao Zhang and Zili Wang and Steven Ka Shing Yau and Zijuan Lin and Liyang Zhou and Chenyu Ran and Lingfeng Xiao and Chenglin Wu and J. MetaGPT: Meta Programming for

work page

[3] [3]

Giulio Starace and Oliver Jaffe and Dane Sherburn and James Aung and Jun Shern Chan and Leon Maksin and Rachel Dias and Evan Mays and Benjamin Kinsella and Wyatt Thompson and Johannes Heidecke and Amelia Glaese and Tejal Patwardhan , title =

work page

[4] [4]

CoRR , volume =

Qingyun Wu and Gagan Bansal and Jieyu Zhang and Yiran Wu and Shaokun Zhang and Erkang Zhu and Beibin Li and Li Jiang and Xiaoyun Zhang and Chi Wang , title =. CoRR , volume =

work page

[5] [5]

NeurIPS , year =

Wei Tao and Yucheng Zhou and Yanlin Wang and Wenqiang Zhang and Hongyu Zhang and Yu Cheng , title =. NeurIPS , year =

work page

[6] [6]

CoRR , volume =

Han Li and Yuling Shi and Shaoxin Lin and Xiaodong Gu and Heng Lian and Xin Wang and Yantao Jia and Tao Huang and Qianxiang Wang , title =. CoRR , volume =

work page

[7] [7]

2026 , eprint=

AgentForge: Execution-Grounded Multi-Agent LLM Framework for Autonomous Software Engineering , author=. 2026 , eprint=

work page 2026

[8] [8]

2026 , eprint=

Effective Strategies for Asynchronous Software Engineering Agents , author=. 2026 , eprint=

work page 2026

[9] [9]

2026 , eprint=

CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery , author=. 2026 , eprint=

work page 2026

[10] [10]

2026 , eprint=

StatsClaw: An AI-Collaborative Workflow for Statistical Software Development , author=. 2026 , eprint=

work page 2026

[11] [11]

Yihong Dong and Xue Jiang and Zhi Jin and Ge Li , title =

work page

[12] [12]

2025 , eprint=

A Survey on Code Generation with LLM-based Agents , author=. 2025 , eprint=

work page 2025

[13] [13]

Xue Jiang and Yihong Dong and Lecheng Wang and Zheng Fang and Qiwei Shang and Ge Li and Zhi Jin and Wenpin Jiao , title =

work page

[14] [14]

NeurIPS , year =

Noah Shinn and Federico Cassano and Ashwin Gopinath and Karthik Narasimhan and Shunyu Yao , title =. NeurIPS , year =

work page

[15] [15]

Patil and Kevin Lin and Sarah Wooders and Joseph E

Charles Packer and Vivian Fang and Shishir G. Patil and Kevin Lin and Sarah Wooders and Joseph E. Gonzalez , title =. CoRR , volume =

work page

[16] [16]

Xu and Xiangru Tang and Mingchen Zhuge and Jiayi Pan and Yueqi Song and Bowen Li and Jaskirat Singh and Hoang H

Xingyao Wang and Boxuan Li and Yufan Song and Frank F. Xu and Xiangru Tang and Mingchen Zhuge and Jiayi Pan and Yueqi Song and Bowen Li and Jaskirat Singh and Hoang H. Tran and Fuqiang Li and Ren Ma and Mingzhang Zheng and Bill Qian and Yanjun Shao and Niklas Muennighoff and Yizhe Zhang and Binyuan Hui and Junyang Lin and et al. , title =

work page

[17] [17]

RepoCoder: Repository-Level Code Completion Through Iterative Retrieval and Generation , booktitle =

Fengji Zhang and Bei Chen and Yue Zhang and Jacky Keung and Jin Liu and Daoguang Zan and Yi Mao and Jian. RepoCoder: Repository-Level Code Completion Through Iterative Retrieval and Generation , booktitle =

work page

[18] [18]

CoRR , volume =

Disha Shrivastava and Denis Kocetkov and Harm de Vries and Dzmitry Bahdanau and Torsten Scholak , title =. CoRR , volume =

work page

[19] [19]

NeurIPS , year =

Yangruibo Ding and Zijian Wang and Wasi Uddin Ahmad and Hantian Ding and Ming Tan and Nihal Jain and Murali Krishna Ramanathan and Ramesh Nallapati and Parminder Bhatia and Dan Roth and Bing Xiang , title =. NeurIPS , year =

work page

[20] [20]

Yingwei Ma and Yongbin Li and Yihong Dong and Xue Jiang and Yanhao Li and Yue Liu and Rongyu Cao and Jue Chen and Fei Huang and Binhua Li , title =

work page

[21] [21]

Jia Li and Ge Li and Yunfei Zhao and Yongmin Li and Huanyu Liu and Hao Zhu and Lecheng Wang and Kaibo Liu and Zheng Fang and Lanshen Wang and Jiazheng Ding and Xuanming Zhang and Yuqi Zhu and Yihong Dong and Zhi Jin and Binhua Li and Fei Huang and Yongbin Li and Bin Gu and Mengfei Yang , title =

work page

[22] [22]

Yihong Dong and Jiazheng Ding and Xue Jiang and Ge Li and Zhuo Li and Zhi Jin , title =

work page

[23] [23]

Guanzhi Wang and Yuqi Xie and Yunfan Jiang and Ajay Mandlekar and Chaowei Xiao and Yuke Zhu and Linxi Fan and Anima Anandkumar , title =. Trans. Mach. Learn. Res. , volume =

work page

[24] [24]

Mengkang Hu and Tianxing Chen and Qiguang Chen and Yao Mu and Wenqi Shao and Ping Luo , title =

work page

[25] [25]

O'Brien and Carrie Jun Cai and Meredith Ringel Morris and Percy Liang and Michael S

Joon Sung Park and Joseph C. O'Brien and Carrie Jun Cai and Meredith Ringel Morris and Percy Liang and Michael S. Bernstein , title =

work page

[26] [26]

Chiu and Claire Cardie and Matthias Gall

Wenting Zhao and Nan Jiang and Celine Lee and Justin T. Chiu and Claire Cardie and Matthias Gall. Commit0: Library Generation from Scratch , booktitle =

work page

[27] [27]

H. T. Kung and John T. Robinson , title =

work page

[28] [28]

2026 , eprint=

MEMCoder: Multi-dimensional Evolving Memory for Private-Library-Oriented Code Generation , author=. 2026 , eprint=

work page 2026

[29] [29]

2025 , eprint=

CodeCRDT: Observation-Driven Coordination for Multi-Agent LLM Code Generation , author=. 2025 , eprint=

work page 2025

[30] [30]

2026 , month=jan, howpublished=

Wilson Lin , title=. 2026 , month=jan, howpublished=

work page 2026

[31] [31]

Xue Jiang and Yihong Dong and Yongding Tao and Huanyu Liu and Zhi Jin and Ge Li , title =

work page

[32] [32]

CoRR , volume =

Xue Jiang and Tianyu Zhang and Ge Li and Mengyang Liu and Taozhi Chen and Zhenhua Xu and Binhua Li and Wenpin Jiao and Zhi Jin and Yongbin Li and Yihong Dong , title =. CoRR , volume =

work page

[33] [33]

2026 , eprint=

From I/O to Code with Discovery Agent , author=. 2026 , eprint=

work page 2026

[34] [34]

Mission Control for AI Agents , howpublished =

work page