arxiv: 2604.01496 · v2 · submitted 2026-04-02 · 💻 cs.SE · cs.CL

Recognition: no theorem link

From SWE-ZERO to SWE-HERO: Execution-free to Execution-based Fine-tuning for Software Engineering Agents

Boris Ginsburg, Nikolai Ludwig, Somshubra Majumdar, Wasi Uddin Ahmad

Authors on Pith no claims yet

Pith reviewed 2026-05-13 21:40 UTC · model grok-4.3

classification 💻 cs.SE cs.CL

keywords SWE-benchsoftware engineering agentsfine-tuningexecution-based refinementdistillationopen-source modelszero-shot transferrepository-level reasoning

0 comments

The pith

A two-stage fine-tuning process converts execution-free code reasoning into execution-based engineering performance, letting 32B open models resolve 62 percent of SWE-bench tasks.

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

The paper presents a two-stage supervised fine-tuning recipe that first uses large-scale execution-free trajectories to master code semantics and repository-level reasoning, then applies targeted execution-backed refinement to turn those intuitions into reliable engineering workflows. This approach distills knowledge from a large frontier model into smaller open-weight models based on the Qwen2.5-Coder series without relying on resource-heavy execution during the initial phase. The resulting SWE-HERO models set new performance marks for open-source agents of comparable size, with the 32B variant reaching 62.2 percent resolution on SWE-bench Verified and 44.1 percent on the multilingual version despite Python-only training. The authors also release the 300k SWE-ZERO and 13k SWE-HERO trajectories to support further work.

Core claim

By first training on 300k execution-free trajectories to master code semantics and repository-level reasoning, then refining with 13k execution-backed trajectories through an evolutionary strategy, the resulting SWE-HERO models convert semantic intuitions into rigorous engineering workflows, with the 32B variant resolving 62.2 percent of tasks on SWE-bench Verified and 44.1 percent on the multilingual benchmark.

What carries the argument

The evolutionary refinement strategy that transitions execution-free semantic mastery into execution-based engineering behavior using targeted execution feedback.

If this is right

Open models of 32B parameters can reach competitive resolution rates on repository-level software engineering tasks through staged distillation.
Training exclusively on Python trajectories still supports strong zero-shot transfer to other programming languages.
The released trajectory datasets make the method reproducible and allow direct extension by other researchers.
Avoiding execution during the first training stage lowers the overall compute cost while preserving final performance.
The two-stage separation shows that semantic understanding can be built before adding execution constraints.

Where Pith is reading between the lines

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

The same staged pattern of first building broad understanding then adding execution feedback could apply to agent tasks outside software engineering, such as data analysis or scientific simulation.
Execution feedback appears most effective once semantic foundations are already present, which may guide data collection priorities in other domains.
Further increases in the number of execution-based trajectories or use of different base models could test whether the reported performance scales further.

Load-bearing premise

The 300k execution-free and 13k execution-based trajectories distilled from the frontier model contain high-quality examples free of systematic biases that the refinement process can reliably convert into correct engineering actions.

What would settle it

Testing the trained 32B agent on a fresh held-out portion of SWE-bench Verified problems and measuring whether the resolution rate stays at or above 62.2 percent when every edit is strictly verified by execution.

Figures

Figures reproduced from arXiv: 2604.01496 by Boris Ginsburg, Nikolai Ludwig, Somshubra Majumdar, Wasi Uddin Ahmad.

**Figure 2.** Figure 2: Comparative overview of Agentic Workflows in OpenHands. The bottom panel [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Cumulative performance and resource efficiency on SWE-bench Verified. Tasks [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Scaling behavior of SWE-ZERO-14B. The resolution rate on SWE-bench increases consistently as execution-free training samples scale from 4k to 150k. Data Scaling Laws in SWE-ZERO Training As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Increasing inference-time compute improves performance on SWE-Bench Verified [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗

**Figure 6.** Figure 6: Cumulative performance and resource efficiency on SWE-bench Verified. Tasks [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗

**Figure 7.** Figure 7: Detection of execution in the execute_bash command. This Python implementation leverages bashlex to generate an Abstract Syntax Tree (AST), ensuring all invoked commands are validated against a predefined WHITELIST. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗

read the original abstract

We introduce SWE-ZERO to SWE-HERO, a two-stage SFT recipe that achieves state-of-the-art results on SWE-bench by distilling open-weight frontier LLMs. Our pipeline replaces resource-heavy dependencies with an evolutionary refinement strategy: (1) SWE-ZERO utilizes large-scale, execution-free trajectories to master code semantics and repository-level reasoning, and (2) SWE-HERO applies targeted, execution-backed refinement to transition these semantic intuitions into rigorous engineering workflows. Our empirical results set a new benchmark for open-source models of comparable size. We release a dataset of 300k SWE-ZERO and 13k SWE-HERO trajectories distilled from Qwen3-Coder-480B, alongside a suite of agents based on the Qwen2.5-Coder series. Notably, SWE-HERO-32B achieves a 62.2% resolution rate on SWE-bench Verified. Furthermore, despite being trained exclusively on Python, our agents demonstrate robust zero-shot transferability on SWE-bench Multilingual, reaching 44.1% and confirming the paradigm's generalizability across diverse languages.

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

The two-stage distillation from execution-free trajectories to execution-based refinement gets a 32B open model to 62% on SWE-bench Verified with multilingual transfer, but the paper does not audit errors in the teacher data or ablate the first stage.

read the letter

The core result is a practical two-stage SFT pipeline that first uses 300k execution-free trajectories to build semantic and repository-level understanding, then refines with 13k execution-backed ones to reach 62.2% resolution on SWE-bench Verified and 44.1% zero-shot on the multilingual version. The release of both trajectory sets from Qwen3-Coder-480B is the most immediately useful part for anyone trying to replicate or extend this kind of agent training.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces a two-stage supervised fine-tuning pipeline called SWE-ZERO to SWE-HERO for software engineering agents. SWE-ZERO distills 300k execution-free trajectories from Qwen3-Coder-480B to instill code semantics and repository-level reasoning, while SWE-HERO applies evolutionary refinement on 13k execution-based trajectories to convert these into reliable engineering workflows. The authors report that the resulting SWE-HERO-32B model reaches 62.2% resolution on SWE-bench Verified and 44.1% zero-shot on SWE-bench Multilingual despite Python-only training data, claiming new state-of-the-art results among open-source models of comparable size. They also release the full set of 300k + 13k trajectories.

Significance. If the results hold after addressing the validation gaps, the work would provide a resource-efficient route to high-performing SE agents by deferring execution costs until the refinement stage. The public release of the 300k SWE-ZERO and 13k SWE-HERO trajectories is a clear strength that supplies the community with large-scale, distilled data for further experimentation. The zero-shot multilingual transfer result further suggests that the learned behaviors may generalize beyond the training language.

major comments (2)

[§4 (Results)] The headline 62.2% resolution rate on SWE-bench Verified (Abstract and §4) is presented without an ablation that trains a comparable model directly on the 13k execution-based trajectories alone. This comparison is required to isolate whether the SWE-ZERO stage contributes beyond what targeted execution-based SFT can achieve on its own.
[§3.1 (SWE-ZERO Trajectory Generation)] No quantitative audit or residual-error analysis is reported for the 300k execution-free trajectories generated in the SWE-ZERO stage (§3.1). Because these trajectories are never execution-verified, it remains unclear whether the subsequent evolutionary refinement reliably corrects teacher-model semantic or repository-level mistakes, which is load-bearing for the pipeline's central claim.

minor comments (1)

[Abstract] The abstract claims 'state-of-the-art results' for open-source models but does not list the exact prior scores or model sizes being surpassed; a one-line reference to the strongest baseline in §4 would improve immediate readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our two-stage SWE-ZERO to SWE-HERO pipeline. The comments highlight important aspects for strengthening the empirical claims, and we address each point below with commitments to revisions where feasible.

read point-by-point responses

Referee: [§4 (Results)] The headline 62.2% resolution rate on SWE-bench Verified (Abstract and §4) is presented without an ablation that trains a comparable model directly on the 13k execution-based trajectories alone. This comparison is required to isolate whether the SWE-ZERO stage contributes beyond what targeted execution-based SFT can achieve on its own.

Authors: We agree that this ablation would strengthen the isolation of the SWE-ZERO stage's contribution. Our design is motivated by the observation that large-scale execution-free trajectories provide foundational semantic and repository-level priors that enable more effective learning from the smaller set of execution-based trajectories; direct SFT on 13k complex trajectories alone risks underfitting due to limited coverage. To address the concern directly, we will add this ablation in the revised manuscript by training a Qwen2.5-Coder-32B baseline solely on the 13k SWE-HERO trajectories and reporting its SWE-bench Verified performance alongside the full pipeline results. revision: yes
Referee: [§3.1 (SWE-ZERO Trajectory Generation)] No quantitative audit or residual-error analysis is reported for the 300k execution-free trajectories generated in the SWE-ZERO stage (§3.1). Because these trajectories are never execution-verified, it remains unclear whether the subsequent evolutionary refinement reliably corrects teacher-model semantic or repository-level mistakes, which is load-bearing for the pipeline's central claim.

Authors: We acknowledge the value of a quantitative audit for transparency on the unverified 300k trajectories. A full audit across all 300k is impractical at this scale, but the evolutionary refinement in SWE-HERO is explicitly designed to use execution feedback to detect and correct semantic and repository-level errors from the teacher. In the revision, we will add a sample-based residual-error analysis: we will randomly sample 500 SWE-ZERO trajectories, categorize common error types (semantic misunderstandings, incorrect repository navigation, etc.), and demonstrate via the corresponding refined SWE-HERO trajectories how execution-based evolution resolves a substantial fraction of these issues. revision: partial

Circularity Check

0 steps flagged

No circularity in empirical distillation pipeline

full rationale

The paper describes a two-stage empirical fine-tuning procedure that generates trajectories from an external frontier model (Qwen3-Coder-480B), applies execution-free then execution-based SFT to smaller Qwen2.5-Coder models, and reports resolution rates on the independent SWE-bench Verified and Multilingual benchmarks. No equations, self-definitional relations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the derivation chain. All performance numbers are produced by standard supervised fine-tuning followed by external evaluation; the pipeline remains falsifiable against held-out test sets and does not reduce any claimed result to a quantity defined by its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim depends on the assumption that trajectories generated by the 480B teacher model are high-quality and that the evolutionary refinement step reliably converts semantic knowledge into executable engineering skill.

axioms (1)

domain assumption Distilled trajectories from Qwen3-Coder-480B provide sufficiently accurate and diverse supervision for both stages
Invoked when stating that the pipeline masters code semantics and transitions to rigorous workflows

pith-pipeline@v0.9.0 · 5507 in / 1219 out tokens · 32073 ms · 2026-05-13T21:40:44.305334+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages · 9 internal anchors

[1]

Claude 4 sonnet

Anthropic . Claude 4 sonnet. https://www.anthropic.com/claude/sonnet, 2025 a . Accessed: 2025-08-31

work page 2025
[2]

Introducing claude opus 4.5

Anthropic . Introducing claude opus 4.5. https://www.anthropic.com/news/claude-opus-4-5, November 2025 b . Accessed: 2026-03-17

work page 2025
[3]

SWE -rebench: An automated pipeline for task collection and decontaminated evaluation of software engineering agents

Ibragim Badertdinov, Alexander Golubev, Maksim Nekrashevich, Anton Shevtsov, Simon Karasik, Andrei Andriushchenko, Maria Trofimova, Daria Litvintseva, and Boris Yangel. SWE -rebench: An automated pipeline for task collection and decontaminated evaluation of software engineering agents. In The Thirty-ninth Annual Conference on Neural Information Processing...

work page 2025
[4]

Qwen3-coder-next technical report.arXiv preprint arXiv:2603.00729, 2026

Ruisheng Cao, Mouxiang Chen, Jiawei Chen, Zeyu Cui, Yunlong Feng, Binyuan Hui, Yuheng Jing, Kaixin Li, Mingze Li, Junyang Lin, Zeyao Ma, Kashun Shum, Xuwu Wang, Jinxi Wei, Jiaxi Yang, Jiajun Zhang, Lei Zhang, Zongmeng Zhang, Wenting Zhao, and Fan Zhou. Qwen3-coder-next technical report, 2026. URL https://arxiv.org/abs/2603.00729

work page arXiv 2026
[5]

Skyrl-agent: Efficient rl training for multi-turn llm agent.arXiv preprint arXiv:2511.16108, 2025

Shiyi Cao, Dacheng Li, Fangzhou Zhao, Shuo Yuan, Sumanth R. Hegde, Connor Chen, Charlie Ruan, Tyler Griggs, Shu Liu, Eric Tang, Richard Liaw, Philipp Moritz, Matei Zaharia, Joseph E. Gonzalez, and Ion Stoica. Skyrl-agent: Efficient rl training for multi-turn llm agent, 2025. URL https://arxiv.org/abs/2511.16108

work page arXiv 2025
[6]

Cwm: An open-weights llm for research on code generation with world models

Jade Copet, Quentin Carbonneaux, Gal Cohen, Jonas Gehring, Jacob Kahn, Jannik Kossen, Felix Kreuk, Emily McMilin, Michel Meyer, Yuxiang Wei, et al. Cwm: An open-weights llm for research on code generation with world models. arXiv preprint arXiv:2510.02387, 2025

work page arXiv 2025
[7]

DeepSeek-AI, Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenhao Xu, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Erhang Li, Fangqi Zhou, Fangyun Lin, Fucong Dai, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Hanwei Xu, Ha...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[8]

davinci-env: Open swe environment synthesis at scale, 2026

Dayuan Fu, Shenyu Wu, Yunze Wu, Zerui Peng, Yaxing Huang, Jie Sun, Ji Zeng, Mohan Jiang, Lin Zhang, Yukun Li, Jiarui Hu, Liming Liu, Jinlong Hou, and Pengfei Liu. davinci-env: Open swe environment synthesis at scale, 2026. URL https://arxiv.org/abs/2603.13023

work page arXiv 2026
[9]

GLM-5: from Vibe Coding to Agentic Engineering

GLM-5-Team, :, Aohan Zeng, Xin Lv, Zhenyu Hou, Zhengxiao Du, Qinkai Zheng, Bin Chen, Da Yin, Chendi Ge, Chenghua Huang, Chengxing Xie, Chenzheng Zhu, Congfeng Yin, Cunxiang Wang, Gengzheng Pan, Hao Zeng, Haoke Zhang, Haoran Wang, Huilong Chen, Jiajie Zhang, Jian Jiao, Jiaqi Guo, Jingsen Wang, Jingzhao Du, Jinzhu Wu, Kedong Wang, Lei Li, Lin Fan, Lucen Zho...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[10]

Training long-context, multi-turn software engineering agents with reinforcement learning, 2025

Alexander Golubev, Maria Trofimova, Sergei Polezhaev, Ibragim Badertdinov, Maksim Nekrashevich, Anton Shevtsov, Simon Karasik, Sergey Abramov, Andrei Andriushchenko, Filipp Fisin, Sergei Skvortsov, and Boris Yangel. Training long-context, multi-turn software engineering agents with reinforcement learning, 2025. URL https://arxiv.org/abs/2508.03501

work page arXiv 2025
[11]

Gemini 3 flash: frontier intelligence built for speed

Google . Gemini 3 flash: frontier intelligence built for speed. https://blog.google/products-and-platforms/products/gemini/gemini-3-flash/, December 2025

work page 2025
[12]

Swe-factory: Your automated factory for issue resolution training data and evaluation benchmarks, 2026

Lianghong Guo, Yanlin Wang, Caihua Li, Wei Tao, Pengyu Yang, Jiachi Chen, Haoyu Song, Duyu Tang, and Zibin Zheng. Swe-factory: Your automated factory for issue resolution training data and evaluation benchmarks, 2026. URL https://arxiv.org/abs/2506.10954

work page arXiv 2026
[13]

SWE-Swiss : A multi-task fine-tuning and RL recipe for high-performance issue resolution

Zhenyu He, Qingping Yang, Wei Sheng, Xiaojian Zhong, Kechi Zhang, Chenxin An, Wenlei Shi, Tianle Cai, Di He, Jiaze Chen, and Jingjing Xu. SWE-Swiss : A multi-task fine-tuning and RL recipe for high-performance issue resolution. Notion Blog / GitHub Repository, 2025. URL https://github.com/zhenyuhe00/SWE-Swiss. Accessed: 2026-03-17

work page 2025
[14]

Qwen2.5-Coder Technical Report

Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, Kai Dang, Yang Fan, Yichang Zhang, An Yang, Rui Men, Fei Huang, Bo Zheng, Yibo Miao, Shanghaoran Quan, Yunlong Feng, Xingzhang Ren, Xuancheng Ren, Jingren Zhou, and Junyang Lin. Qwen2.5-coder technical report, 2024. URL https://arxiv.org/a...

work page internal anchor Pith review Pith/arXiv arXiv 2024
[15]

R2e-gym: Procedural environment generation and hybrid verifiers for scaling open-weights SWE agents

Naman Jain, Jaskirat Singh, Manish Shetty, Tianjun Zhang, Liang Zheng, Koushik Sen, and Ion Stoica. R2e-gym: Procedural environment generation and hybrid verifiers for scaling open-weights SWE agents. In Second Conference on Language Modeling, 2025. URL https://openreview.net/forum?id=7evvwwdo3z

work page 2025
[16]

SWE -bench: Can language models resolve real-world github issues? In The Twelfth International Conference on Learning Representations, 2024

Carlos E Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik R Narasimhan. SWE -bench: Can language models resolve real-world github issues? In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=VTF8yNQM66

work page 2024
[17]

mini-swe-agent-plus: The 100-line AI agent that solves GitHub issues with text-edit tool

Kwai-Klear . mini-swe-agent-plus: The 100-line AI agent that solves GitHub issues with text-edit tool. https://github.com/Kwai-Klear/mini-swe-agent-plus, 2025. GitHub repository

work page 2025
[18]

M. Luo, N. Jain, J. Singh, S. Tan, A. Patel, Q. Wu, A. Ariyak, C. Cai, T. Venkat, S. Zhu, B. Athiwaratkun, M. Roongta, C. Zhang, L. E. Li, R. A. Popa, K. Sen, and I. Stoica. DeepSWE : Training a fully open-sourced, state-of-the-art coding agent by scaling RL . https://www.together.ai/blog/deepswe, Jul. 2025. Together AI Blog post. Accessed: 2025-12-22

work page 2025
[19]

Swe-gpt: A process-centric language model for automated software improvement

Yingwei Ma, Rongyu Cao, Yongchang Cao, Yue Zhang, Jue Chen, Yibo Liu, Yuchen Liu, Binhua Li, Fei Huang, and Yongbin Li. Swe-gpt: A process-centric language model for automated software improvement. Proceedings of the ACM on Software Engineering, 2 0 (ISSTA): 0 2362--2383, 2025

work page 2025
[20]

Minimax m2.5: Built for real-world productivity

MiniMax . Minimax m2.5: Built for real-world productivity. https://www.minimax.io/news/minimax-m25, February 2026. Accessed: 2026-03-17

work page 2026
[21]

Introducing GPT-5.2

OpenAI . Introducing GPT-5.2 . https://openai.com/index/introducing-gpt-5-2/, December 2025

work page 2025
[22]

Training software engineering agents and verifiers with SWE -gym

Jiayi Pan, Xingyao Wang, Graham Neubig, Navdeep Jaitly, Heng Ji, Alane Suhr, and Yizhe Zhang. Training software engineering agents and verifiers with SWE -gym. In Forty-second International Conference on Machine Learning, 2025. URL https://openreview.net/forum?id=Cq1BNvHx74

work page 2025
[23]

YaRN: Efficient Context Window Extension of Large Language Models

Bowen Peng, Jeffrey Quesnelle, Honglu Fan, and Enrico Shippole. Yarn: Efficient context window extension of large language models, 2026. URL https://arxiv.org/abs/2309.00071

work page internal anchor Pith review Pith/arXiv arXiv 2026
[24]

Minh V. T. Pham, Huy N. Phan, Hoang N. Phan, Cuong Le Chi, Tien N. Nguyen, and Nghi D. Q. Bui. Swe-synth: Synthesizing verifiable bug-fix data to enable large language models in resolving real-world bugs, 2025. URL https://arxiv.org/abs/2504.14757

work page arXiv 2025
[25]

Qwen3.5 : Towards native multimodal agents, February 2026

Qwen Team . Qwen3.5 : Towards native multimodal agents, February 2026. URL https://qwen.ai/blog?id=qwen3.5. Accessed: 2026-03-17

work page 2026
[26]

Sera: Soft-verified efficient repository agents, 2026

Ethan Shen, Danny Tormoen, Saurabh Shah, Ali Farhadi, and Tim Dettmers. Sera: Soft-verified efficient repository agents, 2026. URL https://arxiv.org/abs/2601.20789

work page arXiv 2026
[27]

R1-Searcher: Incentivizing the Search Capability in LLMs via Reinforcement Learning

Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. R1-searcher: Incentivizing the search capability in llms via reinforcement learning, 2025. URL https://arxiv.org/abs/2503.05592

work page internal anchor Pith review Pith/arXiv arXiv 2025
[28]

Shuang Sun, Huatong Song, Lisheng Huang, Jinhao Jiang, Ran Le, Zhihao Lv, Zongchao Chen, Yiwen Hu, Wenyang Luo, Wayne Xin Zhao, Yang Song, Hongteng Xu, Tao Zhang, and Ji-Rong Wen

Huatong Song, Lisheng Huang, Shuang Sun, Jinhao Jiang, Ran Le, Daixuan Cheng, Guoxin Chen, Yiwen Hu, Zongchao Chen, Yiming Jia, Wayne Xin Zhao, Yang Song, Tao Zhang, and Ji-Rong Wen. Swe-master: Unleashing the potential of software engineering agents via post-training, 2026. URL https://arxiv.org/abs/2602.03411

work page arXiv 2026
[29]

Bugpilot: Complex bug generation for efficient learning of swe skills, 2025

Atharv Sonwane, Isadora White, Hyunji Lee, Matheus Pereira, Lucas Caccia, Minseon Kim, Zhengyan Shi, Chinmay Singh, Alessandro Sordoni, Marc-Alexandre Côté, and Xingdi Yuan. Bugpilot: Complex bug generation for efficient learning of swe skills, 2025. URL https://arxiv.org/abs/2510.19898

work page arXiv 2025
[30]

Swe-world: Building software engineering agents in docker-free environments, 2026

Shuang Sun, Huatong Song, Lisheng Huang, Jinhao Jiang, Ran Le, Zhihao Lv, Zongchao Chen, Yiwen Hu, Wenyang Luo, Wayne Xin Zhao, Yang Song, Hongteng Xu, Tao Zhang, and Ji-Rong Wen. Swe-world: Building software engineering agents in docker-free environments, 2026. URL https://arxiv.org/abs/2602.03419

work page arXiv 2026
[31]

Swe-lego: Pushing the limits of supervised fine-tuning for software issue resolving, 2026

Chaofan Tao, Jierun Chen, Yuxin Jiang, Kaiqi Kou, Shaowei Wang, Ruoyu Wang, Xiaohui Li, Sidi Yang, Yiming Du, Jianbo Dai, Zhiming Mao, Xinyu Wang, Lifeng Shang, and Haoli Bai. Swe-lego: Pushing the limits of supervised fine-tuning for software issue resolving, 2026. URL https://arxiv.org/abs/2601.01426

work page arXiv 2026
[32]

MiMo-V2-Flash Technical Report

Core Team, Bangjun Xiao, Bingquan Xia, Bo Yang, Bofei Gao, Bowen Shen, Chen Zhang, Chenhong He, Chiheng Lou, Fuli Luo, Gang Wang, Gang Xie, Hailin Zhang, Hanglong Lv, Hanyu Li, Heyu Chen, Hongshen Xu, Houbin Zhang, Huaqiu Liu, Jiangshan Duo, Jianyu Wei, Jiebao Xiao, Jinhao Dong, Jun Shi, Junhao Hu, Kainan Bao, Kang Zhou, Lei Li, Liang Zhao, Linghao Zhang,...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[33]

Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, S. H. Cai, Yuan Cao, Y. Charles, H. S. Che, Cheng Chen, Guanduo Chen, Huarong Chen, Jia Chen, Jiahao Chen, Jianlong Chen, Jun Chen, Kefan Chen, Liang Chen, Ruijue Chen, Xinhao Chen, Yanru Chen, Yanxu Chen, Yicun Chen, Yimin Chen, Yingjiang Chen, Yuankun Chen, Yujie Chen, Yutian Chen, Zhirong Chen, Ziwei Chen...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[34]

Kimi Team, Yifan Bai, Yiping Bao, Y. Charles, Cheng Chen, Guanduo Chen, Haiting Chen, Huarong Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, Zhuofu Chen, Jialei Cui, Hao Ding, Mengnan Dong, Angang Du, Chenzhuang Du, Dikang Du, Yulun Du, Yu Fan, Yichen Feng, Kelin Fu, Bofei Gao, Chenxiao Gao, Hongcheng Gao, Peizhong Ga...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[35]

SWE -dev: Building software engineering agents with training and inference scaling

Haoran Wang, Zhenyu Hou, Yao Wei, Jie Tang, and Yuxiao Dong. SWE -dev: Building software engineering agents with training and inference scaling. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar (eds.), Findings of the Association for Computational Linguistics: ACL 2025, pp.\ 3742--3761, Vienna, Austria, July 2025 a . Associat...

work page doi:10.18653/v1/2025.findings-acl.193 2025
[36]

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

Xingyao Wang, Boxuan Li, Yufan Song, Frank F. Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, Hoang H. Tran, Fuqiang Li, Ren Ma, Mingzhang Zheng, Bill Qian, Yanjun Shao, Niklas Muennighoff, Yizhe Zhang, Binyuan Hui, Junyang Lin, Robert Brennan, Hao Peng, Heng Ji, and Graham Neubig. Openhands: An open platform for AI soft...

work page 2025
[37]

Demystifying llm-based software engineering agents

Chunqiu Steven Xia, Yinlin Deng, Soren Dunn, and Lingming Zhang. Demystifying llm-based software engineering agents. Proc. ACM Softw. Eng., 2 0 (FSE), June 2025. doi:10.1145/3715754. URL https://doi.org/10.1145/3715754

work page doi:10.1145/3715754 2025
[38]

SWE -fixer: Training open-source LLM s for effective and efficient G it H ub issue resolution

Chengxing Xie, Bowen Li, Chang Gao, He Du, Wai Lam, Difan Zou, and Kai Chen. SWE -fixer: Training open-source LLM s for effective and efficient G it H ub issue resolution. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar (eds.), Findings of the Association for Computational Linguistics: ACL 2025, pp.\ 1123--1139, Vienna, Aust...

work page doi:10.18653/v1/2025.findings-acl.62 2025
[39]

Qwen3 Technical Report

An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, ...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[40]

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

John Yang, Carlos E Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik R Narasimhan, and Ofir Press. SWE -agent: Agent-computer interfaces enable automated software engineering. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=mXpq6ut8J3

work page 2024
[41]

SWE -smith: Scaling data for software engineering agents

John Yang, Kilian Lieret, Carlos E Jimenez, Alexander Wettig, Kabir Khandpur, Yanzhe Zhang, Binyuan Hui, Ofir Press, Ludwig Schmidt, and Diyi Yang. SWE -smith: Scaling data for software engineering agents. In The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2025 b . URL https://openreview.net/forum...

work page 2025
[42]

Kimi-dev: Agentless training as skill prior for swe-agents, 2025 c

Zonghan Yang, Shengjie Wang, Kelin Fu, Wenyang He, Weimin Xiong, Yibo Liu, Yibo Miao, Bofei Gao, Yejie Wang, Yingwei Ma, Yanhao Li, Yue Liu, Zhenxing Hu, Kaitai Zhang, Shuyi Wang, Huarong Chen, Flood Sung, Yang Liu, Yang Gao, Zhilin Yang, and Tianyu Liu. Kimi-dev: Agentless training as skill prior for swe-agents, 2025 c . URL https://arxiv.org/abs/2509.23045

work page arXiv 2025
[43]

davinci-dev: Agent-native mid-training for software engineering, 2026

Ji Zeng, Dayuan Fu, Tiantian Mi, Yumin Zhuang, Yaxing Huang, Xuefeng Li, Lyumanshan Ye, Muhang Xie, Qishuo Hua, Zhen Huang, Mohan Jiang, Hanning Wang, Jifan Lin, Yang Xiao, Jie Sun, Yunze Wu, and Pengfei Liu. davinci-dev: Agent-native mid-training for software engineering, 2026. URL https://arxiv.org/abs/2601.18418

work page arXiv 2026
[44]

Skywork-swe: Unveiling data scaling laws for software engineering in llms, 2025

Liang Zeng, Yongcong Li, Yuzhen Xiao, Changshi Li, Chris Yuhao Liu, Rui Yan, Tianwen Wei, Jujie He, Xuchen Song, Yang Liu, and Yahui Zhou. Skywork-swe: Unveiling data scaling laws for software engineering in llms, 2025. URL https://arxiv.org/abs/2506.19290

work page arXiv 2025
[45]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

work page