Research on models that write, repair, and reason about software. Compare programming benchmarks, execution feedback, and developer-facing capabilities.
43 papers · Latest edition October 4, 2026
Three of the latest briefs in this collection. Read the evidence and the original papers alongside them.
Better code documentation can faithfully reconstruct software, but surprisingly does not necessarily help AI coding agents fix real issues when the source code is already available.
GAGAR helps code agents learn not only to pass tests, but to produce cleaner and more targeted implementations by redistributing RL credit according to agentic quality judgments.
IR4RL improves image-to-code generation by rewarding models for making useful visual progress at every intermediate rendering step.
Newest editions first.
Md Shohel Arman, Igor Molybog
Better code documentation can faithfully reconstruct software, but surprisingly does not necessarily help AI coding agents fix real issues when the source code is already available.
Jinhao Dong, Liang Zhao, Zihao Yue, Wenhan Ma, Linghao Zhang, Lei Li, Shicheng Li, Yifan Song, Bowen Ye, Fuli Luo
GAGAR helps code agents learn not only to pass tests, but to produce cleaner and more targeted implementations by redistributing RL credit according to agentic quality judgments.
Omri Kaduri, Kate Feingold, Phillip Isola, Tali Dekel
IR4RL improves image-to-code generation by rewarding models for making useful visual progress at every intermediate rendering step.
Bowen Ye, Lei Li, Shicheng Li, Zihao Yue, Linghao Zhang, Hanglong Lv, Yuanxin Liu, Wenhan Ma, Hao Tian, Rang Li, Jinhao Dong, Yikai Zhao, Xiangwei Deng, Hailin Zhang, Liang Zhao, Qi Liu, Lingpeng Kong, Tong Yang, Fuli Luo
CodeMidas turns existing codebases into scalable, automatically verified RL environments that train coding agents to perform better across diverse software tasks.
Gabrielle O'Brien, Reed Milewicz, Nasir Eisty
A survey of 527 researchers finds that AI-generated scientific code is usually checked informally by individuals rather than through systematic tests or peer review.
Alexander Krentsel, Shubham Agarwal, Mert Cemri, Shu Liu, Sidharth Sankhe, Ziming Mao, Matei Zaharia, Ion Stoica
AI coding agents can pass tests yet fail in the real world, so dependable development requires continuously narrowing the gaps between human intent, evaluation models, and deployment reality.
Leitian Tao, Baolin Peng, Haorui Wang, Hang Wang, Hao Cheng, Wenlin Yao, Qianhui Wu, Tao Ge, Sharon Li, Jianfeng Gao
ExecCritic trains one coding agent to create trustworthy tests and another to repair code from them, substantially improving repository-level bug fixing.
Chang Liu, Edward Raff, Kristopher Micinski
This work shows that LLM decompilers can produce code that compiles and passes known tests while silently changing program behavior or erasing vulnerabilities.
Tongyao Zhu, Wei Hern Lim, Min-Yen Kan
This study shows that coding models often fix bugs by changing too much, and that targeted training can make their edits smaller, safer, and easier to review.
Shangqing Tu, Daniel Zhang-Li, Yucheng Wang, Shiyu Gan, Yanpeng Wang, Huiqiang Rong, Mofei Chen, Shen Yang, Yini Chen, Yinuo Duan, Haoxuan Li, Binglin Liu, Ye He, Danqi Zheng, Zhanxin Hao, Yuxuan Wu, Mengting Tao, Yuqiu Liu, Jifan Yu, Juanzi Li, Bin Xu, Lei Hou, Huiqin Liu, Yu Zhang
CogEvol is a compact, production-trained model that turns course briefs into reliable slides and interactive educational web pages much faster and more cheaply than larger coding models.
Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar, Boris Ginsburg
A specialized language model system combines curated coding data, reinforcement learning, and iterative solution refinement to surpass human gold-medal performance in programming competitions.
Dung Le Quang, Dong Cao Van, Nam Le Hai, Linh Ngo Van, Anh M. T. Bui, Phuong T. Nguyen
XREPOTEST tests whether today’s LLMs can generate genuinely useful unit tests inside real multilingual software repositories.
Yiming Du, Yuxin Jiang, Tao Yuan, Jianbo Dai, Shaowei Wang, Jierun Chen, Chaofan Tao, Xianzhi Yu, Lifeng Shang, Kam-Fai Wong, Xiaohui Li, Haoli Bai
LEGO-RL turns real coding-agent harnesses into trainable reinforcement-learning environments and improves SWE-bench performance across three popular platforms.
Silin Chen, Haoyi Teng, Xiaodong Gu, Yuling Shi, Jiale Huang, Yongpan Wang, Hongyu Zhang, Haibing Guan
Repo0 helps coding agents build entire software repositories by first evolving and stabilizing an explicit modular architecture before writing the code.
Tianyi Xiong, Zhengyuan Yang, Xiaofei Wang, Chung-Ching Lin, Ruichun Ma, Kevin Lin, Zhendong Wang, Linjie Li, Chenxi Liu, Ruibo Chen, Ramani Duraiswami, Heng Huang, Lijuan Wang
RubSE helps vision-language models iteratively repair generated interfaces by turning visual feedback into prioritized, structured rubrics instead of letting each code edit disrupt the whole page.
Rishi Shah, Rishav Shrestha
This paper shows that many AI-generated GPU kernels that pass standard tests are silently wrong, and proposes stricter contracts to catch those failures.
Nimra Noor, Muhammad Bilal, Abdullah Hussain, Hassan Baig
Nova3D turns 3D generation from making static meshes into writing editable, measurable, and animatable Blender programs.
Yunhao Liang, Chengguang Gan, Ruixuan Ying
Security tests can help LLMs repair vulnerable code, but their success depends heavily on test coverage, feedback design, and the model-task combination.
Arshia Soltani Moakhar, Iman Gholami, Max Springer, Mahdi JafariRaviz, MohammadTaghi Hajiaghayi
An agentic system turns advanced mathematical research into verified Lean proofs, even when the required concepts are missing from existing libraries.
Michael Tran, Fred Lewis, Kun Yang, Saksham Thakur, Aditya Kini, Aditya Patil, Milad Hashemi, Parthasarathy Ranganathan
A massive real-world study finds that AI-written C++ can increase maintenance and compute costs, but targeted feedback helps make the code more efficient.
Dongyi Lv, Fushun E, Aichen Cai, Liang Huang, Ya Zhang, Qiuyu Ding, Canhui Wu, Zhi Wang, Yuesong Zhang, Jiaqi Wang, Nan Duan
FailForge teaches code agents to learn from their failures, turning past mistakes into reusable skills that substantially improve software-engineering task performance.
Haomin Qi, Xingliang Wang, Xuanqi Gao, Baihui Sang, Xin Zhang, Minghua Ma, Pengfei Gao, Yu Kang, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang, Qi Zhang
Change2Task turns real pull-request history into reliable, reusable coding-agent tasks for training and evaluation.
Yihao Chen, Shi Chang, Khaled Chawa, Feng Lin, Boyuan Chen, Shaowei Wang, Ahmed E. Hassan
MindForge trains smaller language models to build complete software from scratch by learning through source-free, end-to-end programming environments.
Amir M. Ebrahimi, Mohammed Mehedi Hasan, Aaditya Bhatia, Gopi Krishnan Rajbahadur, Ahmed E. Hassan
LLMs can find the code that needs changing but often avoid deleting it, so this paper measures the problem and explores training models to remove code more reliably.
Fabian Kreppel, Reza Salkhordeh, Ferdinand Schmidt-Kaler, André Brinkmann
Frontier LLMs can autonomously generate competitive trapped-ion shuttling compilers, potentially cutting compiler development from months to days.
Bo Huang, Fengxiang Li, Hao Xu, Haoyang Huang, Hongyi Fu, Jinhua Hao, Kun Yuan, Minglei Zhang, Pengcheng Xu, Shiyang Liu, Wenhao Zhuang, Yuze Shi, Zongxian Feng, Chao Wang, Cheng He, Chongling Rao, Deyu Cao, Fan Yang, Gang Xiong, Haochen Liu, Jiabao Li, Jian Liang, Jinghui Jia, Jingwen Chang, Jun Du, Junyu Shi, Min Li, Mingqi Wu, Qiang Gao, Shangpeng Yan, Shaotong Qi, Shu Xu, Shuo Zhou, Tiankuo Xu, Tong Zheng, Weilun Zhao, Xiancheng Meng, Xianda Sun, Xiaoyu Jiang, Xunhao Jia, Yao Xia, Yimeng Xu, Yinghan Cui, Yingpeng Chen, Yiwen Ning, Yong Wang, Yuxuan Sun, Zhongsheng Liu, Ming Sun, Cheng Luo, Chen Yang, Han Li, Kun Gai
KAT-Coder-V2.5 trains coding agents to autonomously modify and test real software repositories using scalable environments, tool-use trajectories, and reinforcement learning.
Haotian Lin, Silin Chen, Xiaodong Gu, Yuling Shi, Chengxi Pan, Jiaqi Ge, Mengfan Li, Jianghong Huang, Mengchieh Chuang, Beijun Shen, Haibing Guan
ACQUIRE helps coding agents fix software issues more accurately by having them explicitly ask and answer repository-specific questions before writing a patch.
Hao He, Shyam Agarwal, Yegor Denisov-Blanch, Pavel Azaletskiy, Sanmi Koyejo, Bogdan Vasilescu
A large enterprise study finds that an AI coding mandate eventually doubled developer output, while shifting much of code review from humans to automation.
Suzhen Zhong, Shayan Noei, Bram Adams, Ying Zou
A large study finds that AI agents can speed up code reviews, but faster decisions do not necessarily mean higher-quality reviews.
Yubo Wang, Jiarong Liang, Yuxuan Zhang, Xuye Liu, Cong Wei, Yuyu Zhang, Ping Nie, Wenhu Chen
By training code models to fill in dependency-aware function gaps, the paper improves coding-agent performance and preserves broader tool-use abilities after agentic post-training.
Niels Mündler-Sasahara, Hristo Venev, Dawn Song, Martin Vechev, Jingxuan He
This work turns the compiler into a real-time coding partner that catches AI-generated Rust mistakes while the code is still being written.
Zhihao Lin, Mingyi Zhou, Yizhuo Yang, Li Li
This paper shows that giving code agents simple structural hints from static analysis can make them more predictable, more stable across runs, and slightly better at finding the right code.
Ruoyu Wang, Jierun Chen, Shaowei Wang, Chaofan Tao, Sidi Yang, Yuxin Jiang, Kim-Hui Yap, Lifeng Shang, Xiaohui Li, Haoli Bai
This paper teaches coding agents not just to write pull requests, but to review, critique, and revise them in a loop that more closely matches real software development.
Alex Heilman, Alex Kyllo, Emerson Murphy-Hill
The paper finds that engineers write significantly more pull requests in weeks when they use GitHub Copilot heavily, suggesting the coding assistant can boost developer efficiency even after controlling for effort and confounders.
Dongjian Ma, Silin Chen, Yufei Yang, Yulin Shi, Yanfu yan, Xiaodong Gu
This paper shows that coding agents work better when they can 'see' a repository’s structure, and that combining visual graphs with text can cut token use while preserving or improving bug-fixing performance.
Zhihao Lin, Junhua Zhu, Mingyi Zhou, Xin Wang, Zhensu Sun, Renyu Yang, David Lo, Li Li
This paper asks whether running tests during LLM code repair is really worth it, and finds that agents often pay a big execution cost for only modest gains.
Liliana Hotsko, Yinxi Li, Yuntian Deng, Pengyu Nie
Code2LoRA turns code repositories into lightweight, updateable adapters so code models can learn project-specific context without stuffing the prompt or retraining for every repo.
Jiasheng Zheng, Boxi Cao, Boxi Yu, Yuzhong Zhang, Jialun Cao, Yaojie Lu, Hongyu Lin, Xianpei Han, Le Sun
This paper introduces a new way to create harder, more original coding tasks for training LLMs, helping reinforcement learning with verifiable rewards scale beyond hand-crafted or heuristic data.
Tong Ye, Hang Yu, Tengfei Ma, Xuhong Zhang, Jianguo Li, Peng Di, Peiyu Liu, Jianwei Yin, Wenhai Wang
DOMINO learns the hidden structure of a target coding domain from a few examples and uses it to synthesize better training data for large language models.
Thanawat Lodkaew, Johannes Ackermann, Soichiro Nishimori, Nontawat Charoenphakdee, Masashi Sugiyama, Takashi Ishida
This paper proposes a clever way to tell when coding agents are 'gaming the test' by using randomized capped evaluations, and shows how to discourage that cheating during training.
Jianwen Sun, Chuanhao Li, Zizhen Li, Yukang Feng, Fanrui Zhang, Yifei Huang, Yu Dai, Kaipeng Zhang
This paper creates the first large-scale benchmark for whole-game code generation in a real game engine, revealing that today’s AI can often write code that compiles but still fails to produce working games.
Chuan Xiao, Zhengbo Jiao, Shaobo Wang, Wei Wang, Bing Zhao, Hu Wei, Linfeng Zhang, Lin Qu
This paper shows how coding agents can learn from their own past mistakes by turning old solving traces into new training tasks, steadily improving software-engineering performance over multiple iterations.
Zhangyi Hu, Chenhui Liu, Tian Huang, Jindong Li, Yang Yang, Jiemin Wu, Zining Zhong, Menglin Yang, Yutao Yue
CoSPlay lets an LLM improve its own code and tests at inference time, using self-generated unit tests and execution signals to boost code generation without ground-truth labels.