Elamur Research⌕
☰
home / papers / 2026 / EGSS: Entropy-guided Stepwise Scaling for Reliable Software Engineering

EGSS: Entropy-guided Stepwise Scaling for Reliable Software Engineering

Paper · 2026 · Chenhui Mao, Yuanting Lei, Zhixiang Wei et al.

Summary

Agentic test-time scaling (TTS) helps LLM agents resolve repository-level software issues, but running many trajectories is costly, and choosing the best candidate patch is unreliable. The authors, from Ant Group, propose EGSS, which addresses both problems. Its Dynamic Stepwise Search (DSS) computes tool entropy, a measure of how uncertain the agent is when choosing a tool at each step. Most steps are near-deterministic routine actions, so DSS spends extra compute, including calls to an auxiliary judge, only at the rare high-entropy decision points. Its Test Consolidation Augmentation (TCA) merges the tests that different trajectories wrote into a single executable suite and ranks candidate patches by how many of those tests they pass. This counters "self-deceptive debugging", where an agent checks its own fix yet still accepts a wrong one. On SWE-Bench-Verified, EGSS improves every evaluated model by 5–10 points and sets a new open-source best with GLM-4.6. It also needs only half the ensemble size of the baselines. The work shows that targeted compute allocation can beat brute-force sampling on both accuracy and cost.

Key points

  • Preliminary analysis (Kimi-K2-Instruct, 500 SWE-Bench-Verified tasks):
    • 440 trajectories included self-verification, yet 157 of them (35.7%) still ended with a wrong patch.
    • Tool entropy follows a right-skewed distribution, with most mass at low values.
  • Accuracy gains:
    • Kimi-K2-Instruct rises from 63.2% to 72.2% resolved.
    • GLM-4.6 rises from 65.8% to 74.6%, the best result among open-source LLMs.
  • Efficiency at K = 4 versus baseline (Repeat + Dei Aug) at K = 8:
    • GLM-4.6 reaches 73.8% with about 291k tokens, versus 68.2% with about 502k tokens (42.0% fewer tokens).
    • Kimi-K2-Instruct reaches 68.4% with about 296k tokens, versus 66.4% with about 481k tokens (38.4% fewer tokens).
    • At a fixed K, TCA alone adds 3–13% token overhead.
  • Ablation: the consolidated test suite is the main source of gains. Top-1 and Top-3 selection by pass rate comes close to the Oracle upper bound.
  • Availability and limitations: code is released in the codefuse-ai/CodeFuse-Agent GitHub repository. The method has not been tested on open-ended tasks such as feature implementation or API design.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🏢Ant GroupHangzhou · China42 works · 1,131 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksInformation Systems · 1,437 works

🧭 Topics 1Research subjects and languages the work is about.Code & programming languages · 129 works

Cite

Cite as Chenhui Mao, Yuanting Lei, Zhixiang Wei, Ming Liang, Zhixiang Wang, Jingxuan Xu, Dajun Chen, Wei Jiang and Yong Li. 2026. EGSS: Entropy-guided Stepwise Scaling for Reliable Software Engineering. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1359

@inproceedings{mao2026egss,
  title = {{EGSS: Entropy-guided Stepwise Scaling for Reliable Software Engineering}},
  author = {Chenhui Mao, Yuanting Lei, Zhixiang Wei, Ming Liang, Zhixiang Wang, Jingxuan Xu, Dajun Chen, Wei Jiang and Yong Li},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1359},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1359}
}

References

30 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ibrahim Abdelaziz, Kinjal Basu, Mayank Agarwal, Sadhana Kumaravel, Matthew Stallone, Rameswar Panda, Yara Rizk, GP Shrivatsa Bhargav, Maxwell Crouse, Chulaka Gunasekara, and 1 others. 2024. Granite-function calling model: Introducing function calling abilities via multi-task learning of granular tas… find
  2. Antonis Antoniades, Albert Örwall, Kexun Zhang, Yuxi Xie, Anirudh Goyal, and William Wang. 2024. Swe-search: Enhancing software agents with monte carlo tree search and iterative refinement. arXiv preprint arXiv:2410.20285. find
  3. Islem Bouzenia, Premkumar Devanbu, and Michael Pradel. 2024. Repairagent: An autonomous, llm-based agent for program repair. arXiv preprint arXiv:2403.17134. — [arXiv:2403.17134](https://arxiv.org/abs/2403.17134) find
  4. Mark Chen. 2021. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374.
  5. Tongfei Chen and Colin Flaherty. 2025. \#1 open-source agent on SWE-Bench Verified by combining Claude 3.7 and O1. Augment Blog. Accessed: 2025-12-30. find
  6. Guanting Dong, Hangyu Mao, Kai Ma, Licheng Bao, Yifei Chen, Zhongyuan Wang, Zhongxia Chen, Jiazhen Du, Huiyang Wang, Fuzheng Zhang, and 1 others. 2025. Agentic reinforced policy optimization. arXiv preprint arXiv:2507.19849. — [arXiv:2507.19849](https://arxiv.org/abs/2507.19849) find
  7. Shubham Gandhi, Jason Tsay, Jatin Ganhotra, Kiran Kate, and Yara Rizk. 2025. When agents go astray: Course-correcting swe agents with prms. arXiv preprint arXiv:2509.02360. — [arXiv:2509.02360](https://arxiv.org/abs/2509.02360) find
  8. Pengfei Gao, Zhao Tian, Xiangxin Meng, Xinchen Wang, Ruida Hu, Yuanan Xiao, Yizhou Liu, Zhao Zhang, Junjie Chen, Cuiyun Gao, and 1 others. 2025. Trae agent: An llm-based agent for software engineering with test-time scaling. arXiv preprint arXiv:2507.23370. — [arXiv:2507.23370](https://arxiv.org/abs/2507.23370) find
  9. Carlos E Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. 2023. Swe-bench: Can language models resolve real-world github issues? arXiv preprint arXiv:2310.06770. 152 others. 2025a. Glm-4.5: Agentic, reasoning, and coding (arc) foundation models. Prepri…
  10. Kimi Team, Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, and 1 others. 2025b. Kimi k2: Open agentic intelligence. arXiv preprint arXiv:2507.20534.
  11. Muhammad Khalifa, Rishabh Agarwal, Lajanugen Logeswaran, Jaekyeom Kim, Hao Peng, Moontae Lee, Honglak Lee, and Lu Wang. 2025. Process reward models that think. arXiv preprint arXiv:2504.16828. — [arXiv:2504.16828](https://arxiv.org/abs/2504.16828) find
  12. Hanlin Wang, Jian Wang, Chak Tou Leong, and Wenjie Li. 2025. Steca: Step-level trajectory calibration for llm agent learning. arXiv preprint arXiv:2502.14276. find
  13. Jiaolong Kong, Xiaofei Xie, Mingfei Cheng, Shangqing Liu, Xiaoning Du, and Qi Guo. 2025. Contrastrepair: Enhancing conversation-based automated program repair via contrastive test case pairs. ACM Transactions on Software Engineering and Methodology, 34(8):1–31. find
  14. Taiyi Wang, Zhihao Wu, Jianheng Liu, Jianye Hao, Jun Wang, and Kun Shao. 2024a. Distrl: An asynchronous distributed reinforcement learning framework for on-device control agents. arXiv preprint arXiv:2410.14803. — [arXiv:2410.14803](https://arxiv.org/abs/2410.14803) find
  15. Dacheng Li, Shiyi Cao, Chengkun Cao, Xiuyu Li, Shangyin Tan, Kurt Keutzer, Jiarong Xing, Joseph E Gonzalez, and Ion Stoica. 2025a. S*: Test time scaling for code generation. arXiv preprint arXiv:2502.14382. find
  16. Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhattacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, and 1 others. 2025b. From generation to judgment: Opportunities and challenges of llm-asa-judge. In Proceedings of the 2025 Conference on Empirical Methods… find
  17. Qingyao Li, Xinyi Dai, Xiangyang Li, Weinan Zhang, Yasheng Wang, Ruiming Tang, and Yong Yu. 2025c. Codeprm: Execution feedback-enhanced process reward model for code generation. In Findings of the Association for Computational Linguistics: ACL 2025, pages 8169–8182. find
  18. Yingwei Ma, Yongbin Li, Yihong Dong, Xue Jiang, Rongyu Cao, Jue Chen, Fei Huang, and Binhua Li. 2025. Thinking longer, not larger: Enhancing software engineering agents via scaling test-time compute. arXiv preprint arXiv:2503.23803. — [arXiv:2503.23803](https://arxiv.org/abs/2503.23803) find
  19. Jiayi Pan, Xingyao Wang, Graham Neubig, Navdeep Jaitly, Heng Ji, Alane Suhr, and Yizhe Zhang. 2024. Training software engineering agents and verifiers with swe-gym. arXiv preprint arXiv:2412.21139. find
  20. Lin Sun, Chuang Liu, Xiaofeng Ma, Tao Yang, Weijia Lu, and Ning Wu. 2025. Freeprm: Training process reward models without ground truth process labels. arXiv preprint arXiv:2506.03570. 5 Team, Aohan Zeng, Xin Lv, Qinkai Zheng, Zhenyu Hou, Bin Chen, Chengxing Xie, Cunxiang Wang, Da Yin, Hao Zeng, Jiaj…
  21. Scott Wu. 2024. Introducing devin, the first ai software engineer. Cognition. ai Blog. find
  22. Chunqiu Steven Xia, Yinlin Deng, Soren Dunn, and Lingming Zhang. 2024. Agentless: Demystifying llm-based software engineering agents. arXiv preprint arXiv:2407.01489. find
  23. Yuntong Xie, Juyong Jiang, Hung Nguyen, Pinjia Tan, Zehua Gao, Jiazhen He, and David Lo. 2024. AutoCodeRover: Autonomous program improvement. In Proceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis (ISSTA). find
  24. 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, and 41 others. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388.
  25. John Yang, Carlos E Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. 2024. SWE-agent: Agent-computer interfaces enable automated software engineering. arXiv preprint arXiv:2405.15793.
  26. Karina Zainullina, Alexander Golubev, Maria Trofimova, Sergei Polezhaev, Ibragim Badertdinov, Daria Litvintseva, Simon Karasik, Filipp Fisin, Sergei Skvortsov, Maksim Nekrashevich, and 1 others. 2025. Guided search strategies in non-serializable environments with applications to software engineering… — [arXiv:2505.13652](https://arxiv.org/abs/2505.13652) find
  27. Guangtao Zeng, Maohao Shen, Delin Chen, Zhenting Qi, Subhro Das, Dan Gutfreund, David Cox, Gregory Wornell, Wei Lu, Zhang-Wei Hong, and 1 others. 2025. Satori-swe: Evolutionary test-time scaling for sample-efficient software engineering. arXiv preprint arXiv:2505.23604. — [arXiv:2505.23604](https://arxiv.org/abs/2505.23604) find
  28. Dan Zhang, Sining Zhoubian, Ziniu Hu, Yisong Yue, Yuxiao Dong, and Jie Tang. 2024a. Rest-mcts*: Llm self-training via process reward guided tree search. Advances in Neural Information Processing Systems, 37:64735–64772.
  29. Kexun Zhang, Weiran Yao, Zuxin Liu, Yihao Feng, Zhiwei Liu, Rithesh Murthy, Tian Lan, Lei Li, Renze Lou, Jiacheng Xu, and 1 others. 2024b. Diversity empowers intelligence: Integrating expertise of software engineering agents. arXiv preprint arXiv:2408.07060. — [arXiv:2408.07060](https://arxiv.org/abs/2408.07060) find
  30. Mingchen Zhuge, Changsheng Zhao, Dylan Ashley, Wenyi Wang, Dmitrii Khizbullin, Yunyang Xiong, Zechun Liu, Ernie Chang, Raghuraman Krishnamoorthi, Yuandong Tian, Yangyang Shi, Vikas Chandra, and Jürgen Schmidhuber. 2024. Agent-asa-judge: Evaluate agents with agents. Preprint, arXiv:2410.10934. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mao-2026-egss-entropy-guided-stepwise-scaling/