Elamur Research⌕
☰
home / papers / 2026 / FuseSearch: Learning Adaptive Parallel Execution for Efficient Code Localization

FuseSearch: Learning Adaptive Parallel Execution for Efficient Code Localization

Paper · 2026 · Ke Xu, Siyang Xiao, Ming Liang et al.

Summary

Finding the files and functions that must change to fix a software issue, known as code localization, consumes much of the compute that coding agents spend. Under strict limits on interaction turns, sequential agents run out of turns before they gather enough evidence. Issuing many tool calls per turn helps, but a fixed number of parallel calls makes about a third of them redundant. FuseSearch addresses this by treating localization as a joint problem of quality and efficiency. The authors define "tool efficiency" as the average share of new, previously unseen code entities that each tool call returns. Compact Qwen3-based models (4B and 30B) are first trained with SFT on trajectories filtered for both F1 and efficiency, then with RL using a combined reward. The trained agent decides how many calls to run in parallel at each step, using only grep, glob and read_file. On SWE-bench Verified, FuseSearch-4B reaches state-of-the-art accuracy at a fraction of the cost. This shows that penalizing redundant search also improves accuracy, which makes cheap and fast localization agents practical.

Key points

  • Fixed-breadth parallel tool execution wastes 34.9% of tool calls. FuseSearch's information-gain metric counts how many returned entities were not already in the search history.
  • On SWE-bench Verified with a Qwen3-4B backbone, FuseSearch raises file-level F1 from 38.1% (RepoSearcher) to 84.7% and function-level F1 from 21.7% to 56.4%. It is also 93.6% faster and uses 67.7% fewer turns and 68.9% fewer tokens.
  • The reward ablation shows the multiplicative design (F1 + F1·e) works best: 84.65 file F1, 56.45 function F1, efficiency 0.69, about 5.4 s and 30.9k tokens. It outperforms SFT alone, F1 alone, and the additive F1 + e reward.
  • As a helper for Kimi-K2-Instruct on issue resolution, the pre-search mode keeps the pass rate roughly unchanged (68.1% vs 68.4%). It cuts turns from 41.1 to 31.6, time from 312 s to 223 s, and tokens from about 1053k to 562k.
  • The approach needs no code graphs, ASTs or language-specific parsers. However, evaluation covers only Python repositories and issue-driven tasks.

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

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksInformation Systems · 1,730 worksSoftware · 119 works

Cite

Cite as Ke Xu, Siyang Xiao, Ming Liang, Yichen Yu, Zhixiang Wang, Jingxuan Xu, Dajun Chen, Wei Jiang and Yong Li. 2026. FuseSearch: Learning Adaptive Parallel Execution for Efficient Code Localization. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.143

@inproceedings{xu2026fusesearch,
  title = {{FuseSearch: Learning Adaptive Parallel Execution for Efficient Code Localization}},
  author = {Ke Xu, Siyang Xiao, Ming Liang, Yichen Yu, Zhixiang Wang, Jingxuan Xu, Dajun Chen, Wei Jiang and Yong Li},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.143},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.143}
}

References

19 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Zhaoling Chen, Xiangru Tang, Gangda Deng, Fang Wu, Jialong Wu, Zhiwei Jiang, Viktor Prasanna, Arman Cohan, and Xingyao Wang. 2025. Locagent: Graph-guided llm agents for code localization. Preprint, arXiv:2503.09089.
  2. Pengfei Gao and Chao Peng. 2025. More with less: An empirical study of turn-control strategies for efficient coding agents. arXiv preprint arXiv:2510.16786. — [arXiv:2510.16786](https://arxiv.org/abs/2510.16786) find
  3. Zhonghao Jiang, Xiaoxue Ren, Meng Yan, Wei Jiang, Yong Li, and Zhongxin Liu. 2025. Issue localization via llm-driven iterative code graph searching. Preprint, arXiv:2503.22424. — [arXiv:2503.22424](https://arxiv.org/abs/2503.22424) find
  4. Carlos E Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik R Narasimhan. 2024. SWE-bench: Can language models resolve real-world github issues? In The Twelfth International Conference on Learning Representations.
  5. Dayoon Ko, Jihyuk Kim, Haeju Park, Sohyeon Kim, Dahyun Lee, Yongrae Jo, Gunhee Kim, Moontae Lee, and Kyungjae Lee. 2025. Hybrid deep searcher: Integrating parallel and sequential search reasoning. Preprint, arXiv:2508.19113. — [arXiv:2508.19113](https://arxiv.org/abs/2508.19113) find
  6. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles.
  7. Weichen Li, Xiaotong Huang, Jianwu Zheng, Zheng Wang, Chaokun Wang, Li Pan, and Jianhua Li. 2025. rllm: Relational table learning with llms. Preprint, arXiv:2407.20157. — [arXiv:2407.20157](https://arxiv.org/abs/2407.20157) find
  8. Zexiong Ma, Chao Peng, Qunhong Zeng, Pengfei Gao, Yanzhen Zou, and Bing Xie. 2025. Tool-integrated reinforcement learning for repo deep search. Preprint, arXiv:2508.03012. — [arXiv:2508.03012](https://arxiv.org/abs/2508.03012) find
  9. Ben Pan, Carlo Baronio, Albert Tam, Pietro Marsella, Mokshit Jain, Daniel Chiu, Swyx, and Silas Alberti. 2025. Introducing swe-grep and swe-grep-mini: Rl for multi-turn, fast context retrieval. find
  10. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. Preprint, arXiv:2402.03300.
  11. Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. 2024. Hybridflow: A flexible and efficient rlhf framework. arXiv preprint arXiv: 2409.19256.
  12. Tarun Suresh, Revanth Gangi Reddy, Yifei Xu, Zach Nussbaum, Andriy Mulyar, Brandon Duderstadt, and Heng Ji. 2025. Cornstack: High-quality contrastive data for better code retrieval and reranking. Preprint, arXiv:2412.01007. — [arXiv:2412.01007](https://arxiv.org/abs/2412.01007) find
  13. Zhiwen Tan, Jiaming Huang, Qintong Wu, Hongxuan Zhang, Chenyi Zhuang, and Jinjie Gu. 2025. Rag-r1: Incentivizing the search and reasoning capabilities of llms through multi-query parallelism. Preprint, arXiv:2507.02962. — [arXiv:2507.02962](https://arxiv.org/abs/2507.02962) find
  14. Qwen Team. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388.
  15. 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, and 5 others. 2025. Openhands: An open platform f…
  16. Chunqiu Steven Xia, Yinlin Deng, Soren Dunn, and Lingming Zhang. 2024. Agentless: Demystifying llm-based software engineering agents. Preprint, arXiv:2407.01489.
  17. 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. Preprint, arXiv:2405.15793.
  18. Eric Zelikman, Georges Harik, Yijia Shao, Varuna Jayasiri, Nick Haber, and Noah D. Goodman. 2024. Quiet-star: Language models can teach themselves to think before speaking. Preprint, arXiv:2403.09629.
  19. Shu Zhao, Tan Yu, Anbang Xu, Japinder Singh, Aaditya Shukla, and Rama Akkiraju. 2025. Parallelsearch: Train your llms to decompose query and search sub-queries in parallel with reinforcement learning. Preprint, arXiv:2508.09303. — [arXiv:2508.09303](https://arxiv.org/abs/2508.09303) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/xu-2026-fusesearch-learning-adaptive-parallel-execution/