Elamur Research⌕
☰
home / papers / 2026 / RepoDistill: Distilling Repository Knowledge through Compression-Aware Budget Allocation and Policy Optimization

RepoDistill: Distilling Repository Knowledge through Compression-Aware Budget Allocation and Policy Optimization

Paper · 2026 · Xin Yin, Zixiang Ding, Yiang Zhang et al.

Summary

Repository-level coding tasks require LLMs to work with long, noisy contexts that span many files. Retrieving more code often buries the relevant parts and lowers accuracy. In the authors' pilot study with function-chunking RAG, scores rose until about 30k tokens of context and then fell sharply at 100k–200k tokens, even for Gemini-3-Pro.

RepoDistill addresses this in three stages:

  • GraphRAG: a lightweight, swappable module builds a dependency graph of classes and functions. It retrieves anchor nodes, their call neighbours and the paths connecting them.
  • CABA: each retrieved function is cut down to its own token budget, keeping only the most useful code blocks.
  • CAPO: a learned policy sets these budget ratios (0–100%) by treating context management as a sequential decision problem.

The method was tested on SWE-QA, CoderEval and LongCodeU and outperformed baselines in every setting. A small trained model can also serve as a cheap pre-compressor for commercial LLMs. This makes the approach practical for cost-sensitive long-context code assistants.

Key points

  • Results:
    • SWE-QA: gains reach +7.00 points for open-source models and +7.61 for closed-source ones.
    • CoderEval: Pass@10 up to +24.1% (Python) and +24.4% (Java).
    • LongCodeU: up to +0.25 on LDU.
  • Compressor for closed models: Qwen3-4B-Instruct trained with RepoDistill cuts input tokens by up to 66% for closed LLMs with similar quality. With GPT-5.2 it reaches 43.03 on SWE-QA, only slightly below the untrained variant.
  • Graph construction: code is parsed with tree-sitter into a graph with CONTAIN and INVOKE edges. Nodes are located with Qwen3-Embedding-0.6B, and candidates are re-ranked with λ = 0.5.
  • Compression mechanics: this step builds on LongCodeZip. It splits functions into blocks at perplexity jumps (α = 0.2) and picks blocks by MMR rather than with fixed budgets. Per the overview figure, training combines SFT via knowledge distillation with multi-turn RL.
  • Limitations: RL training is GPU-intensive, and evaluation covers only Python and Java.

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

🎓Shanghai Jiao Tong UniversityShanghai · China212 works · 19,544 citations🎓Zhejiang UniversityHangzhou · China224 works · 13,587 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksInformation Systems · 1,730 works

Cite

Cite as Xin Yin, Zixiang Ding, Yiang Zhang, Qiang Wang, Rui Wang, Chao Ni and Zhe Cui. 2026. RepoDistill: Distilling Repository Knowledge through Compression-Aware Budget Allocation and Policy Optimization. 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.217

@inproceedings{yin2026repodistill,
  title = {{RepoDistill: Distilling Repository Knowledge through Compression-Aware Budget Allocation and Policy Optimization}},
  author = {Xin Yin, Zixiang Ding, Yiang Zhang, Qiang Wang, Rui Wang, Chao Ni and Zhe Cui},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.217},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.217}
}

References

55 references, 22 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. \#\# Limitations Despite the promising results, our work has several limitations: • Dependency on retrieval quality. RepoDistill relies on the initial quality of GraphRAG retrieval. If the retrieval module fails to capture relevant code segments due to complex implicit dependencies or ambiguous querie… find
  2. Anthropic. 2025. Introducing claude sonnet 4.5. find
  3. Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, and 1 others. 2022. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073.
  4. Guanzheng Chen, Xin Li, Zaiqiao Meng, Shangsong Liang, and Lidong Bing. Clex: Continuous length extrapolation for large language models. In The Twelfth International Conference on Learning Representations. find
  5. Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuandong Tian. 2023. Extending context window of large language models via positional interpolation. arXiv preprint arXiv:2306.15595. find
  6. Yuhan Chen, Ang Lv, Ting-En Lin, Changyu Chen, Yuchuan Wu, Fei Huang, Yongbin Li, and Rui Yan. 2024. Fortify the shortest stave in attention: Enhancing context awareness of large language models for effective tool use. In Proceedings of the 62nd Annual Meeting of the Association for Computational Li… find
  7. Google DeepMind. 2024. Gemini 2.0 flash thinking. find
  8. Google DeepMind. 2025. Gemini 3 pro: Best for complex tasks and bringing creative concepts to life. find
  9. Jiayu Ding, Shuming Ma, Li Dong, Xingxing Zhang, Shaohan Huang, Wenhui Wang, Nanning Zheng, and Furu Wei. 2023. Longnet: Scaling transformers to 1,000,000,000 tokens. arXiv preprint arXiv:2307.02486. llm context window without tuning. In Proceedings of the 41st International Conference on Machine Le…
  10. Jia Li, Xuyuan Guo, Lei Li, Kechi Zhang, Ge Li, Zhengwei Tao, Fang Liu, Chongyang Tao, Yuqi Zhu, and Zhi Jin. 2025. Benchmarking long-context language models on long code understanding. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)… find
  11. Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, and 1 others. 2025a. Deepseek-v3.2: Pushing the frontier of open large language models. arXiv preprint arXiv:2512.02556.
  12. Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12:157–173.
  13. Yang Liu, Li Zhang, Fang Liu, Zhuohang Wang, Donglin Wei, Zhishuo Yang, Kechi Zhang, Jia Li, and Lin Shi. 2025b. Reposcope: Leveraging call chain-aware multi-view context for repository-level code generation. arXiv preprint arXiv:2507.14791. — [arXiv:2507.14791](https://arxiv.org/abs/2507.14791) find
  14. Lang Feng, Zhenghai Xue, Tingcong Liu, and Bo An. 2025. Group-in-group policy optimization for llm agent training. arXiv preprint arXiv:2505.10978.
  15. Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. 2025c. Understanding r1-zero-like training: A critical perspective. arXiv preprint arXiv:2503.20783.
  16. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  17. Daye Nam, Andrew Macvean, Vincent Hellendoorn, Bogdan Vasilescu, and Brad Myers. 2024. Using an llm to help with code understanding. In Proceedings of the IEEE/ACM 46th International Conference on Software Engineering, pages 1–13. find
  18. Chi Han, Qifan Wang, Wenhan Xiong, Yu Chen, Heng Ji, and Sinong Wang. 2023. Lm-infinite: Simple on-the-fly length generalization for large language models. arXiv preprint arXiv:2308.16137. — [arXiv:2308.16137](https://arxiv.org/abs/2308.16137) find
  19. Jian Hu. 2025. Reinforce++: A simple and efficient approach for aligning large language models. arXiv preprint arXiv:2501.03262.
  20. Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, and 1 others. 2024. Qwen2.5-coder technical report. arXiv preprint arXiv:2409.12186.
  21. Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. Search-r1: Training llms to reason and leverage search engines with reinforcement learning. arXiv preprint arXiv:2503.09516.
  22. Hongye Jin, Xiaotian Han, Jingfeng Yang, Zhimeng Jiang, Zirui Liu, Chia-Yuan Chang, Huiyuan Chen, and Xia Hu. 2024. Llm maybe longlm: Selfextend Chao Ni, Xin Yin, Kaiwen Yang, Dehai Zhao, Zhenchang Xing, and Xin Xia. 2023. Distinguishing look-alike innocent and vulnerable code by subtle semantic rep… find
  23. OpenAI. 2024. Learning to reason with llms. find
  24. OpenAI. 2025. Introducing gpt-5.2. find
  25. Jie Ouyang, Ruiran Yan, Yucong Luo, Mingyue Cheng, Qi Liu, Zirui Liu, Shuo Yu, and Daoyu Wang. 2025a. Training powerful llm agents with end-to-end reinforcement learning. find
  26. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. 2022. Tra… — Training language models to follow instructions with human feedback
  27. Siru Ouyang, Wenhao Yu, Kaixin Ma, Zilin Xiao, Zhihan Zhang, Mengzhao Jia, Jiawei Han, Hongming Zhang, and Dong Yu. 2025b. Repograph: Enhancing ai software engineering with repository-level code graph. In 13th International Conference on Learning Representations, ICLR 2025, pages 30361–30384. Intern… find
  28. Bowen Peng, Jeffrey Quesnelle, Honglu Fan, and Enrico Shippole. Yarn: Efficient context window extension of large language models. In The Twelfth International Conference on Learning Representations. find
  29. Weihan Peng, Yuling Shi, Yuhang Wang, Xinyun Zhang, Beijun Shen, and Xiaodong Gu. 2025a. Swe-qa: Can language models answer repository-level code questions? arXiv preprint arXiv:2509.14635.
  30. Zhiyuan Peng, Xin Yin, Rui Qian, Peiqin Lin, Yongkang Liu, Hao Zhang, Chenhao Ying, and Yuan Luo. 2025b. Soleval: Benchmarking large language models for repository-level solidity smart contract generation. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pag… find
  31. Qwen. 2025a. Qwen3-coder: Agentic coding in the world. find
  32. Qwen. 2025b. Qwen3-max: Just scale it. find
  33. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347. find
  34. ByteDance Seed, Jiaze Chen, Tiantian Fan, Xin Liu, Lingjun Liu, Zhiqi Lin, Mingxuan Wang, Chengyi Wang, Xiangpeng Wei, Wenyuan Xu, and 1 others. 2025. Seed1. 5-thinking: Advancing superb reasoning models with reinforcement learning. arXiv preprint arXiv:2504.13914. — Seed1.5-Thinking: Advancing Superb Reasoning Models with Reinforcement Learning
  35. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, and 1 others. 2024a. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.
  36. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, YK Li, Y Wu, and Daya Guo. 2024b. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.
  37. Yuling Shi, Yichun Qian, Hongyu Zhang, Beijun Shen, and Xiaodong Gu. 2025. Longcodezip: Compress long context for code language models. arXiv preprint arXiv:2510.00446. — [arXiv:2510.00446](https://arxiv.org/abs/2510.00446) find
  38. Weisong Sun, Yun Miao, Yuekang Li, Hongyu Zhang, Chunrong Fang, Yi Liu, Gelei Deng, Yang Liu, and Zhenyu Chen. 2025. Source code summarization in the era of large language models. In 2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE), pages 1882–1894. IEEE. find
  39. Kimi Team, Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, and 1 others. 2025. Kimi k2: Open agentic intelligence. arXiv preprint arXiv:2507.20534.
  40. Xingliang Wang, Baoyi Wang, Chen Zhi, Junxiao Han, Xinkui Zhao, Jianwei Yin, and Shuiguang Deng. 2025a. Grace: Graph-guided repository-aware code completion through hierarchical code fusion. arXiv preprint arXiv:2509.05980. — [arXiv:2509.05980](https://arxiv.org/abs/2509.05980) find
  41. Yanlin Wang, Yanli Wang, Daya Guo, Jiachi Chen, Ruikai Zhang, Yuchi Ma, and Zibin Zheng. 2025b. Rlcoder: Reinforcement learning for repository-level code completion. In 2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE), pages 1140–1152. IEEE. find
  42. Zihan Wang, Kangrui Wang, Qineng Wang, Pingyue Zhang, Linjie Li, Zhengyuan Yang, Xing Jin, Kefan Yu, Minh Nhat Nguyen, Licheng Liu, Eli Gottlieb, Yiping Lu, Kyunghyun Cho, Jiajun Wu, Li Fei-Fei, Lijuan Wang, Yejin Choi, and Manling Li. 2025c. Ragen: Understanding self-evolution in llm agents via mul…
  43. Jeff Wu, Long Ouyang, Daniel M Ziegler, Nisan Stiennon, Ryan Lowe, Jan Leike, and Paul Christiano. 2021. Recursively summarizing books with human feedback. arXiv preprint arXiv:2109.10862.
  44. Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. In The Twelfth International Conference on Learning Representations. Wenhan Xiong, Jingyu Liu, Igor Molybog, Hejia Zhang, Prajjwal Bhargava, Rui Hou, Louis Martin, Rashi Run… find
  45. Ruiyang Xu, Jialun Cao, Yaojie Lu, Ming Wen, Hongyu Lin, Xianpei Han, Ben He, Shing-Chi Cheung, and Le Sun. 2025. Cruxeval-x: A benchmark for multilingual code reasoning, understanding and execution. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1… find
  46. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv:2505.09388. arXiv preprint Zhen Yang, Fang Liu, Zhongxing Yu, Jacky Wai Keung, Jia Li, Shuo Liu, Yifan Hong, Xiaoxue Ma, Zhi Jin, a…
  47. Xin Yin, Chao Ni, Xinrui Li, Liushan Chen, Guojun Ma, and Xiaohu Yang. Enhancing llm’s ability to generate more repository-aware unit tests through precise context injection. find
  48. Xin Yin, Chao Ni, Tien N Nguyen, Shaohua Wang, and Xiaohu Yang. 2024. Rectifier: Code translation with corrector via llms. arXiv preprint arXiv:2407.07472. — [arXiv:2407.07472](https://arxiv.org/abs/2407.07472) find
  49. Hao Yu, Bo Shen, Dezhi Ran, Jiaxin Zhang, Qi Zhang, Yuchi Ma, Guangtai Liang, Ying Li, Qianxiang Wang, and Tao Xie. 2024. Codereval: A benchmark of pragmatic code generation with generative pre-trained models. In Proceedings of the 46th IEEE/ACM International Conference on Software Engineering, page… find
  50. Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, and 1 others. 2025. Dapo: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476.
  51. Fengji Zhang, Bei Chen, Yue Zhang, Jacky Keung, Jin Liu, Daoguang Zan, Yi Mao, Jian-Guang Lou, and Weizhu Chen. Repocoder: Repository-level code completion through iterative retrieval and generation. In The 2023 Conference on Empirical Methods in Natural Language Processing.
  52. Kechi Zhang, Ge Li, Huangzhao Zhang, and Zhi Jin. 2024. Hirope: Length extrapolation for code models using hierarchical position. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13615–13627. find
  53. Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, and 1 others. 2025. Qwen3 embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176.
  54. Qihao Zhu, Qingyuan Liang, Zeyu Sun, Yingfei Xiong, Lu Zhang, and Shengyu Cheng. 2024. Grammart5: Grammar-integrated pretrained encoder-decoder neural model for code. In Proceedings of the IEEE/ACM 46th International Conference on Software Engineering, pages 1–13. find
  55. Siyuan Zhu, Chengdong Xu, Kaiqiang Ke, and Chao Yu. 2025. Context-picker: Dynamic context selection using multi-stage reinforcement learning. arXiv preprint arXiv:2512.14465. — [arXiv:2512.14465](https://arxiv.org/abs/2512.14465) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yin-2026-repodistill-distilling-repository-knowledge-through/