Pith. sign in

REVIEW 2 major objections 7 minor 62 references

CoRet: Improved Retriever for Code Editing

T0 review · 2 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read CoRet, a retrieval model that joins file structure and call-graph context to a likelihood-based training loss, raises perfect recall@5 on SWE-bench Verified from 0.34 to 0.54 and improves recall by at least 15 percentage points over…

desk verdict Useful incremental retriever with good ablations, but the 'at least 15pp' claim is overstated on LCA and should be fixed before publication. read the letter →

arxiv 2505.24715 v1 pith:TO7VYHHU submitted 2025-05-30 cs.LG cs.AIcs.CL

classification cs.LGcs.AIcs.CL
keywords denseretrievalcodeeditingrepository-levelcallgraphcontextlikelihoodtraininglossSWE-benchLongArenabuglocalisation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces CoRet, a dense retrieval model for code-editing tasks that returns the code chunks of a repository most relevant to a natural-language bug report or feature request. The central claim is that CoRet's design—training with a repository-level likelihood loss instead of a contrastive in-batch loss, prefixing each chunk with its file path, and enriching each chunk with its downstream call-graph neighbours—raises retrieval recall by at least 15 percentage points over existing retrievers on SWE-bench Verified and Long Code Arena bug localisation. If correct, this matters because retrieval is the first step for coding agents and human editors: a retriever that finds all the edited chunks at small $k$ would make downstream repair far more reliable. The paper further shows that each design choice contributes: call-graph context helps multi-chunk retrieval, in-instance negatives outperform across-instance ones, and file paths are a learned signal rather than a neutral prefix.

What carries the argument

The central object is the trained dense retriever $f(q,c)=\text{cosine\ similarity}$ between an encoded query and an encoded chunk, operating on chunks defined as individual functions, methods, and class declarations, each prefixed with its repository file path. The training loss is the mean log-likelihood of retrieving the ground-truth chunks under a softmax over the whole repository, approximated by sampling up to 1024 in-instance negatives; this replaces the standard in-batch contrastive objective. The call-graph context is implemented as string concatenation of a chunk $c_i$ with its downstream neighbours, separated by a special [DOWN] token and distinguished by trainable segment-type embeddings, so the encoder sees the chunk together with the code it invokes.

What would settle it

One concrete check is to run CoRet on the Long Code Arena instances while allowing the ground-truth set to include every chunk that the gold patch reads or calls, not only the chunks it edits; if perfect recall drops materially, the edited-only labelling is hiding a substantial part of the retrieval problem. A second check, which the paper itself names, is to test on a multi-language dataset such as SWE-PolyBench to see whether the 15-point margin survives outside Python.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that repository-level code retrieval for editing should be optimised as a per-repository categorical likelihood—the probability of picking the ground-truth edited chunks out of all chunks in that repository—rather than with contrastive losses whose negatives come from across the batch. The paper reports that CoRet achieves perfect recall@5 of 0.54 on SWE-bench Verified versus 0.34 for the best baseline (CodeSage Small), a comparable jump on Long Code Arena, and that the gain is composed of three ingredients: the likelihood loss with up to 1024 sampled in-instance negatives, call-graph context appended to each chunk with a dedicated [DOWN] token, and file paths prepended to every chunk. The paper also shows, via ablations, that removing call-graph context, switching to across-instance negatives, or dropping file paths all degrade performance, with file paths mattering only after fine-tuning.

Load-bearing premise

The evaluation assumes that the only code that must be retrieved is the set of chunks edited in the pull request's patch; if a correct fix also requires unedited context—callers, tests, constants, or configuration—the recall numbers do not measure the retriever's true usefulness.

Editorial extensions

If this is right

  • On SWE-bench Verified, perfect recall@5 rises from 0.34 (CodeSage Small) to 0.54 with CoRet, a 52.9% relative improvement, and recall@20 rises from 0.51 to 0.71.
  • On Long Code Arena, perfect recall@5 goes from 0.26 to 0.32, recall@20 from 0.34 to 0.47, and MRR from 0.28 to 0.47.
  • The paper's ablations show each design piece is load-bearing: replacing the likelihood loss with across-instance negatives, removing call-graph context, or testing without file paths each lowers performance, and recall improves monotonically as the number of in-instance negatives grows from 8 to 1024.
  • A retriever that needs to surface far fewer chunks to an agent or user is the practical payoff: CoRet reaches high chunk recall at small $k$ where the baseline encoders require much larger $k$ to approach the same level.
  • Because retrieval is the first step in agentic code editing, a stable 15-point recall gain at the retrieval stage is a direct upper-bound improvement for any downstream repair pipeline that consumes the retrieved chunks.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that the edited-chunks-only ground truth makes the reported recall a lower bound on the retriever's value: a code-editing agent also needs unedited callers, tests, constants, and configuration, so extending the label set to everything the gold patch reads would likely change both absolute numbers and the ranking of methods.
  • The per-repository likelihood loss is a general recipe for any retrieval task over a structured corpus with an internal graph—API documentation, scientific literature with citation networks, or configuration files—so the method transfers beyond code editing.
  • The attention-map evidence that CoRet learns to exploit file-path tokens suggests a testable extension: deliberately masking or leaking paths in the query at training time would quantify how much of the 15-point gain is path-matching rather than semantic understanding.
  • The paper restricts experiments to Python and names SWE-PolyBench as future work; whether the margin survives in other languages is the direct empirical test of the method's generality, since the chunking and call-graph extraction are language-dependent.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 7 minor

Summary. The paper introduces CoRet, a dense retriever for code editing that fine-tunes the CodeSage S encoder with a multi-class log-likelihood loss using negatives sampled from the same repository. The model also incorporates file-path prefixes and call-graph context into the chunk representation. CoRet is evaluated on SWE-bench Verified and Long Code Arena (LCA) bug localisation, where it is reported to improve chunk-level perfect recall and MRR over off-the-shelf baselines. The paper includes ablations of the loss, negative sampling, call graph context, and file path, plus detailed appendices on chunking, metrics, and hyperparameters.

Significance. CoRet addresses a practical problem—repository-level retrieval for code-editing agents—and the paper provides a systematic empirical study with a clear set of ablations. If the results are confirmed, CoRet offers a strong retrieval baseline and a useful training recipe. The paper is transparent about its scope (Python only, two datasets) and provides reproducibility details. The main weakness is that the headline quantitative claim is overstated for LCA, and the absence of uncertainty estimates and a fine-tuned contrastive baseline makes the evidence less definitive. These issues are correctable and should be addressed before the paper is accepted.

major comments (2)
  1. [Abstract; Section 1; Table 1] The abstract and Section 1 claim improvements in retrieval recall of at least 15 percentage points on both SWE-bench and Long Code Arena. Table 1 does not support this for LCA: relative to CodeSage S, CoRet improves perfect recall@5 from 0.26 to 0.32 (6 points) and perfect recall@20 from 0.34 to 0.47 (13 points); only the MRR gain (0.19) exceeds 15 points, and the claim explicitly concerns recall. In addition, LCA has only 34 test instances and the paper reports no error bars or significance tests, so the 6-point difference at recall@5 corresponds to roughly two instances and could be within noise. Please either restrict the 'at least 15 percentage points' claim to SWE-bench Verified, or add uncertainty quantification and significance testing on LCA, or rephrase the claim to name the metric for which it holds on both datasets.
  2. [Section 4.4; Figure 5] The main results in Table 1 compare CoRet only against off-the-shelf baselines; no fine-tuned baseline with a standard contrastive loss is evaluated on the test benchmarks. The ablations in Figure 5 compare in-instance and across-instance negatives on a hold-out set from SWE-bench train, but this comparison is not repeated on SWE-bench Verified or LCA. Because fine-tuning on retrieval data is expected to produce large gains over off-the-shelf encoders, the reported improvements may not be attributable to the proposed loss alone. Please add test-set results for CodeSage S fine-tuned with a standard contrastive/across-instance loss, or explicitly discuss why the hold-out comparison is sufficient.
minor comments (7)
  1. [Section 4.4] The sentence 'On SWE-bench Verified, recall@5 improves by 52.9% compared to CodeSage S for recall@5 and by about 35% for recall@20' appears to describe CoRet-CG rather than CoRet; the numbers in Table 1 are 0.52 vs 0.34 (52.9% relative) and 0.69 vs 0.51 (35.3% relative). Please clarify which model the sentence refers to.
  2. [Section 4.4; Table 1] The text says the file-context ablation 'CoRet−CG + file' shows 'a substantial decrease in performances throughout,' but Table 1 shows that adding file context improves SWE-bench Verified recall@5 from 0.52 to 0.54 and leaves recall@20 at 0.69; the statement is not accurate for that dataset.
  3. [Conclusion] The word 'constrastive' should be 'contrastive' in the sentence 'this problem differs from the traditional constrastive representation learning problem' (the same typo appears in Appendix A).
  4. [Figure 4 caption] In the caption, 'chuck level' should be 'chunk level' (both occurrences).
  5. [Appendix A] The sentence 'LLM2Vec comes at a cost of a much higher parameter count for the model model' contains a duplicated word 'model'.
  6. [Table 2 caption] The caption reads 'Chunk-level accuracy on SWE-bench Verified,' but the table appears to report recall or perfect recall, not accuracy. Please use a consistent metric name.
  7. [Section 3.2] The call graph context uses only downstream neighbours indicated by [DOWN]; the paper does not explain why upstream/incoming edges are excluded. Consider clarifying or justifying this choice.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central claim is a held-out empirical benchmark result, and the paper's self-citations are not load-bearing.

full rationale

CoRet's central claim is an empirical retrieval benchmark result, not a derivation from first principles. The model is trained on SWE-bench train and evaluated on held-out SWE-bench Verified and Long Code Arena instances (Section 4.1: 'We evaluate our trained models on SWE-bench Verified, and Long Code Arena (LCA) bug localisation'), so the reported recall numbers are not forced by construction. The ground-truth label set is defined as the edited code chunks (Section 4.1: 'The ground truth pull requests are parsed to obtain the ground truth chunks C* which correspond to the edited code chunks'), which is a task definition rather than a circular step. Hyperparameters were selected on a hold-out from SWE-bench train (Appendix H), and no fitted parameter is renamed as a prediction. The call graph context and file path features are architectural choices that are ablated, not smuggled in via a self-citation chain. The only self-citations (SWE-PolyBench in the Limitations, Cesa-Bianchi et al. in a speculative sentence about call graph topology) are not load-bearing for the empirical claim. The abstract's 'at least 15 percentage points' statement is contradicted by Table 1 on LCA (+6pp at recall@5 and +13pp at recall@20), but that is an empirical-consistency and statistical-significance concern, not a circularity concern, and it does not raise the circularity score. The paper is self-contained against external benchmarks, so no circularity is identified.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper's empirical claims rest on hyperparameter choices (temperature, negatives, learning rate) and dataset assumptions (PR patch as ground truth, negative sampling). No new physical entities are introduced.

free parameters (4)
  • temperature tau = 0.05
    Set following Zhang et al. (2024a); no sensitivity analysis reported, so it is a fixed constant in the loss.
  • number of negatives |B| = 1024
    Ablation shows larger is better; 1024 is the maximum tested, indicating this value was chosen as an upper bound rather than an optimum.
  • learning rate = 5e-4
    Selected from a grid in Appendix G; it is a training hyperparameter, not a scientific constant.
  • training epochs = 4 (Full), 8 (Late Fusion)
    Different epochs for the two variants; chosen for convergence.
assumptions (4)
  • standard math The softmax over repository chunks is a valid likelihood for retrieval
    Equation 1 uses a categorical likelihood; standard in metric learning.
  • domain assumption PR-patch edited chunks constitute complete and correct ground truth
    Section 4.1 uses edited chunks as C*; this is an assumption about label quality.
  • domain assumption Randomly sampled in-repository chunks are true negatives
    Equation 2 samples B as negatives; if some sampled chunks are actually relevant, training is biased.
  • domain assumption Downstream call-graph neighbours are useful context
    Section 3.2 encodes only downstream neighbours with a [DOWN] token; this is an architectural choice tested by ablation but not derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CoRet: Improved Retriever for Code Editing." pith.science (2026). https://pith.science/paper/TO7VYHHU

@misc{pith2026250524715,
  author       = {Pith},
  title        = {Pith review of: CoRet: Improved Retriever for Code Editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TO7VYHHU}},
  note         = {Machine review of arXiv:2505.24715}
}
read the original abstract

In this paper, we introduce CoRet, a dense retrieval model designed for code-editing tasks that integrates code semantics, repository structure, and call graph dependencies. The model focuses on retrieving relevant portions of a code repository based on natural language queries such as requests to implement new features or fix bugs. These retrieved code chunks can then be presented to a user or to a second code-editing model or agent. To train CoRet, we propose a loss function explicitly designed for repository-level retrieval. On SWE-bench and Long Code Arena's bug localisation datasets, we show that our model substantially improves retrieval recall by at least 15 percentage points over existing models, and ablate the design choices to show their importance in achieving these results.

Figures

Figures reproduced from arXiv: 2505.24715 by the authors.

Figure 1
Figure 1. Top: Code repository before processing. Bot￾tom: Code chunks after filtering and chunking. negative pairs. We optimise the following loss function: L(θ) = 1 N X N i 1 |C∗ i | X c ∗∈C∗ i log exp  qi·c ∗ τ  Γ(q, Ci) , (1) where θ = (θq, θc) are the parameters of the model and Γ(q, C) = P c∈C exp q·c τ  is the normalising factor. This is the mean likelihood of retrieving the code chunk per model, and is akin to the … view at source ↗
Figure 2
Figure 2. Given a function ci and its downstream neigh￾bour cout, we concatenate the strings as ci ; [DOWN]; cout, including the special separation token, and fine-tune the model to obtain CoRet. endowed by the call graph (Ahn et al., 2009; Bansal et al., 2023). The neighbours in the call graph are the code entities that are invoked by or invoke the current chunk of interest. We propose to enrich each chunk ci in C with its c… view at source ↗
Figure 3
Figure 3. Recall@k for the baseline models and our proposed method on instances of SWE-bench Verified. CodeSage-family models substantially outperform other baselines. CoRet, described in Section 3, consistently outperforms baselines across all k. The dashed line corresponds to the trained CoRet model, whereas the solid lines correspond to the untrained baselines. SWE Verified LCA Model @5 @20 MRR @5 @20 MRR CodeSage S 0.34 0… view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: The influence of number of negatives |B| and their source. reproducibility. This Transformer encoder has 6 layers with 8 attention heads per layer. The size for the word embedding vectors and model projections are 1024, feed-forward dimensions are 4096 which leads to m…
Figure 4
Figure 4. Figure 4: Performance of various models on SWE￾bench Verified before fine-tuning. Top: chuck level re￾call; bottom: file-level recall. It is evident that the mod￾ern encoder models like the ones from the CodeSage family perform substantially better than other baselines. However,…
Figure 6
Figure 6. Figure 6: Attention map: query (problem statement) containing file path. Top: full average attention map (best [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Attention map: code chunk containing file path. Top: full average attention map (best viewed on a screen); [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 27 canonical work pages

  1. [1]

    Shin-Young Ahn, Sungwon Kang, Jongmoon Baik, and Ho‐Jin Choi. 2009. https://api.semanticscholar.org/CorpusID:17682894 A weighted call graph approach for finding relevant components in source code . 2009 10th ACIS International Conference on Software Engineering, Artificial Intelligences, Networking and Parallel/Distributed Computing, pages 539--544

  2. [2]

    Aakash Bansal, Zachary Eberhart, Zachary Karas, Yu Huang, and Collin McMillan. 2023. https://api.semanticscholar.org/CorpusID:258906815 Function call graph context encoding for neural source code summarization . IEEE Transactions on Software Engineering, 49:4268--4281

  3. [3]

    Aakash Bansal, Sakib Haque, and Collin McMillan. 2021. https://doi.org/10.1109/ICPC52881.2021.00032 Project-level encoding for neural source code summarization of subroutines . In 2021 IEEE/ACM 29th International Conference on Program Comprehension (ICPC), pages 253--264

  4. [4]

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. https://openreview.net/forum?id=IW1PR7vEBf LLM 2vec: Large language models are secretly powerful text encoders . In First Conference on Language Modeling

  5. [5]

    Egor Bogomolov, Aleksandra Eliseeva, Timur Galimzyanov, Evgeniy Glukhov, Anton Shapkin, Maria Tigina, Yaroslav Golubev, Alexander Kovrigin, Arie van Deursen, Maliheh Izadi, and Timofey Bryksin. 2024. https://arxiv.org/abs/2406.11612 Long code arena: a set of benchmarks for long-context code models . Preprint, arXiv:2406.11612

  6. [6]

    Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego De Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, Tom Hennigan, Saffron Huang, Loren Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, ...

  7. [7]

    Nicolo Cesa-Bianchi, Claudio Gentile, Fabio Vitale, and Giovanni Zappella. 2012. A correlation clustering approach to link classification in signed networks. In Conference on Learning Theory, pages 34--1. JMLR Workshop and Conference Proceedings

  8. [8]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597--1607. PMLR

Show all 62 references
  1. [9]

    Kai-Yang Chiang, Cho-Jui Hsieh, Nagarajan Natarajan, Inderjit S Dhillon, and Ambuj Tewari. 2014. Prediction and clustering in signed networks: a local to global perspective. The Journal of Machine Learning Research, 15(1):1177--1213

  2. [10]

    Chopra, R

    S. Chopra, R. Hadsell, and Y. LeCun. 2005. https://doi.org/10.1109/CVPR.2005.202 Learning a similarity metric discriminatively, with application to face verification . In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), volume 1, page...

  3. [11]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Associat...

  4. [12]

    Yangruibo Ding, Zijian Wang, Wasi Ahmad, Murali Krishna Ramanathan, Ramesh Nallapati, Parminder Bhatia, Dan Roth, and Bing Xiang. 2024. https://aclanthology.org/2024.lrec-main.305/ C o C o MIC : Code completion by jointly modeling in-file and cross-file context . In Proceeding...

  5. [13]

    Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, and Ming Zhou. 2020. https://doi.org/10.18653/v1/2020.findings-emnlp.139 C ode BERT : A pre-trained model for programming and natural languages . In Findings ...

  6. [14]

    Luyu Gao, Zhuyun Dai, Tongfei Chen, Zhen Fan, Benjamin Van Durme, and Jamie Callan. 2021. https://doi.org/10.1007/978-3-030-72113-8_10 Complement lexical retrieval model with semantic residual embeddings . In Advances in Information Retrieval: 43rd European Conference on IR Re...

  7. [15]

    Daya Guo, Shuai Lu, Nan Duan, Yanlin Wang, Ming Zhou, and Jian Yin. 2022. https://doi.org/10.18653/v1/2022.acl-long.499 U ni X coder: Unified cross-modal pre-training for code representation . In Proceedings of the 60th Annual Meeting of the Association for Computational Lingu...

  8. [16]

    Daya Guo, Shuo Ren, Shuai Lu, Zhangyin Feng, Duyu Tang, Shujie LIU, Long Zhou, Nan Duan, Alexey Svyatkovskiy, Shengyu Fu, Michele Tufano, Shao Kun Deng, Colin Clement, Dawn Drain, Neel Sundaresan, Jian Yin, Daxin Jiang, and Ming Zhou. 2021. https://openreview.net/forum?id=jLoC...

  9. [17]

    Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, Y. K. Li, Fuli Luo, Yingfei Xiong, and Wenfeng Liang. 2024. https://api.semanticscholar.org/CorpusID:267211867 Deepseek-coder: When the large language model meets programming -...

  10. [18]

    Michael Günther, Isabelle Mohr, Daniel James Williams, Bo Wang, and Han Xiao. 2024. https://arxiv.org/abs/2409.04701 Late chunking: Contextual chunk embeddings using long-context embedding models . Preprint, arXiv:2409.04701

  11. [19]

    Junjie Huang, Duyu Tang, Linjun Shou, Ming Gong, Ke Xu, Daxin Jiang, Ming Zhou, and Nan Duan. 2021. https://doi.org/10.18653/v1/2021.acl-long.442 C o SQA : 20,000+ web queries for code search and question answering . In Proceedings of the 59th Annual Meeting of the Association...

  12. [20]

    Hamel Husain, Ho - Hsiang Wu, Tiferet Gazit, Miltiadis Allamanis, and Marc Brockschmidt. 2019. https://arxiv.org/abs/1909.09436 Codesearchnet challenge: Evaluating the state of semantic code search . CoRR, abs/1909.09436

  13. [21]

    Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. https://api.semanticscholar.org/CorpusID:249097975 Unsupervised dense information retrieval with contrastive learning . Trans. Mach. Learn. Res., 2022

  14. [22]

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

  15. [23]

    Aditya Kanade, Petros Maniatis, Gogul Balakrishnan, and Kensen Shi. 2020. https://openreview.net/forum?id=rygoURNYvS Pre-trained contextual embedding of source code

  16. [24]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.550 Dense passage retrieval for open-domain question answering . In Proceedings of the 2020 Conference on Empiric...

  17. [25]

    Omar Khattab and Matei Zaharia. 2020. https://doi.org/10.1145/3397271.3401075 Colbert: Efficient and effective passage search via contextualized late interaction over bert . In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Informatio...

  18. [26]

    Mike Lewis, Marjan Ghazvininejad, Gargi Ghosh, Armen Aghajanyan, Sida Wang, and Luke Zettlemoyer. 2020 a . https://proceedings.neurips.cc/paper_files/paper/2020/file/d6f1dd034aabde7657e6680444ceff62-Paper.pdf Pre-training via paraphrasing . In Advances in Neural Information Pr...

  19. [27]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K\" u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\" a schel, Sebastian Riedel, and Douwe Kiela. 2020 b . Retrieval-augmented generation for knowledge-intensive nlp tasks. In P...

  20. [28]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K\" u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\" a schel, Sebastian Riedel, and Douwe Kiela. 2020 c . https://proceedings.neurips.cc/paper_files/paper/2020/file/6b493230205...

  21. [29]

    Jia Li, Ge Li, Xuanming Zhang, Yunfei Zhao, Yihong Dong, Zhi Jin, Binhua Li, Fei Huang, and Yongbin Li. 2024 a . https://openreview.net/forum?id=kvjbFVHpny Evocodebench: An evolving code generation benchmark with domain-specific evaluations . In The Thirty-eight Conference on ...

  22. [30]

    Raymond Li, Loubna Ben Allal, Yangtian Zi, Niklas Muennighoff, Denis Kocetkov, Chenghao Mou, Marc Marone, Christopher Akiki, Jia Li, Jenny Chim, Qian Liu, Evgenii Zheltonozhskii, Terry Yue Zhuo, Thomas Wang, Olivier Dehaene, Mishig Davaadorj, Joel Lamy-Poirier, João Monteiro, ...

  23. [31]

    Xiangyang Li, Kuicai Dong, Yi Quan Lee, Wei Xia, Yichun Yin, Hao Zhang, Yong Liu, Yasheng Wang, and Ruiming Tang. 2024 b . https://arxiv.org/abs/2407.02883 Coir: A comprehensive benchmark for code information retrieval models . Preprint, arXiv:2407.02883

  24. [32]

    Dianshu Liao, Shidong Pan, Xiaoyu Sun, Xiaoxue Ren, Qing Huang, Zhenchang Xing, Huan Jin, and Qinying Li. 2024. https://arxiv.org/abs/2312.05772 A3-codgen: A repository-level code generation framework for code reuse with local-aware, global-aware, and third-party-library-aware...

  25. [33]

    Tianyang Liu, Canwen Xu, and Julian McAuley. 2024. https://openreview.net/forum?id=pPjZIOuQuF Repobench: Benchmarking repository-level code auto-completion systems . In The Twelfth International Conference on Learning Representations

  26. [34]

    Xiangyan Liu, Bo Lan, Zhiyuan Hu, Yang Liu, Zhicheng Zhang, Fei Wang, Michael Qizhe Shieh, and Wenmeng Zhou. 2025. https://aclanthology.org/2025.naacl-long.7/ C odex G raph: Bridging large language models and code repositories via code graph databases . In Proceedings of the 2...

  27. [35]

    Shuai Lu, Daya Guo, Shuo Ren, Junjie Huang, Alexey Svyatkovskiy, Ambrosio Blanco, Colin Clement, Dawn Drain, Daxin Jiang, Duyu Tang, Ge Li, Lidong Zhou, Linjun Shou, Long Zhou, Michele Tufano, MING GONG, Ming Zhou, Nan Duan, Neel Sundaresan, Shao Kun Deng, Shengyu Fu, and Shuj...

  28. [36]

    Yi Luan, Jacob Eisenstein, Kristina Toutanova, and Michael Collins. 2021. https://doi.org/10.1162/tacl_a_00369 Sparse, dense, and attentional representations for text retrieval . Transactions of the Association for Computational Linguistics, 9:329--345

  29. [37]

    Yingwei Ma, Qingping Yang, Rongyu Cao, Binhua Li, Fei Huang, and Yongbin Li. 2024. https://api.semanticscholar.org/CorpusID:270215039 How to understand whole software repository? ArXiv, abs/2406.01422

  30. [38]

    Siru Ouyang, Wenhao Yu, Kaixin Ma, Zilin Xiao, Zhihan Zhang, Mengzhao Jia, Jiawei Han, Hongming Zhang, and Dong Yu. 2024. Repograph: Enhancing ai software engineering with repository-level code graph. arXiv preprint arXiv:2410.14684

  31. [39]

    Radev, Hong Qi, Harris Wu, and Weiguo Fan

    Dragomir R. Radev, Hong Qi, Harris Wu, and Weiguo Fan. 2002. https://aclanthology.org/L02-1301/ Evaluating web-based question answering systems . In Proceedings of the Third International Conference on Language Resources and Evaluation ( LREC `02) , Las Palmas, Canary Islands ...

  32. [40]

    Muhammad Shihab Rashid, Christian Bock, Yuan Zhuang, Alexander Buchholz, Tim Esler, Simon Valentin, Luca Franceschi, Martin Wistuba, Prabhu Teja Sivaprasad, Woo Jung Kim, Anoop Deoras, Giovanni Zappella, and Laurent Callot. 2025. https://arxiv.org/abs/2504.08703 Swe-polybench:...

  33. [41]

    Nils Reimers and Iryna Gurevych. 2019. https://api.semanticscholar.org/CorpusID:201646309 Sentence-bert: Sentence embeddings using siamese bert-networks . In Conference on Empirical Methods in Natural Language Processing

  34. [42]

    Joshua David Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. 2021. https://openreview.net/forum?id=CR1XOQ0UTh- Contrastive learning with hard negative samples . In International Conference on Learning Representations

  35. [43]

    Haifeng Ruan, Yuntong Zhang, and Abhik Roychoudhury. 2024. https://arxiv.org/abs/2408.02232 Specrover: Code intent extraction via llms . Preprint, arXiv:2408.02232

  36. [44]

    Disha Shrivastava, Denis Kocetkov, Harm de Vries, Dzmitry Bahdanau, and Torsten Scholak. 2023. https://arxiv.org/abs/2306.10998 Repofusion: Training code models to understand your repository . Preprint, arXiv:2306.10998

  37. [45]

    Kihyuk Sohn. 2016 a . https://proceedings.neurips.cc/paper_files/paper/2016/file/6b180037abbebea991d8b1232f8a8ca9-Paper.pdf Improved deep metric learning with multi-class n-pair loss objective . In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc

  38. [46]

    Kihyuk Sohn. 2016 b . https://proceedings.neurips.cc/paper_files/paper/2016/file/6b180037abbebea991d8b1232f8a8ca9-Paper.pdf Improved deep metric learning with multi-class n-pair loss objective . In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc

  39. [47]

    Tarun Suresh, Revanth Gangi Reddy, Yifei Xu, Zach Nussbaum, Andriy Mulyar, Brandon Duderstadt, and Heng Ji. 2025. https://openreview.net/forum?id=iyJOUELYir Co RNS tack: High-quality contrastive data for better code retrieval and reranking . In The Thirteenth International Con...

  40. [48]

    Chongyang Tao, Tao Shen, Shen Gao, Junshuo Zhang, Zhen Li, Zhengwei Tao, and Shuai Ma. 2024. Llms are also effective embedding models: An in-depth overview. arXiv preprint arXiv:2412.12591

  41. [49]

    Andrew Trotman, Antti Puurula, and Blake Burgess. 2014. https://api.semanticscholar.org/CorpusID:207220720 Improvements to bm25 and language models examined . Proceedings of the 19th Australasian Document Computing Symposium

  42. [50]

    Charalampos E Tsourakakis, Michael Mitzenmacher, Kasper Green Larsen, Jaros aw B asiok, Ben Lawson, Preetum Nakkiran, and Vasileios Nakos. 2017. Predicting positive and negative links with noisy queries: Theory & practice. arXiv preprint arXiv:1709.07308

  43. [51]

    A \"a ron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. https://api.semanticscholar.org/CorpusID:49670925 Representation learning with contrastive predictive coding . ArXiv, abs/1807.03748

  44. [52]

    Xu, Yiqing Xie, Graham Neubig, and Daniel Fried

    Zora Zhiruo Wang, Akari Asai, Xinyan Velocity Yu, Frank F. Xu, Yiqing Xie, Graham Neubig, and Daniel Fried. 2024. https://arxiv.org/abs/2406.14497 Coderag-bench: Can retrieval augment code generation? Preprint, arXiv:2406.14497

  45. [53]

    Chunqiu Steven Xia, Yinlin Deng, Soren Dunn, and Lingming Zhang. 2024. Agentless: Demystifying llm-based software engineering agents. arXiv preprint

  46. [54]

    Bennett, Junaid Ahmed, and Arnold Overwijk

    Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. https://openreview.net/forum?id=zeFrfgyZln Approximate nearest neighbor negative contrastive learning for dense text retrieval . In International Conference o...

  47. [55]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze...

  48. [56]

    Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press

    John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. 2024 b . https://arxiv.org/abs/2405.15793 Swe-agent: Agent-computer interfaces enable automated software engineering . Preprint, arXiv:2405.15793

  49. [57]

    Dejiao Zhang, Wasi Uddin Ahmad, Ming Tan, Hantian Ding, Ramesh Nallapati, Dan Roth, Xiaofei Ma, and Bing Xiang. 2024 a . https://openreview.net/forum?id=vfzRRjumpX CODE REPRESENTATION LEARNING AT SCALE . In The Twelfth International Conference on Learning Representations

  50. [58]

    Chen, Yue Zhang, Jin Liu, Daoguang Zan, Yi Mao, Jian-Guang Lou, and Weizhu Chen

    Fengji Zhang, B. Chen, Yue Zhang, Jin Liu, Daoguang Zan, Yi Mao, Jian-Guang Lou, and Weizhu Chen. 2023 a . https://api.semanticscholar.org/CorpusID:257663528 Repocoder: Repository-level code completion through iterative retrieval and generation . In Conference on Empirical Met...

  51. [59]

    Fengji Zhang, Bei Chen, Yue Zhang, Jacky Keung, Jin Liu, Daoguang Zan, Yi Mao, Jian-Guang Lou, and Weizhu Chen. 2023 b . https://openreview.net/forum?id=q09vTY1Cqh Repocoder: Repository-level code completion through iterative retrieval and generation . In The 2023 Conference o...

  52. [60]

    Shuai Zhang, Wang Lijie, Xinyan Xiao, and Hua Wu. 2022. https://doi.org/10.18653/v1/2022.findings-acl.191 Syntax-guided contrastive learning for pre-trained language model . In Findings of the Association for Computational Linguistics: ACL 2022, pages 2430--2440, Dublin, Irela...

  53. [61]

    Yuntong Zhang, Haifeng Ruan, Zhiyu Fan, and Abhik Roychoudhury. 2024 b . https://arxiv.org/abs/2404.05427 Autocoderover: Autonomous program improvement . Preprint, arXiv:2404.05427

  54. [62]

    Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Haonan Chen, Zheng Liu, Zhicheng Dou, and Ji-Rong Wen. 2024. https://arxiv.org/abs/2308.07107 Large language models for information retrieval: A survey . Preprint, arXiv:2308.07107

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.