Pith. sign in

REVIEW 4 major objections 4 minor 19 references

A dataset of 5,264 real Go code-and-test pairs lifts fine-tuned LLMs past a 75% win rate on unit-test generation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-03 22:20 UTC pith:CMAATGMK

load-bearing objection Go-UT-Bench is a genuinely useful Go unit-test fine-tuning dataset with a reproducible pipeline, but the headline win rates only prove that fine-tuned models better imitate repository test style under one LLM judge, not that they generate better tests. the 4 major comments →

arxiv 2511.10868 v2 pith:CMAATGMK submitted 2025-11-14 cs.LG

Go-UT-Bench: A Fine-Tuning Dataset for LLM-Based Unit Test Generation in Go

classification cs.LG
keywords Go unit test generationfine-tuning datasetcode LLMLLM-as-a-judgesupervised fine-tuningLoRAGolangunit test generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces Go-UT-Bench, an open dataset of 5,264 {Go source file, unit test} pairs curated from ten permissively licensed, industrial-scale Go projects, with commit hashes so each pair can be traced to a specific version. Its central claim is that fine-tuning LLMs on this dataset markedly improves unit-test generation: a Mixture-of-Experts coder fine-tuned with low-rank adaptation (LoRA) wins 81.9% of pairwise comparisons against its base version, and a smaller dense instruction model fine-tuned fully wins 76.7%, as judged by GPT-4o-mini's alignment with the repository's ground-truth tests. The authors argue this shows domain-specific fine-tuning can move code LLMs from autocompletion toward a real developer workflow — writing tests for Go's concurrency-heavy, statically typed code. The dataset is designed for reproducibility, and the paper is candid that no generated test is compiled or executed.

Core claim

On its own terms, the paper's discovery is that a modest, carefully curated set of real code-test pairs — not synthetic examples — is sufficient to shift an LLM's unit-test generation toward the style and coverage of repository-maintained tests. Concretely, the low-rank-adapted (LoRA) DeepSeek-Coder-V2-Lite-Instruct beats its base on 81.9% of 790 validation tasks, and the fully tuned Llama-3.2-3B-Instruct beats its base on 76.7%, where 'beats' means GPT-4o-mini judged the output more aligned with the ground-truth test file. The same pipeline contributes an Abstract Syntax Tree (AST)-guided chunking method for files too long for reliable single-pass generation.

What carries the argument

The load-bearing object is Go-UT-Bench itself: 5,264 {code, unit test} pairs, each tagged with repository, file path, commit hashes, and a SHA256 identifier, drawn from ten real Go projects. The argument is carried by an oracle-based pairwise evaluation protocol (Algorithm 1) in which GPT-4o-mini scores how closely a generated test matches the ground-truth test; the model with more 'closer alignment' judgments wins. For long files, an Abstract Syntax Tree (AST)-based chunker partitions source into dependency-connected segments before test generation.

Load-bearing premise

The whole evaluation rests on GPT-4o-mini's 'closer alignment to the ground-truth test' being a valid proxy for unit-test quality, because no generated test is compiled, run, or judged by a human — a limitation the paper itself flags as Evaluation Bias.

What would settle it

Take the 790 validation inputs, compile the generated test files with `go test`, and compare pass rates of fine-tuned vs base models. If fine-tuned tests pass at or below base rates, the claimed improvement is not supported by working tests. A cheaper first check: have two human Go developers independently rank a random sample of 50 pairs and compare their rankings to GPT-4o-mini's.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Fine-tuning on Go-UT-Bench raises win rates from 14.2% to 81.9% (DeepSeek-Coder-V2-Lite-Instruct, LoRA) and from 22.0% to 76.7% (Llama-3.2-3B-Instruct, full fine-tuning) on 790 validation pairs.
  • The gains hold across most of the ten repositories; the weakest results appear for repos with few validation pairs (Moby, Gin), suggesting per-domain data scale matters.
  • Because every pair carries commit hashes, future evaluations can be re-run against the exact code versions — a reproducibility property the paper argues is rare in this space.
  • The AST-guided chunking pipeline makes the approach applicable to real Go files that exceed context limits.
  • The reported numbers are alignment-with-ground-truth judgments, not proof that generated tests compile or pass.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The large win-rate gap may partly reflect the judge rewarding stylistic imitation of repository tests; a test that looks like the ground truth is not necessarily a better test. An execution-based filter (e.g., `go test`) would separate imitation from correctness.
  • The AST chunking plus fine-tuning recipe could transfer to other statically typed, concurrency-heavy languages (e.g., Rust) where unit-test data is also scarce.
  • The dataset's severe repository imbalance (e.g., Kubernetes contributes roughly 37% of pairs) may make fine-tuned models best at infrastructure-style testing; a balanced or stratified version could change the win rates.
  • Pairing Go-UT-Bench with mutation testing — deleting a line of code and checking whether the generated test fails — would give an execution-based quality signal the current oracle does not provide.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces Go-UT-Bench, a dataset of 5,264 {Go source file, unit test} pairs extracted from 10 permissively licensed Go repositories, with commit-hash metadata for reproducibility. The authors propose a two-stage curation pipeline and evaluate the dataset's usefulness for supervised fine-tuning by fine-tuning two models—DeepSeek-Coder-V2-Lite-Instruct (LoRA) and Llama-3.2-3B-Instruct (full fine-tuning)—and comparing them to their base counterparts using an LLM-as-a-judge protocol (GPT-4o-mini) that scores alignment between generated and repository ground-truth tests. They report win rates of 81.9% and 76.7% over 790 validation pairs, respectively, and interpret this as evidence that fine-tuning on Go-UT-Bench substantially improves Go unit test generation. The paper also describes an AST-guided chunking pipeline for long files and discusses limitations including evaluation bias, data leakage risk, and representation imbalance.

Significance. If the central claim were supported by execution-based or human-validated evaluation, Go-UT-Bench would be a useful resource for a relatively under-served language in code LLM research. The dataset design is sensible: permissively licensed repositories, commit-hash provenance, stratified splits, and a reproducible curation pipeline. The authors also deserve credit for explicitly acknowledging the main weaknesses of their evaluation in Section 6, including the absence of compile/run or human assessment, lack of confidence intervals, and leakage risk. However, the paper's headline result—that fine-tuned models outperform base models on more than 75% of benchmark tasks—is currently supported only by GPT-4o-mini's pairwise similarity judgments against ground-truth tests drawn from the same repositories used for SFT. This does not establish that the generated tests compile, run, or assert correct behavior, and it may partially measure imitation of repository-specific test style. The resource is potentially valuable, but the effectiveness claim needs substantiation or careful reframing.

major comments (4)
  1. [Section 4 and Algorithm 1] The central effectiveness claim rests entirely on the oracle in Algorithm 1: GPT-4o-mini assigns alignment scores of each generated test to the repository ground-truth test, and the model with more high-scoring outputs wins. No generated test is compiled or executed, and no human evaluation is reported. Because the SFT training pairs are extracted from the same ten repositories used for scoring, the fine-tuned model can inflate oracle alignment by reproducing repository-specific test idioms and structural templates rather than by producing tests that are correct in any executable sense. The win rates 81.9% and 76.7% therefore support 'better mimicry of the ground-truth distribution under one LLM judge,' not the broader claim of improved unit-test generation stated in the abstract and Section 4. I recommend adding at least one execution-based signal (e.g., compilation success or pass rate
  2. [Section 6 (Data Leakage and Overfitting Risks)] The paper acknowledges that pretrained LLMs may have seen portions of the public repositories, but this risk is not addressed in the experimental protocol. Since GPT-4o-mini is also likely to have seen popular repositories such as Kubernetes and TiDB, its judgments of 'alignment with ground truth' may favor outputs that resemble memorized tests. More importantly, random stratified splitting does not prevent leakage from training into validation if a pretrained model already memorized the validation ground truth; comparing base and fine-tuned versions does not remove this confound. A temporal split, deduplication against model training corpora, or at least a demonstration that the judge's scores correlate with an external correctness signal would materially strengthen the claim.
  3. [Section 3.2 and Figures 4-5] The reported 'win rate' is a single aggregate number over 790 validation pairs, with no confidence intervals, significance testing, or error analysis beyond an undefined 'error rate' (3.9% for DeepSeek, 1.3% for Llama). The repository-level results are also fragile for low-count repositories: gin-Gonic/gin has only 4 validation pairs, so a 25% win rate there is not a reliable signal. I recommend reporting per-repository counts alongside CIs or bootstrap intervals, and clarifying what constitutes an 'error' in the evaluation workflow.
  4. [Section 3.1, Equations (1)-(3)] The AST-chunking and concatenation pipeline is described only at a high level. Equation (3) simply appends generated unit tests for chunks; it is not shown whether the resulting test files have valid Go syntax, duplicate test function names, or import conflicts. Since the evaluation never compiles the generated output, the pipeline itself is not validated. If execution-based evaluation is added, chunk boundary handling will likely need more detail (e.g., how dependencies and context across chunks are preserved).
minor comments (4)
  1. [Throughout] There are naming inconsistencies: 'GO UT Bench', 'Go-UT-Bench', and 'GO-UNITTEST-BENCH' are used interchangeably; repository names are inconsistently capitalized (e.g., 'Golang/Go', 'GohuGoio/huGo', 'hashicorp/terraform'). Please unify terminology and repository casing.
  2. [References] Two references are malformed: 'et al. Li' and 'et al. Lin' replace author lists. Full author information is needed for these entries. Also, the arXiv ID and date are missing for the Becker et al. RCT report.
  3. [Section 7] The reproducibility statement gives only a Hugging Face URL and no license or access instructions for the dataset. Since permissive licensing of the underlying repositories is a stated design goal, please state the exact license(s) of the dataset artifacts and how the commit hashes can be used to reconstruct the extraction.
  4. [Section 3.2] The term 'error rate' in Figures 4 and 5 is not defined in the text or Algorithm 1. It appears in the caption and results discussion but no formula or description is provided. Please define it or remove it.

Circularity Check

1 steps flagged

Evaluation's only quality signal is LLM-judged similarity to ground-truth tests drawn from the same repositories used for SFT, so the win-rate claim partially reduces to measuring imitation of the training distribution.

specific steps
  1. self definitional [Section 3 (Algorithm 1) and Section 3.2; acknowledged in Section 6 (Evaluation Bias)]
    "s A ←J(r A, gi)▷Alignment score of A to ground truth ... In our pairwise evaluation approach, the win rate for a model is defined as the percentage of instances in which its response is judged to be more closely aligned with the ground truth than that of the competing model."

    The SFT objective trains on {code, unit-test} pairs whose test side is the same ground-truth distribution later used as the reference g_i in the oracle (Algorithm 1). The only quality signal is GPT-4o-mini's judgment of 'alignment with the ground truth'; no generated test is compiled/run and no human judge is used (Section 6 admits both). Thus the reported win-rate improvement largely measures the fine-tuned model's ability to imitate the repository test style present in its training distribution, rather than an externally validated improvement in unit-test correctness. The held-out split prevents direct memorization of the specific validation pairs, so this is partial: the result is not fully forced, but the central claim is partly circular because the evaluation metric is defined in term

full rationale

The paper contains no self-citation load-bearing steps, no imported uniqueness theorems, and no ansatz smuggled in via citation. The dataset construction and fine-tuning pipeline are self-contained and externally reproducible via commit hashes. The only circularity-like issue is in the evaluation design: fine-tuning optimizes toward ground-truth unit tests from ten repositories, and the oracle then scores outputs by their judged alignment with ground-truth unit tests from the same repositories. This makes the win-rate numbers partially a measure of how well the model reproduces repository-specific test style. However, the validation and test splits are held out, so the result is not purely tautological, and the paper candidly lists the missing triangulation in Section 6. Consistent with the rubric, this is a partial evaluation-metric circularity rather than a fully forced derivation, so a score of 4 is appropriate.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

No physically new entities are introduced. The load-bearing assumptions are about evaluation validity, code-test pairing, model representativeness, chunking effectiveness, and licensing. The free parameters are hand-chosen pipeline/training settings rather than fitted scientific constants.

free parameters (2)
  • AST chunking token threshold = 15,360 tokens
    Chosen empirically ('we have empirically observed') as the size above which files are chunked; this affects both generation and evaluation inputs and is not systematically tuned.
  • Fine-tuning hyperparameters = epochs=2, lr=5e-05, batch_size=8, LoRA rank=8, alpha=16
    Hand-selected settings for LoRA and full fine-tuning; the reported win rates are conditional on these choices and no hyperparameter search is reported.
axioms (5)
  • domain assumption GPT-4o-mini pairwise alignment to ground truth is a valid measure of unit-test quality
    Used in Algorithm 1 and Section 3; the authors acknowledge in Section 6 that it lacks human and execution-based validation.
  • domain assumption Grouping Go files by basename yields correct code-test pairings
    Section 2.1 pairs source files with test files by base name only; tests with different basenames or unconventional test-file layouts would be missed or mispaired, and this is not validated.
  • domain assumption The two selected models represent the relevant LLM families
    Section 3.2 tests one MoE model and one dense decoder; conclusions about the dataset's general effectiveness rest on this small sample.
  • domain assumption AST-based chunking preserves the model's ability to generate coherent tests across chunk boundaries
    Section 3.1 describes chunking and concatenation (Equations 1-3) but provides no ablation or correctness check for chunked inputs.
  • domain assumption Permissive licenses permit redistribution of the code-test pairs
    Section 1 asserts permissive licensing, but the paper presents no license audit or per-repository license breakdown.

pith-pipeline@v1.3.0-alltime-deepseek · 10861 in / 14164 out tokens · 126240 ms · 2026-08-03T22:20:24.464766+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Go-UT-Bench: A Fine-Tuning Dataset for LLM-Based Unit Test Generation in Go." pith.science (2026). https://pith.science/paper/CMAATGMK

@misc{pith2026251110868,
  author       = {Pith},
  title        = {Pith review of: Go-UT-Bench: A Fine-Tuning Dataset for LLM-Based Unit Test Generation in Go},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CMAATGMK}},
  note         = {Machine review of arXiv:2511.10868}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Training data imbalance poses a major challenge for code LLMs. Most available data heavily over represents raw opensource code while underrepresenting broader software engineering tasks, especially in low resource languages like Golang. As a result, models excel at code autocompletion but struggle with real world developer workflows such as unit test generation. To address this gap, we introduce GO UT Bench, a benchmark dataset of 5264 pairs of code and unit tests, drawn from 10 permissively licensed Golang repositories spanning diverse domain. We evaluate its effectiveness as a fine tuning dataset across two LLM families i.e. mixture of experts and dense decoders. Our results show that finetuned models outperform their base counterparts on more than 75% of benchmark tasks.

Figures

Figures reproduced from arXiv: 2511.10868 by Debojyoti Dutta, Hritik Raj, Rajat Ghosh, Vaishnavi Bhargava, Yashshi Pipalani.

Figure 1
Figure 1. Figure 1: PCA based analysis of Go-UT-Bench revealing the internal structure of the dataset. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The diversity in {code, unit test} pairs in terms of line lengths across 10 different opensource [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Prompt for pairwise evaluation of two LLM generated responses (Assistant A and Assistant [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Fine-tuning results for deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct. The left plot com [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Fine-tuning results for meta-llama/Llama-3.2-3B-Instruct. The left plot compares the win [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Example of a code and unit test pair from the dataset [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Prompt template for the unit test generation in Golang. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The training dataset consists of 3,684 code, unit test pairs extracted from 10 diverse [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Distribution of 790 test set samples across repositories in Go-UT-Bench. Kubernetes has [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Distribution of validation set samples across repositories in Go-UT-Bench. The dataset [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

19 extracted references · 12 linked inside Pith

  1. [4]

    Mike Conover, Matt Hayes, Ankit Mathur, Jianwei Xie, Jun Wan, Sam Shah, Ali Ghodsi, Patrick Wendell, Matei Zaharia, and Reynold Xin

    Accessed: 2025-08-27. Mike Conover, Matt Hayes, Ankit Mathur, Jianwei Xie, Jun Wan, Sam Shah, Ali Ghodsi, Patrick Wendell, Matei Zaharia, and Reynold Xin. Free dolly: Introducing the world’s first truly open instruction-tuned llm,

  2. [5]

    Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Bingxiang He, Wei Zhu, Yuan Ni, Guotong Xie, Ruobing Xie, Yankai Lin, et al

    URL https://www.databricks.com/blog/2023/04/12/ dolly-first-open-commercially-viable-instruction-tuned-llm. Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Bingxiang He, Wei Zhu, Yuan Ni, Guotong Xie, Ruobing Xie, Yankai Lin, et al. Ultrafeedback: Boosting language models with scaled ai feedback. arXiv preprint arXiv:2310.01377,

  3. [9]

    10 et al

    URL https://arxiv.org/abs/ 2411.15124. 10 et al. Li. Api-guided dataset synthesis to finetune large code models.arXiv preprint arXiv:2408.08343,

  4. [10]

    URLhttps://arxiv.org/abs/2305.06161. et al. Lin. Domain-specific fine-tuning in a retrieval-augmented generation framework.Preprints.org,

  5. [11]

    URLhttps://arxiv.org/abs/2301.13688. Anton Lozhkov, Raymond Li, Loubna Ben Allal, Federico Cassano, Joel Lamy-Poirier, Nouamane Tazi, Ao Tang, Dmytro Pykhtar, Jiawei Liu, Yuxiang Wei, Tianyang Liu, Max Tian, Denis Kocetkov, Arthur Zucker, Younes Belkada, Zijian Wang, Qian Liu, Dmitry Abulkhanov, Indraneil Paul, Zhuang Li, Wen-Ding Li, Megan Risdal, Jia Li...

  6. [12]

    Niels Mündler, Mark Niklas Müller, Jingxuan He, and Martin Vechev

    URLhttps://arxiv.org/abs/2402.19173. Niels Mündler, Mark Niklas Müller, Jingxuan He, and Martin Vechev. SWT-bench: Testing and validating real-world bug-fixes with code agents. InNeural Information Processing Systems (NeurIPS),

  7. [13]

    Predibase

    URLhttps://arxiv.org/abs/2203.13474. Predibase. Fine-tuning overview. https://docs.predibase.com/guides/fine-tuning/ overview,

  8. [14]

    Spaceo.ai

    Accessed: 2025-08-27. Spaceo.ai. Llm fine-tuning: Complete guide to optimizing language. https://www.spaceo.ai/ blog/llm-fine-tuning/,

  9. [15]

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B

    Accessed: 2025-08-27. Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Alpaca: A strong, replicable instruction-following model. Stanford CRFM blog post,

  10. [16]

    Yiqing Xie, Alex Xie, Divyanshu Sheth, Pengfei Liu, Daniel Fried, and Carolyn Rose

    Describes fine-tuning LLaMA-7B using 52K self-instruct demonstrations, demonstrating performance on par with text-davinci-003. Yiqing Xie, Alex Xie, Divyanshu Sheth, Pengfei Liu, Daniel Fried, and Carolyn Rose. Codebenchgen: Creating scalable execution-based code generation benchmarks.arXiv preprint arXiv:2404.00566,

  11. [17]

    Instruction tuning for large language models: A survey.arXiv preprint arXiv:2308.10792,

    11 Shengyu Zhang et al. Instruction tuning for large language models: A survey.arXiv preprint arXiv:2308.10792,

  12. [18]

    Judging llm-as-a-judge with mt-bench and chatbot arena.Advances in Neural Information Processing Systems, 36:46595–46623, 2023a

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena.Advances in Neural Information Processing Systems, 36:46595–46623, 2023a. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi L...

  13. [19]

    linear" seed 42 logging_steps 100 save_steps 1000 save_strategy

    12 8 Appendix 8.1 Example of Code and Unit test pair from the dataset Figure 6: Example of a code and unit test pair from the dataset 13 8.2 Unit Test Generation Prompt Figure 7: Prompt template for the unit test generation in Golang. 14 8.3 Evaluation Workflow We do comparison between the base-generated unit tests and the LoRA trained DeepSeek-Coder-V2- ...

  14. [20]

    none" task_type

    (Looped over this value) lora_dropout 0.05 bias "none" task_type "CAUSAL_LM" target_modules ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"] 16 8.5 Parameters used in Full Fine-Tuning of meta-llama/Llama-3.2-3B-Instruct Table 3: Llama-3.2 Full fine-tuning Parameters Parameter Value epochs 2 learning_rate 5e-05 batch_size 8 wei...

  15. [2021]

    Carlos E

    URL https: //arxiv.org/abs/2106.09685. Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik R. Narasimhan. SWE-bench: Can language models resolve real-world github issues? InThe Twelfth International Conference on Learning Representations (ICLR),

  16. [2022]

    Fine-tuning language models with limited data.arXiv preprint arXiv:2406.11201,

    Steven Barnett et al. Fine-tuning language models with limited data.arXiv preprint arXiv:2406.11201,

  17. [2023]

    Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374,

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374,

  18. [2024]

    Measuring the impact of early-2025 ai on experienced open-source developer productivity

    Joel Becker, Nate Rush, Elizabeth Barnes, and David Rein. Measuring the impact of early-2025 ai on experienced open-source developer productivity. Randomized Controlled Trial report,

  19. [2025]

    arXiv preprint arXiv:2304.08247v3, last revised March 18,