pith. machine review for the scientific record. sign in

arxiv: 2604.13725 · v1 · submitted 2026-04-15 · 💻 cs.SE

Recognition: unknown

On the Effectiveness of Context Compression for Repository-Level Tasks: An Empirical Investigation

Jia Feng , Zhanyue Qin , Cuiyun Gao , Ruiqi Wang , Chaozheng Wang , Yingwei Ma , Xiaoyuan Xie
Authors on Pith no claims yet

Pith reviewed 2026-05-10 12:56 UTC · model grok-4.3

classification 💻 cs.SE
keywords context compressionrepository-level code taskscode completioncode generationlarge language modelsempirical evaluationcontinuous latent vectorsinference efficiency
0
0 comments X

The pith

Context compression improves repository-level code tasks by filtering noise, with continuous latent vector methods outperforming full context at 4x ratios.

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

Repository-level code intelligence requires models to handle long multi-file inputs that often bury key details in noise, exceed context windows, and raise latency. The paper conducts the first broad empirical test of eight compression techniques grouped into discrete token sequences, continuous latent vectors, and visual tokens. It applies them to code completion and generation benchmarks and measures both accuracy and speed. Results indicate that compression succeeds overall, and at moderate ratios the continuous vector approaches actually raise scores by removing irrelevant material rather than losing signal. Efficiency improves across all methods, with some reaching half the latency of uncompressed inputs while approaching the cost of running without repository context.

Core claim

Context compression is effective for repository-level code intelligence tasks. Continuous latent vector methods at 4x compression surpass full-context performance by up to 28.3 percent in BLEU score on code completion and generation, demonstrating that they filter noise rather than merely truncate. All three paradigms reduce inference cost, and visual plus text-based approaches achieve up to 50 percent end-to-end latency reduction at high ratios, nearing the expense of inference without repository context.

What carries the argument

Three paradigms of context compression—discrete token sequences, continuous latent vectors, and visual tokens—applied to condense long multi-file repository inputs for language models performing code completion and generation.

If this is right

  • Continuous latent vector compression can raise task accuracy instead of lowering it by removing noise from repository context.
  • All compression paradigms lower inference costs relative to uncompressed full-context inputs.
  • Visual and text-based methods can reduce end-to-end latency by up to 50 percent at high ratios while maintaining or exceeding full-context accuracy.
  • Compressed inputs can approach the cost of running without any repository context while preserving or improving quality on the tested tasks.

Where Pith is reading between the lines

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

  • The noise-filtering benefit observed here could apply to other repository-scale code activities such as bug localization or refactoring if similar irrelevant context patterns appear.
  • Alternative evaluation measures beyond BLEU, such as execution success or human preference, would clarify whether the reported gains matter in actual development workflows.
  • Hybrid systems that combine compression with selective retrieval might produce still larger efficiency or accuracy gains on very large repositories.

Load-bearing premise

That the chosen code completion and generation benchmarks together with BLEU scores adequately reflect real repository-level task difficulty and usefulness.

What would settle it

Evaluating the same compression methods on fresh repository tasks or with human developer judgments where higher BLEU scores do not translate into measurable improvements in task success or reduced developer effort.

Figures

Figures reproduced from arXiv: 2604.13725 by Chaozheng Wang, Cuiyun Gao, Jia Feng, Ruiqi Wang, Xiaoyuan Xie, Yingwei Ma, Zhanyue Qin.

Figure 1
Figure 1. Figure 1: Overview of the unified context compression framework (top) and the three compression paradigms (bottom) . [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Edit Similarity (%) on code completion across compression ratios for QC-3B (T2V, T2T) and QV-3B (T2I). Dashed lines [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Edit Similarity (%) on code generation across compression ratios for QC-3B (T2V, T2T) and QV-3B (T2I). Dashed lines [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: T2T (LLMLingua) compression applied to a representative Java source file. Left: original content. Middle and right: [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
read the original abstract

Repository-level code intelligence tasks require large language models (LLMs) to process long, multi-file contexts. Such inputs introduce three challenges: crucial context can be obscured by noise, truncated due to limited windows, and increased inference latency. Context compression mitigates these risks by condensing inputs. While studied in NLP, its applicability to code tasks remains largely unexplored. We present the first systematic empirical study of context compression for repository-level code intelligence, organizing eight methods into three paradigms: discrete token sequences, continuous latent vectors, and visual tokens. We evaluate them on code completion and generation, measuring performance and efficiency. Results show context compression is effective: at 4x compression, continuous latent vector methods surpass full-context performance by up to 28.3% in BLEU score, indicating they filter noise rather than just truncating. On efficiency, all paradigms reduce inference cost. Both visual and text-based compression achieve up to 50% reduction in end-to-end latency at high ratios, approaching the cost of inference without repository context. These findings establish context compression as a viable approach and provide guidance for paradigm selection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents the first systematic empirical study of context compression for repository-level code intelligence tasks. It organizes eight methods into three paradigms (discrete token sequences, continuous latent vectors, and visual tokens) and evaluates them on code completion and generation benchmarks. Key results indicate that at 4x compression, continuous latent vector methods exceed full-context performance by up to 28.3% BLEU score (interpreted as noise filtering rather than truncation) while all paradigms reduce end-to-end latency by up to 50%.

Significance. If the results hold, this work is significant as the first systematic application of context compression to repository-level code tasks, providing evidence that compression can improve performance beyond efficiency gains and offering guidance on paradigm selection. Credit is due for the clear organization into paradigms and the joint reporting of performance (BLEU) and efficiency metrics.

major comments (2)
  1. [Abstract] Abstract: The claim that the 28.3% BLEU improvement at 4x compression demonstrates noise filtering (rather than mere truncation) is load-bearing for the central effectiveness conclusion, yet it relies solely on n-gram overlap without supporting analyses such as ablation on noise levels, information retention checks, or comparisons to random truncation baselines.
  2. [Results and Evaluation sections] Results and Evaluation sections: BLEU is used as the primary metric for code completion and generation, but it does not measure functional correctness, compilation, test passage, or semantic equivalence—metrics more relevant to repository-level tasks with cross-file dependencies. This weakens the interpretation that higher BLEU equates to better task usefulness or noise removal.
minor comments (2)
  1. [Experimental Setup] The manuscript would benefit from explicit details on benchmark selection criteria, data splits, and any statistical tests used to support the reported gains.
  2. Figure and table captions could more clearly label the three paradigms and compression ratios for easier cross-reference with the text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for your thorough review of our manuscript. We appreciate the positive assessment of the work's significance and the constructive major comments. We address each point below and have made revisions to strengthen the paper where appropriate.

read point-by-point responses

  1. Referee: [Abstract] The claim that the 28.3% BLEU improvement at 4x compression demonstrates noise filtering (rather than mere truncation) is load-bearing for the central effectiveness conclusion, yet it relies solely on n-gram overlap without supporting analyses such as ablation on noise levels, information retention checks, or comparisons to random truncation baselines.

    Authors: We thank the referee for this observation. Our interpretation that the methods filter noise stems from the empirical result that compressed contexts achieve higher BLEU scores than the full context at 4x compression. This outperformance would be unlikely under simple truncation, as removing context should not improve n-gram overlap with the target. To address the request for additional support, we have added in the revised manuscript a baseline comparison with random truncation at equivalent compression rates, showing that our compression methods significantly outperform random selection. This provides evidence that the improvement is due to selective retention rather than truncation. We have also tempered the language in the abstract to present this as an indication rather than a definitive demonstration. revision: partial

  2. Referee: [Results and Evaluation sections] BLEU is used as the primary metric for code completion and generation, but it does not measure functional correctness, compilation, test passage, or semantic equivalence—metrics more relevant to repository-level tasks with cross-file dependencies. This weakens the interpretation that higher BLEU equates to better task usefulness or noise removal.

    Authors: We agree that BLEU has well-known limitations and does not directly evaluate functional correctness or semantic equivalence, which are indeed critical for repository-level code tasks. Our use of BLEU follows common practice in code generation research to enable quantitative comparison of generation quality via n-gram overlap. In the revised manuscript, we have added a dedicated paragraph in the Evaluation section discussing the choice of metrics, their limitations, and the implications for interpreting our results on noise filtering and task performance. We also highlight in the conclusion that future investigations should incorporate execution-based metrics to assess practical utility more comprehensively. revision: partial

Circularity Check

0 steps flagged

No circularity: purely empirical evaluation

full rationale

The paper is an empirical comparison of eight context compression methods across three paradigms on code completion and generation benchmarks. It reports direct measurements of BLEU scores and latency at varying compression ratios, with no derivation chain, first-principles predictions, fitted parameters renamed as predictions, or load-bearing self-citations. All claims rest on observable experimental outcomes rather than any reduction to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on standard assumptions of empirical ML evaluation (representative benchmarks, BLEU as proxy for code quality) rather than new axioms or invented entities. No free parameters are introduced in the abstract.

pith-pipeline@v0.9.0 · 5509 in / 1078 out tokens · 32172 ms · 2026-05-10T12:56:54.099629+00:00 · methodology

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

39 extracted references · 37 canonical work pages · 12 internal anchors

  1. [1]

    Aydar Bulatov, Yuri Kuratov, and Mikhail S. Burtsev. 2022. Recurrent Memory Transformer. arXiv:2207.06881 [cs.CL] https://arxiv.org/abs/2207.06881

    work page arXiv 2022

  2. [2]

    Cheng Cui, Ting Sun, Suyin Liang, Tingquan Gao, Zelun Zhang, Jiaxuan Liu, Xueqing Wang, Changda Zhou, Hongen Liu, Manhui Lin, Yue Zhang, Yubo Zhang, Handong Zheng, Jing Zhang, Jun Zhang, Yi Liu, Dianhai Yu, and Yanjun Ma. 2025. PaddleOCR-VL: Boosting Multilingual Document Parsing via a 0.9B Ultra-Compact Vision-Language Model. arXiv:2510.14528 [cs.CV] htt...

    work page arXiv 2025

  3. [3]

    DeepSeek-AI. 2025. DeepSeek-V3 Technical Report. arXiv:2412.19437 [cs.CL] https://arxiv.org/abs/2412.19437

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    Mark Chen et al. 2021. Evaluating Large Language Models Trained on Code. arXiv:2107.03374 [cs.LG] https://arxiv.org/abs/2107.03374

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  5. [5]

    Raymond Li et al. 2023. StarCoder: may the source be with you! arXiv:2305.06161 [cs.CL] https://arxiv.org/abs/2305.06161

    work page internal anchor Pith review arXiv 2023

  6. [6]

    Angela Fan, Beliz Gokkaya, Mark Harman, Mitya Lyubarskiy, Shubho Sen- gupta, Shin Yoo, and Jie M. Zhang. 2023. Large Language Models for Soft- ware Engineering: Survey and Open Problems. arXiv:2310.03533 [cs.SE] https: //arxiv.org/abs/2310.03533

    work page arXiv 2023

  7. [7]

    Jia Feng, Jiachen Liu, Cuiyun Gao, Chun Yong Chong, Chaozheng Wang, Shan Gao, and Xin Xia. 2024. ComplexCodeEval: A Benchmark for Evaluating Large Code Models on More Complex Code. InProceedings of the 39th IEEE/ACM International Conference on Automated Software Engineering (ASE ’24). ACM, 1895–1906. doi:10.1145/3691620.3695552

    work page doi:10.1145/3691620.3695552 2024

  8. [8]

    Lang Feng, Fuchao Yang, Feng Chen, Xin Cheng, Haiyang Xu, Zhenglin Wan, Ming Yan, and Bo An. 2026. AgentOCR: Reimagining Agent History via Optical Self-Compression. arXiv:2601.04786 [cs.LG] https://arxiv.org/abs/2601.04786

    work page arXiv 2026

  9. [9]

    Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, and Ming Zhou. 2020. CodeBERT: A Pre-Trained Model for Programming and Natural Languages. arXiv:2002.08155 [cs.CL] https://arxiv.org/abs/2002.08155

    work page internal anchor Pith review arXiv 2020

  10. [10]

    Tao Ge, Jing Hu, Lei Wang, Xun Wang, Si-Qing Chen, and Furu Wei. 2024. In-context Autoencoder for Context Compression in a Large Language Model. arXiv:2307.06945 [cs.CL] https://arxiv.org/abs/2307.06945

    work page arXiv 2024

  11. [11]

    Daya Guo, Shuai Lu, Nan Duan, Yanlin Wang, Ming Zhou, and Jian Yin

  12. [12]

    Unixcoder: Unified cross-modal pre-training for code representation,

    UniXcoder: Unified Cross-Modal Pre-training for Code Representation. arXiv:2203.03850 [cs.CL] https://arxiv.org/abs/2203.03850

    work page arXiv

  13. [13]

    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. GraphCodeBERT: Pre-training Code Representations with Data Flow. arXiv:2009.08366 [cs.SE] https://arxiv.org/abs/2009.08366

    work page arXiv 2021

  14. [14]

    Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y. Wu, Y. K. Li, Fuli Luo, Yingfei Xiong, and Wen- feng Liang. 2024. DeepSeek-Coder: When the Large Language Model Meets Programming – The Rise of Code Intelligence. arXiv:2401.14196 [cs.SE] https: //arxiv.org/abs/2401.14196

    work page internal anchor Pith review arXiv 2024

  15. [15]

    Hasi Hays. 2026. Attention mechanisms in neural networks. arXiv:2601.03329 [cs.LG] https://arxiv.org/abs/2601.03329

    work page arXiv 2026

  16. [16]

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, Kai Dang, Yang Fan, Yichang Zhang, An Yang, Rui Men, Fei Huang, Bo Zheng, Yibo Miao, Shanghaoran Quan, Yunlong Feng, Xingzhang Ren, Xuancheng Ren, Jingren Zhou, and Junyang Lin. 2024. Conference’17, July 2017, Washington, DC, USA Trovato e...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  17. [17]

    Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2023. LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models. arXiv:2310.05736 [cs.CL] https://arxiv.org/abs/2310.05736

    work page arXiv 2023

  18. [18]

    Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2024. LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression. arXiv:2310.06839 [cs.CL] https://arxiv.org/abs/2310.06839

    work page arXiv 2024

  19. [19]

    Juyong Jiang, Fan Wang, Jiasi Shen, Sungju Kim, and Sunghun Kim. 2026. A Survey on Large Language Models for Code Generation.ACM Transactions on Software Engineering and Methodology35, 2 (Jan. 2026), 1–72. doi:10.1145/3747588

    work page doi:10.1145/3747588 2026

  20. [20]

    Cornelius Lanczos. 1964. Evaluation of noisy data.Journal of the Society for Industrial and Applied Mathematics, Series B: Numerical Analysis1, 1 (1964), 76–85

    1964

  21. [21]

    Yucheng Li, Bo Dong, Frank Guerin, and Chenghua Lin. 2023. Compressing Context to Enhance Inference Efficiency of Large Language Models. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore, 6342–6353. doi:10.18653/v1...

    work page doi:10.18653/v1/2023.emnlp-main.391 2023

  22. [22]

    Zongqian Li, Yixuan Su, and Nigel Collier. 2024. 500xCompressor: Generalized Prompt Compression for Large Language Models. arXiv:2408.03094 [cs.CL] https://arxiv.org/abs/2408.03094

    work page arXiv 2024

  23. [23]

    Fanfan Liu and Haibo Qiu. 2025. Context Cascade Compression: Exploring the Upper Limits of Text Compression. arXiv:2511.15244 [cs.CL] https://arxiv.org/ abs/2511.15244

    work page arXiv 2025

  24. [24]

    Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

    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 (2024), 157–173. doi:10.1162/tacl_a_00638

    work page doi:10.1162/tacl_a_00638 2024

  25. [25]

    Jesse Mu, Xiang Lisa Li, and Noah Goodman. 2024. Learning to Compress Prompts with Gist Tokens. arXiv:2304.08467 [cs.CL] https://arxiv.org/abs/2304.08467

    work page arXiv 2024

  26. [26]

    OpenAI. 2024. GPT-4 Technical Report. arXiv:2303.08774 [cs.CL] https://arxiv. org/abs/2303.08774

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  27. [27]

    Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression.arXiv preprint arXiv:2403.12968, 2024

    Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Menglin Xia, Xufang Luo, Jue Zhang, Qingwei Lin, Victor Rühle, Yuqing Yang, Chin-Yew Lin, H. Vicky Zhao, Lili Qiu, and Dongmei Zhang. 2024. LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression. arXiv:2403.12968 [cs.CL] https: //arxiv.org/abs/2403.12968

    work page arXiv 2024

  28. [28]

    Fei Qi, Yingnan Hou, Ning Lin, Shanshan Bao, and Nuo Xu. 2024. A Survey of Testing Techniques Based on Large Language Models. InProceedings of the 2024 International Conference on Computer and Multimedia Technology(Sanming, China)(ICCMT ’24). Association for Computing Machinery, New York, NY, USA, 280–284. doi:10.1145/3675249.3675298

    work page doi:10.1145/3675249.3675298 2024

  29. [29]

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li,...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiao- qing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cris- tian Canton Ferrer, Aaron Grattafiori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, ...

    work page internal anchor Pith review arXiv 2024

  31. [31]

    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. arXiv:2407.07959 [cs.SE] https://arxiv.org/ abs/2407.07959

    work page arXiv 2025

  32. [32]

    Yutao Sun, Zhenyu Li, Yike Zhang, Tengyu Pan, Bowen Dong, Yuyi Guo, and Jianyong Wang. 2026. Efficient Attention Mechanisms for Large Language Models: A Survey. arXiv:2507.19595 [cs.CL] https://arxiv.org/abs/2507.19595

    work page arXiv 2026

  33. [33]

    Gemini Team. 2025. Gemini: A Family of Highly Capable Multimodal Models. arXiv:2312.11805 [cs.CL] https://arxiv.org/abs/2312.11805

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [34]

    Linformer: Self-Attention with Linear Complexity

    Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. 2020. Linformer: Self-Attention with Linear Complexity. arXiv:2006.04768 [cs.LG] https://arxiv.org/abs/2006.04768

    work page internal anchor Pith review arXiv 2020

  35. [35]

    Xiaoqiang Wang, Suyuchen Wang, Yun Zhu, and Bang Liu. 2025. R 3Mem: Bridging Memory Retention and Retrieval via Reversible Compression. arXiv:2502.15957 [cs.CL] https://arxiv.org/abs/2502.15957

    work page arXiv 2025

  36. [36]

    Yue Wang, Weishi Wang, Shafiq Joty, and Steven C. H. Hoi. 2021. CodeT5: Identifier-aware Unified Pre-trained Encoder-Decoder Models for Code Under- standing and Generation. arXiv:2109.00859 [cs.CL] https://arxiv.org/abs/2109. 00859

    work page arXiv 2021

  37. [37]

    Haoran Wei, Yaofeng Sun, and Yukun Li. 2025. DeepSeek-OCR: Contexts Optical Compression. arXiv:2510.18234 [cs.CV] https://arxiv.org/abs/2510.18234

    work page internal anchor Pith review arXiv 2025

  38. [38]

    Haoran Wei, Yaofeng Sun, and Yukun Li. 2026. DeepSeek-OCR 2: Visual Causal Flow. arXiv:2601.20552 [cs.CV] https://arxiv.org/abs/2601.20552

    work page arXiv 2026

  39. [39]

    Bissyandé, Yang Liu, and Haoye Tian

    Boyang Yang, Zijian Cai, Fengling Liu, Bach Le, Lingming Zhang, Tegawendé F. Bissyandé, Yang Liu, and Haoye Tian. 2025. A Survey of LLM-based Au- tomated Program Repair: Taxonomies, Design Paradigms, and Applications. arXiv:2506.23749 [cs.SE] https://arxiv.org/abs/2506.23749

    work page arXiv 2025