Pith. sign in

REVIEW 4 major objections 5 minor 33 references

EfficientXLang: Towards Improving Token Efficiency Through Cross-Lingual Reasoning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Prompting reasoning models in non-English languages cuts thinking tokens by 20–40% with little accuracy loss.

desk verdict A real token-savings effect in multilingual reasoning, wrapped in an overstated accuracy claim and a confounded test for 'genuine reasoning shifts.' read the letter →

arxiv 2507.00246 v1 pith:2JM73CNC submitted 2025-06-30 cs.CL

Sanchit Ahuja , Praneetha Vaddamanu , Barun Patra This is my paper
classification cs.CL
keywords multilingualreasoningtokenefficiencylanguagemodelschain-of-thoughtpromptingcross-lingualmathematicaltargetconsistencyinferencecost
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 asks whether English is the most token-efficient language for a reasoning language model's internal thinking. Across three open-weight reasoning models, four math benchmarks, and seven typologically varied languages, it reports that prompting the model to reason in a non-English language while answering in English cuts thinking-token use by 20–40% and usually keeps accuracy intact. The savings survive machine-translating the traces back to English, which the paper reads as evidence of genuinely different reasoning paths rather than tokenizer or script compactness. The benefit is conditional: it appears for models with strong target-language adherence, and is clearest on easier problems, while harder competition problems show accuracy drops. The claimed payoff is a retraining-free lever for cheaper inference in multilingual models.

What carries the argument

The load-bearing mechanism is the cross-lingual reasoning prompt together with two new metrics that make the comparison fair. The prompt fixes the final answer in English while forcing the internal chain-of-thought into a target language, isolating language from output format. Target Language Consistency (TLC) measures, via a language-identification model, what fraction of sampled traces actually obey the language instruction, and Target Language Pass@k (TLP@k) extends the standard Pass@k estimator to count only traces that are both correct and in the target language. The decisive control is translation-back: non-English traces are translated to English with GPT-4o and compared token-for-token with native English traces, so that any remaining brevity reflects the reasoning path itself. Pairwise GPT-4o judging, run in both orders to cancel position bias, converts the length difference into a quality preference.

What would settle it

Run the authors' protocol in reverse: take English reasoning traces, translate them into each non-English language with GPT-4o, and translate them back to English, then compare the round-tripped token counts with the originals. If back-translation systematically removes steps or shortens text, translation compression is present and the savings claimed in RQ2 may not reflect genuine reasoning differences. A simpler check is to count semantic units—explicit reasoning steps, checks, and restatements—in matched English and translated non-English traces; the paper's claim predicts the translated traces contain fewer such units, not merely shorter sentences.

Watch

Extended reading notes

Core claim

The paper's central claim is that the language in which a reasoning model thinks is a controllable efficiency dial: instructing the model to reason inside <think> tags in Chinese, Russian, Spanish, Arabic, Turkish, Hindi, or Korean yields shorter traces than English at comparable accuracy, with mean token reductions of roughly 20–40% for models that actually follow the language instruction. Because the token gap persists after GPT-4o translates the non-English traces into English, the paper attributes the saving not to tokenizer fertility or linguistic conciseness but to the model adopting a more direct reasoning strategy when prompted in another language. A GPT-4o judge also prefers the translated non-English traces over the original English ones for clarity, structure, and conciseness in most languages, with win rates above 90% for Arabic, Korean, and Spanish on DeepSeek R1. The authors conclude that English is not always the optimal reasoning language, and that multilingual prompting is a simple, retraining-free way to improve inference efficiency whenever the base model has strong multilingual alignment.

Load-bearing premise

The whole comparison rests on trusting GPT-4o's translations to be faithful: if the translator quietly shortens or polishes the non-English traces it converts to English, the observed token savings and judge preferences could be artifacts of translation rather than properties of the model's original reasoning.

Editorial extensions

If this is right

  • For models with high target-language consistency, switching the thinking language is a retraining-free way to cut inference tokens by roughly 20–40%.
  • Easy math benchmarks (MATH500, AMC23) tolerate the language switch with little or no accuracy loss, so low-stakes applications can capture the saving directly.
  • On hard competition problems (AIME), non-English reasoning tends to lose accuracy, so the strategy should be applied selectively rather than universally.
  • Because savings survive translation, token efficiency and reasoning quality travel together: shorter translated traces are also judged clearer and better structured.
  • The model's multilingual strength, measured by TLC, predicts who benefits; improving base multilingual alignment is a prerequisite for broader gains.

Reading between the lines

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

  • The same prompting trick could be tested on non-math reasoning such as coding, planning, or fact-checking; the paper's proposed mechanism is language-general, though only math was tested.
  • A practical extension is a difficulty-aware language router that picks the cheapest language per problem, since easy problems show savings and hard ones show accuracy risk.
  • The results suggest token savings may compound with other efficiency levers like length-aware rewards, because multilingual prompting changes the reasoning path rather than just the surface text.
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

4 major / 5 minor

Summary. The paper asks whether prompting open-weight reasoning language models to reason in non-English languages reduces inference token usage while preserving accuracy. The authors evaluate DeepSeek R1, Qwen 2.5 (QwQ-32B), and Qwen 3 on four math datasets across seven languages, introducing Target Language Consistency (TLC) and Target Language Pass@k (TLP@k) metrics. They report consistent token savings for non-English reasoning, with accuracy largely preserved for models with high TLC, and they use GPT-4o translation and judging in RQ2/RQ3 to argue that the savings reflect genuine differences in reasoning behavior rather than mere linguistic compactness. The conclusion claims a 20-40% token reduction without compromising accuracy. The code is provided. The core token-efficiency observation is a useful empirical contribution, but the stronger claim about genuine reasoning shifts rests on an untested translation assumption, and the accuracy-preservation claim is overstated relative to the paper's own tables.

Significance. If the strongest claims were fully supported, the paper would make a practical contribution by showing a zero-retraining inference-efficiency lever for multilingual RLMs, with a new evaluation methodology (TLC, TLP@k) for multilingual reasoning. The paper's main strengths are the clearly defined metrics, the breadth of languages and datasets, the machine-checkable experimental setup with released code, and the identification of TLC as a mediating factor for accuracy retention. However, the paper's significance as a claim about reasoning behavior rather than tokenizer/linguistic effects depends on RQ2/RQ3, and those experiments are currently confounded by the translator-and-judge asymmetry. The direct token-savings result (RQ1) is credible and useful even if the reasoning-shift interpretation is weakened. The paper would be acceptable for publication after the load-bearing issues are addressed, because the empirical core is reproducible and the contribution would remain valuable with appropriately narrowed claims.

major comments (4)
  1. [§3, RQ2 and RQ3] The claim that token savings persist after translation and that multilingual traces are 'better' is not yet established because GPT-4o both translates the non-English traces and judges them. In RQ2, non-English traces are translated to English with GPT-4o and compared by token count; GPT-4o's own style may compress or clean the text regardless of the source reasoning. In RQ3, the judge prompt explicitly rewards 'Efficiency & Conciseness' and 'Clarity', and the comparison is between a raw English trace and a GPT-4o-polished translated trace, so a preference for the translated trace is almost inevitable. The paper provides no control (e.g., translating the English traces with GPT-4o) and no content-preservation check. Please add such a control or explicitly restrict the conclusions to the untranslated token-savings result.
  2. [Abstract and Conclusion vs. Table 1 and §3 RQ1] The statement that multilingual prompting 'reduces token usage by 20-40% without compromising accuracy' overstates the results. Table 1 shows large TLP@4 drops for Qwen 2.5 on several languages (e.g., the reported absolute TLP@4 values drop substantially for Spanish, Arabic, Hindi, and Korean), and §3 reports a relative accuracy drop of up to 12.52% on AIME2025 even for DeepSeek R1. The accuracy-preservation claim should be restricted to models with high TLC and easier tasks, and the 20-40% token-savings figure should be replaced with the observed range or a distribution, since Table 1 shows savings from 7.7% to 90.1% depending on model and language.
  3. [Appendix A.3, Tables 5 and 6] The TLA tables for Qwen 2.5 and Qwen 3 are numerically identical across all datasets and languages. This is either a copy-paste error or a data-reporting error; it must be corrected and the numbers re-verified, because the analysis text references these tables as distinct evidence for each model.
  4. [§3 RQ2 and Figure 1] The translated-token comparison in RQ2 is under-specified: the paper does not state the translation prompt, whether the translation is sampled with temperature, or which tokenizer is used to count tokens in the translated output. Since GPT-4o is an external API model, these details are needed for reproducibility, and without them the reader cannot assess whether the translation step introduces systematic compression.
minor comments (5)
  1. [Throughout] The paper contains inconsistent model name formatting (e.g., 'DEEP SEEK R1' vs. 'DeepSeek R1' and 'QWEN' vs. 'Qwen'); please standardize.
  2. [Table 2] Table 2 is visually garbled in the manuscript; the column headers and row labels should be reformatted so that the percentage changes are legible and correctly aligned.
  3. [Appendix A.2] The code snippet for pass_at_k is truncated and does not show the complete function body; the reader cannot verify the implementation from the appendix.
  4. [Figure 1 caption] The caption 'Original Language Tokens Translated to English' is ambiguous; it should explicitly state that the two series are (a) tokens of the original non-English trace and (b) tokens of the GPT-4o English translation of that trace.
  5. [Conclusion] The conclusion's '20-40%' claim should be qualified with the actual range and the dependence on model and language, as noted in the major comment above.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical measurement study whose claims rest on external model outputs and standard estimators, not on definitional identities or load-bearing self-citations.

full rationale

The paper is an empirical measurement study with no fitted parameters, no definitional identity between inputs and outputs, and no load-bearing self-citations. The central results (RQ1 token counts and TLP@4 per language; RQ2 post-translation token counts; RQ3 GPT-4o pairwise judgments) are all external measurements of generated traces. TLC and TLP@k are explicitly defined metrics computed from generated traces and a language-identification model, and no claim is derived solely from their definitions. The TLP@k estimator is an adaptation of the Chen et al. (2021) Pass@k estimator, which is external and not authored by the present authors. The only methodological vulnerability is that RQ2 and RQ3 use GPT-4o as both translator and judge, so the observed brevity and preference could partly reflect GPT-4o's polishing behavior; that is a validity threat, not a circularity, because the paper does not fit any parameter to force this outcome and does not invoke its own prior work to justify the step. No equation reduces to an input by construction, and no self-citation carries a load-bearing premise. Accordingly the circularity score is 0.

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

The paper introduces no new physical or mathematical entities. Its central claims rest on domain assumptions about the reliability of LID classification, faithfulness of GPT-4o translations, neutrality of GPT-4o as a judge, and token count as a cost proxy. The new metrics TLC and TLP@k are operational definitions, not postulated entities, and are computable by any researcher. These assumptions are not verified with independent evidence, which moderates the strength of the conclusions.

assumptions (4)
  • domain assumption The fastText LID model (Joulin et al., 2016) correctly identifies the language of generated reasoning traces.
    TLC and TLP@k both rely on language identification; systematic LID errors would bias the reported consistency and accuracy metrics across languages.
  • domain assumption GPT-4o translation preserves the semantic content and reasoning structure of the original non-English traces.
    RQ2's conclusion that token savings persist after translation and RQ3's preference judgments both assume translations are faithful and do not systematically compress or alter reasoning.
  • domain assumption GPT-4o serves as an impartial and accurate judge of reasoning quality in the pairwise comparison.
    RQ3 uses GPT-4o with a detailed rubric to prefer non-English traces; the judge is a proprietary model with its own potential biases, and no calibration or human validation is reported.
  • domain assumption Token count in each model's tokenizer is a valid proxy for inference cost across languages.
    The central efficiency claim equates fewer tokens with lower cost; while standard, tokenizer fertility varies across languages and the paper does not measure actual latency or compute.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EfficientXLang: Towards Improving Token Efficiency Through Cross-Lingual Reasoning." pith.science (2026). https://pith.science/paper/2JM73CNC

@misc{pith2026250700246,
  author       = {Pith},
  title        = {Pith review of: EfficientXLang: Towards Improving Token Efficiency Through Cross-Lingual Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2JM73CNC}},
  note         = {Machine review of arXiv:2507.00246}
}
read the original abstract

Despite recent advances in Language Reasoning Models (LRMs), most research focuses solely on English, even though many models are pretrained on multilingual data. In this work, we investigate: Is English the most token-efficient language for reasoning? We evaluate three open-source RLMs: DeepSeek R1, Qwen 2.5 and Qwen 3, across four math datasets and seven typologically diverse languages. We find that reasoning in non-English languages not only reduces token usage, but also preserves accuracy. These gains persist even after translating the reasoning traces into English, suggesting genuine shifts in reasoning behavior rather than surface-level linguistic effects. The extent of improvement, however, depends on the models multilingual strength. Our findings motivate a broader view of reasoning in language models, highlighting the potential of multilingual reasoning and the importance of strong multilingual foundations. The code for our work can be found: https://github.com/microsoft/EfficientXLang.

Figures

Figures reproduced from arXiv: 2507.00246 by the authors.

Figure 1
Figure 1. Token Count Ratio vs. number of problems with at least one correct answer in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. DEEPSEEK R1 analysis: TLA versus mean thinking tokens per language, with point size indicating how often the target language was used (out of 16 samples). 2500 5000 7500 Mean Thinking Tokens 0.50 0.75 TLA R: 0.79 AIME2024 5000 10000 Mean Thinking Tokens 0.50 0.75 R: 0.56 AIME2025 2000 4000 6000 Mean Thinking Tokens 0.70 0.80 0.90 R: 0.47 AMC23 1000 2000 3000 Mean Thinking Tokens 0.70 0.80 R: 0.33 MATH500 QWEN 3 ar z… view at source ↗
Figure 3
Figure 3. QWEN 3 analysis 11 [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: QWEN 2.5 analysis 12 [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Example problem solved by QWEN 3 13 [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 10 canonical work pages

  1. [1]

    L1: Controlling how long a reasoning model thinks with reinforcement learning

    Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning. arXiv preprint arXiv:2503.04697, 2025

    arXiv 2025

  2. [2]

    2023 amc 12a

    Art of Problem Solving . 2023 amc 12a. https://artofproblemsolving.com/wiki/index.php/2023_AMC_12A, 2023 a . Accessed: 2025-05-04

    work page 2023

  3. [3]

    2023 amc 12b, 2023 b

    Art of Problem Solving . 2023 amc 12b, 2023 b . URL https://artofproblemsolving.com/wiki/index.php/2023_AMC_12B. Accessed: 2025-05-04

    work page 2023

  4. [4]

    2025 aime i, 2025 a

    Art of Problem Solving . 2025 aime i, 2025 a . URL https://artofproblemsolving.com/wiki/index.php/2025_AIME_I. Accessed: 2025-05-04

    work page 2025

  5. [5]

    2025 aime ii, 2025 b

    Art of Problem Solving . 2025 aime ii, 2025 b . URL https://artofproblemsolving.com/wiki/index.php/2025_AIME_II. Accessed: 2025-05-04

    work page 2025

  6. [6]

    Multilingual test-time scaling via initial thought transfer, 2025

    Prasoon Bajpai and Tanmoy Chakraborty. Multilingual test-time scaling via initial thought transfer, 2025. URL https://arxiv.org/abs/2505.15508

    arXiv 2025

  7. [7]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian...

    arXiv 2021

  8. [8]

    Do not think that much for 2+3=? on the overthinking of o1-like llms, 2025

    Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, and Dong Yu. Do not think that much for 2+3=? on the overthinking of o1-like llms, 2025. URL https://arxiv.org/abs/2412.21187

    arXiv 2025

Show all 33 references
  1. [10]

    The danger of overthinking: Examining the reasoning-action dilemma in agentic tasks

    Alejandro Cuadron, Dacheng Li, Wenjie Ma, Xingyao Wang, Yichuan Wang, Siyuan Zhuang, Shu Liu, Luis Gaspar Schroeder, Tian Xia, Huanzhi Mao, et al. The danger of overthinking: Examining the reasoning-action dilemma in agentic tasks. arXiv preprint arXiv:2502.08235, 2025 b

    2025 arXiv

  2. [11]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025

    DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025. URL https://arxiv.org/abs/2501.12948

    2025 arXiv

  3. [12]

    Could thinking multilingually empower llm reasoning?, 2025

    Changjiang Gao, Xu Huang, Wenhao Zhu, Shujian Huang, Lei Li, and Fei Yuan. Could thinking multilingually empower llm reasoning?, 2025. URL https://arxiv.org/abs/2504.11833

    2025 arXiv

  4. [13]

    Chinese and english mathematics language: The relation between linguistic clarity and mathematics performance

    Yi et.al Han. Chinese and english mathematics language: The relation between linguistic clarity and mathematics performance. Mathematical Thinking and Learning, 3 0 (2-3): 0 201--220, 2001

    2001

  5. [14]

    Dnr bench: Benchmarking over-reasoning in reasoning llms, 2025

    Masoud Hashemi, Oluwanifemi Bamgbose, Sathwik Tejaswi Madhusudhan, Jishnu Sethumadhavan Nair, Aman Tiwari, and Vikas Yadav. Dnr bench: Benchmarking over-reasoning in reasoning llms, 2025. URL https://arxiv.org/abs/2503.15793

    2025 arXiv

  6. [15]

    Bag of tricks for efficient text classification

    Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016

    2016 arXiv

  7. [16]

    Understand, solve and translate: Bridging the multilingual mathematical reasoning gap, 2025

    Hyunwoo Ko, Guijin Son, and Dasol Choi. Understand, solve and translate: Bridging the multilingual mathematical reasoning gap, 2025. URL https://arxiv.org/abs/2501.02448

    2025 arXiv

  8. [17]

    Let's verify step by step

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. arXiv preprint arXiv:2305.20050, 2023

    2023 arXiv

  9. [18]

    Reasoning models can be effective without thinking, 2025

    Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. Reasoning models can be effective without thinking, 2025. URL https://arxiv.org/abs/2504.09858

    2025 arXiv

  10. [19]

    Deepseek-r1 thoughtology: Let's think about llm reasoning, 2025

    Sara Vera Marjanović, Arkil Patel, Vaibhav Adlakha, Milad Aghajohari, Parishad BehnamGhader, Mehar Bhatia, Aditi Khandelwal, Austin Kraft, Benno Krojer, Xing Han Lù, Nicholas Meade, Dongchan Shin, Amirhossein Kazemnejad, Gaurav Kamath, Marius Mosbach, Karolina Stańczak, and Si...

    2025

  11. [20]

    OpenAI, :, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander Mądry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, and Alex Kirillov et. al. Gpt-4o sys...

    2024 arXiv

  12. [21]

    Openai o1 system card, 2024 b

    OpenAI, :, Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, Alex Iftimie, Alex Karpenko, Alex Tachard Passos, Alexander Neitz, Alexander Prokofiev, Alexander Wei, Allison Tam, and Ally Be...

    2024 arXiv

  13. [22]

    Cross-lingual prompting: Improving zero-shot chain-of-thought reasoning across languages

    Libo Qin, Qiguang Chen, Fuxuan Wei, Shijue Huang, and Wanxiang Che. Cross-lingual prompting: Improving zero-shot chain-of-thought reasoning across languages. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural...

    2023 doi

  14. [23]

    Language models are multilingual chain-of-thought reasoners, 2022

    Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush Vosoughi, Hyung Won Chung, Yi Tay, Sebastian Ruder, Denny Zhou, Dipanjan Das, and Jason Wei. Language models are multilingual chain-of-thought reasoners, 2022. URL https://arxiv.org/abs/2210.03057

    2022 arXiv

  15. [25]

    Stop overthinking: A survey on efficient reasoning for large language models

    Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Hanjie Chen, et al. Stop overthinking: A survey on efficient reasoning for large language models. arXiv preprint arXiv:2503.16419, 2025 b

    2025 arXiv

  16. [26]

    Qwen3, April 2025 a

    Qwen Team. Qwen3, April 2025 a . URL https://qwenlm.github.io/blog/qwen3/

    2025

  17. [27]

    Qwq-32b: Embracing the power of reinforcement learning, March 2025 b

    Qwen Team. Qwq-32b: Embracing the power of reinforcement learning, March 2025 b . URL https://qwenlm.github.io/blog/qwq-32b/

    2025

  18. [28]

    Aime problem set 1983-2024, 2023

    Hemish Veeraboina. Aime problem set 1983-2024, 2023. URL https://www.kaggle.com/datasets/hemishveeraboina/aime-problem-set-1983-2024

    1983

  19. [29]

    Do llamas work in english? on the latent language of multilingual transformers

    Chris Wendler, Veniamin Veselovsky, Giovanni Monea, and Robert West. Do llamas work in english? on the latent language of multilingual transformers. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 15366--...

    2024

  20. [30]

    Qwen2.5 technical report

    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, Me...

    2024 arXiv

  21. [31]

    Gonzalez, Clark Barrett, and Ying Sheng

    Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E. Gonzalez, Clark Barrett, and Ying Sheng. Sglang: Efficient execution of structured language model programs. In A. Globerson, L. Mackey, D. Bel...

    2024

  22. [32]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  23. [33]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  24. [34]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  25. [35]

    better\_response\_label

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.