Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

OpenCodeReasoning-II: A Simple Test Time Scaling Approach via Self-Critique

T0 review · 3 major / 4 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 A two-stage fine-tuning recipe on 2.5M question–solution–critique triples lets a coding model select its best answer from ten tries and beat larger rivals.

desk verdict A genuinely useful dataset and benchmark extension, with a test-time scaling method that is plausible but needs a stronger contamination screen before the headline gains are taken at face value. read the letter →

arxiv 2507.09075 v1 pith:CJTROUTQ submitted 2025-07-11 cs.CL

Wasi Uddin Ahmad , Somshubra Majumdar , Aleksander Ficek , Sean Narenthiran , Mehrzad Samadi , Jocelyn Huang , Siddhartha Jain , Vahid Noroozi
show 1 more author
Boris Ginsburg
This is my paper
classification cs.CL
keywords test-timescalingself-critiquecodegenerationsyntheticreasoningdatadistillationcompetitiveprogrammingLiveBenchsupervisedfine-tuning
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 claims that a simple supervised fine-tuning recipe can give coding models a genuine test-time scaling ability without reinforcement learning. Its central contribution is OpenCodeReasoning-II, a dataset of roughly 2.5 million question–solution–critique triples over about 35,000 competitive programming problems, where both solutions and critiques are written as chain-of-thought reasoning traces ending in a binary right/wrong judgment. The authors fine-tune an instruction-tuned model family in two stages—first for code generation, then jointly for generation and self-critique—and show that generating ten candidate solutions and letting the model's own critique pick the best one lifts pass@1 by up to about six points. On their LiveCodeBench evaluations, the 32B model reaches 62.1 pass@1 in Python and 58.1 in C++, rising to 67.4 and 60.6 with self-critique selection. They also contribute a C++ extension of LiveCodeBench with 279 problems.

What carries the argument

The load-bearing object is the question–solution–critique triple: for each of about 35,000 unique competitive programming problems, the dataset records multiple sampled solutions in Python and C++, each paired with a chain-of-thought critique and a binary right/wrong judgment. Two mechanisms carry the argument: the two-stage fine-tuning schedule, which first teaches generation and then jointly teaches generation plus critique so that the same weights perform both roles at inference; and the test-time selection heuristic, which among all solutions a model labels 'right' picks the one with the shortest critique reasoning trace. The dataset's execution pass rates for a 60% subsample are an auxiliary resource for future work such as offline reinforcement learning.

What would settle it

Run the OCR-2-32B model on a fresh set of competitive programming problems published after the training data was collected, applying the same ten-sample self-critique selection; if the pass@1 gain over single-shot generation largely disappears on those future problems, the benchmark gains reflect contamination rather than a general test-time scaling ability.

Watch

Extended reading notes

Core claim

The central claim is that a model can learn to verify its own programming solutions well enough that sampling several candidates and selecting by self-critique produces large accuracy gains, without any reinforcement learning. The evidence is a two-stage fine-tuning procedure on a 2.5M-triple dataset: stage one trains pure code generation, stage two trains generation and critique jointly, so the same model can produce and judge its own outputs at inference. With ten samples and a selection heuristic that picks the 'right'-labeled solution with the shortest critique trace, the 32B model's LiveCodeBench-Python pass@1 rises from 61.3 to 67.4, and its C++ pass@1 from 59.8 to 60.6. The authors attribute the remaining gap to critique accuracy, which they measure at about 47% on medium and under 14% on hard LiveCodeBench problems, and they identify better selection as the main lever for future gains. The paper also introduces a C++ extension of LiveCodeBench covering 279 problems, on which models trained on the dataset dramatically outperform open-weight rivals.

Load-bearing premise

The paper's results depend on the assumption that its automatic contamination check—cosine similarity with a 0.7 cutoff judged by one LLM—removed every meaningful overlap between the 34,125 training questions and the LiveCodeBench evaluation set.

Editorial extensions

If this is right

  • If the central claim holds, self-critique selection is a cheap RL-free route to test-time scaling: a single fine-tuned model can generate and verify its own candidates, narrowing the gap between pass@1 and pass@10.
  • Smaller models benefit most from scaling synthetic reasoning data: the 7B model shows substantial gains as the training set grows from 25K to 1.4M samples, while the 14B and 32B models plateau.
  • Joint training on comparable volumes of Python and C++ data yields strong performance in both languages, with C++-only training transferring to Python more effectively than Python-only training transfers to C++.
  • Critique accuracy is the binding constraint on selection-based test-time scaling, so improving the judge (not the generator) is the most direct route to further gains.
  • The C++ extension of LiveCodeBench exposes large performance gaps for models not trained on C++, making it a more discriminating evaluation surface than the Python split.

Reading between the lines

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

  • The shortest-critique-trace heuristic is an oddly brittle selection rule; a natural extension is to make several independent critiques per solution and select by agreement, which the paper's data could support without new generation.
  • Because the contamination check only compares problem statements, it would not catch overlap via hidden test cases or editorial solutions; a stronger audit would execute generated solutions on the evaluation harness's own tests.
  • The cross-language transfer asymmetry may be driven by tokenization or pretraining exposure rather than by problem content; controlling for per-token dataset size would clarify whether the effect is about language or data volume.
  • If critique accuracy is the bottleneck, scaling critique computation per solution (e.g., self-consistency over judgments) may be a more compute-effective way to improve selection than sampling more solutions.
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

3 major / 4 minor

Summary. This paper introduces OpenCodeReasoning-II, a dataset of roughly 2.5M question-solution-critique triples derived from about 35K competitive programming problems, and presents a two-stage supervised fine-tuning recipe applied to Qwen2.5-Instruct models of 7B, 14B, and 32B. The first stage trains code generation; the second jointly trains generation and self-critique. At inference, the model samples k solutions and selects the one with the shortest critique trace among solutions judged 'right'. The paper reports that OCR-2-32B reaches 62.1 pass@1 on LiveCodeBench-Python and 67.4 pass@1|select@10, with a +6.1 gain from self-critique, and similarly 58.1/60.6 on a new LiveCodeBench-C++ extension. The authors also release the dataset and the C++ benchmark extension.

Significance. If the empirical claims hold, the paper provides a valuable public resource (the largest code reasoning dataset to date) and a simple, reproducible recipe for test-time scaling without reinforcement learning. Notably, the paper discloses the limitation that self-critique accuracy is low on medium and hard problems (47% and below 14%, respectively), which bounds the achievable gains. The strongest evidence, however, rests on LiveCodeBench results, and the adequacy of the contamination screening is not fully demonstrated; the absence of error bars further tempers the claim of 'significant improvements.' Because the dataset and benchmark are released, the work is likely to be useful to the community regardless of these caveats.

major comments (3)
  1. [Section 2.1.1] The decontamination step (cosine similarity cutoff 0.7 with Llama-3.3-70B-Instruct judge, removing 674 questions) may not detect rephrased or near-duplicate problems that share source platforms with LiveCodeBench's recent window (2408-2502). Since the training sources include CodeForces, AtCoder, and LeetCode and the evaluation uses the same platforms, any surviving overlap could inflate both pass@1 and the self-critique gains reported in Tables 2 and 3. Please report per-benchmark similarity distributions, a manual audit of pairs near the 0.7 threshold, or a date-filtered/contamination-hardened evaluation, and show the main results on that subset.
  2. [Section 4, Benchmarks and Metrics] The sentence 'This benchmark contains 67 easy, 89 medium, and 279 hard coding questions' is numerically impossible for a 279-problem split; the hard count should be 123 (67+89+123=279). This typo creates ambiguity in all per-difficulty numbers and in Section 5.1's statement that 212 problems fall under the Medium/Hard category. Please correct the typo and re-verify all derived statistics.
  3. [Tables 2-3] The paper reports averages over 10 runs (Table 3) or 64/16 runs (Table 2) without variance or significance testing. The self-critique gains range from +0.8 to +6.1 and cross-model comparisons are close (e.g., Python Pass@1|Select@10 is 67.4 for both Qwen3-32B and OCR-2-32B). Add standard errors or paired bootstrap confidence intervals for the headline comparisons, or explicitly state that observed differences are within noise; otherwise the claim of 'significant improvements' is not statistically supported.
minor comments (4)
  1. [Section 2.1.3] There is a duplicated word in 'We prompted QwQ-32B to to generate critiques'; please fix.
  2. [Section 4, Benchmarks and Metrics] The word 'mentioend' should be 'mentioned' in the definition of critique accuracy.
  3. [Appendix A, Figure 5 caption] The caption says 'using OCR-2-32B on LiveCodeBench-Python' but the figure displays several models; please clarify that OCR-2-32B is the model used for the selection comparison or update the caption to cover all plotted models.
  4. [Table 1] The total sample count in Table 1 (1,398,166 + 1,174,475 = 2,572,641) is described as approximately 2.5M in the abstract; consider stating the exact total in the dataset description for precision.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central test-time-scaling claim is evaluated on the external LiveCodeBench benchmark, and no fitted parameter is renamed as a prediction.

full rationale

The paper's load-bearing results (Tables 2 and 3) are pass@1 and pass@1|select@10 scores on LiveCodeBench-Python and an extended LiveCodeBench-C++. These numbers are produced by fine-tuning Qwen2.5-Instruct on a separately constructed dataset and then evaluating on externally maintained benchmark problems; the target metric is not defined in terms of the training labels. The self-critique selector is an explicit heuristic (shortest 'right' critique trace), and Appendix A compares it against random selection, so the reported improvement is an empirical finding rather than a definitional one. The contamination assessment in Section 2.1.1 follows an external protocol (Yang et al., 2023) with a stated cosine-similarity threshold; whether that threshold is strong enough to ensure cleanliness is a validity concern, not circularity. Self-citations to prior OCR work and related NVIDIA papers are used for dataset provenance and context, but they are not the load-bearing justification for the main results, which stand on external benchmark comparisons and ablation tables. No equation or metric reduces by construction to the training data, and no fitted parameter is relabeled as a prediction.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claims depend on the synthetic data pipeline, the decontamination check, and the selection heuristic. These are empirical assumptions rather than derived facts. No new physical entities, dimensions, or conserved quantities are proposed.

free parameters (4)
  • k (number of samples at test time) = 10 (up to 100 in Fig. 3)
    Chosen by hand for main evaluation; larger k improves pass@k but selection saturates (Fig. 3).
  • Selection heuristic: shortest critique trace
    Heuristic chosen based on prior underthinking work; compared against random selection in Appendix A, but no principled optimization.
  • Temperature for solution generation and inference = 0.6
    Ablation showed minimal effect between 0.2 and 0.7, so 0.6 was used; a hand choice, not fitted to the test set.
  • top-p = 0.95
    Standard nucleus sampling choice used for data generation and inference.
assumptions (3)
  • domain assumption The contamination assessment protocol (cosine similarity cutoff 0.7 plus Llama-3.3-70B-Instruct judge) identifies every meaningful overlap between the 34,125 training questions and the evaluation benchmarks, including LiveCodeBench.
    Invoked in Section 2.1.1 to justify removing only 674 questions; if near-duplicates remain, benchmark scores could be inflated.
  • domain assumption DeepSeek-R1 and QwQ-32B generate solutions and critiques of sufficient quality for distillation, without human verification on the 40% of data lacking execution results.
    The dataset is fully synthetic; critique labels are filtered only for binary format, and the authors report critique accuracy around 47% on medium problems, so label noise is a known risk.
  • domain assumption The 'shortest critique trace' heuristic selects the correct solution among the right-labeled candidates.
    Motivated by Wang et al. (2025a) on underthinking; the paper compares it to random selection in Appendix A but does not prove its general validity. Its effectiveness varies by model (e.g., small gain for R1 in Figure 5).

how reviews work

0 comments
Cite this review

Pith. "Pith review of OpenCodeReasoning-II: A Simple Test Time Scaling Approach via Self-Critique." pith.science (2026). https://pith.science/paper/CJTROUTQ

@misc{pith2026250709075,
  author       = {Pith},
  title        = {Pith review of: OpenCodeReasoning-II: A Simple Test Time Scaling Approach via Self-Critique},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CJTROUTQ}},
  note         = {Machine review of arXiv:2507.09075}
}
read the original abstract

Recent advancements in reasoning-based Large Language Models (LLMs), particularly their potential through test-time scaling, have created significant opportunities for distillation in code generation and critique. However, progress in both areas fundamentally depends on large-scale, high-quality datasets. In this work, we introduce OpenCodeReasoning-II, a dataset consists of 2.5M question-solution-critique triples (approx. 35K unique programming questions), making it nearly twice the size of the previous largest publicly available code reasoning dataset. In this work, we employ a two-stage supervised fine-tuning strategy. The first stage focuses on fine-tuning for code generation, while the second stage involves the joint training of models for both code generation and critique. Our resulting finetuned Qwen2.5-Instruct models achieve performance in code generation that either exceeds or equals the best prior open-weight distilled models. Notably, the integration of our code generation and critique models leads to significant improvements in competitive coding performance. Furthermore, we present an extension of the LiveCodeBench benchmark to specifically support the C++ programming language, thereby facilitating more comprehensive LLM evaluation using this benchmark.

Figures

Figures reproduced from arXiv: 2507.09075 by the authors.

Figure 1
Figure 1. Demonstrating performance gains on LiveCodeBench, achieved through test-time scaling [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the OPENCODEREASONING-II development stages. 2 Development of OPENCODEREASONING-II and LiveCodeBench-C++ 2.1 Construction of OPENCODEREASONING-II The construction of the OPENCODEREASONING-II dataset involved a four-stage approach which is demonstrated in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Performance gap between pass@1, pass@1|select@k, and pass@k under test-time scaling - large number of samples drawn from OCR-2-32B. improvements. We leave the exploration of more sophisticated heuristics and methods to enhance the accuracy of self-critique-based selection for future research. 5.2 Impact of Temperature on Self Critique We tested how the critic LLMs responded to different decoding temperatures by re-e… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Impact of scaling up data from 25k to 1.4M samples in O [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Differences in pass@1 scores between randomly selecting the final output vs. choosing [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Prompt template used for solution generation using R1 for O [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Prompt template used for critique data generation for O [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Length Value Model: Scalable Value Pretraining for Token-Level Length Modeling

    cs.CL 2026-04 unverdicted novelty 6.0 of 10

    LenVM trains a token-level value head to predict discounted remaining length, enabling length control and efficiency steering on LLMs and VLMs.

Reference graph

Works this paper leans on

66 extracted references · 21 canonical work pages · cited by 1 Pith paper

  1. [1]

    , author Ficek, A

    author Ahmad, W.U. , author Ficek, A. , author Samadi, M. , author Huang, J. , author Noroozi, V. , author Majumdar, S. , author Ginsburg, B. , year 2025 a. title Opencodeinstruct: A large-scale instruction tuning dataset for code llms . https://arxiv.org/abs/2504.04030, http://arxiv.org/abs/2504.04030 arXiv:2504.04030

    arXiv 2025

  2. [2]

    , author Narenthiran, S

    author Ahmad, W.U. , author Narenthiran, S. , author Majumdar, S. , author Ficek, A. , author Jain, S. , author Huang, J. , author Noroozi, V. , author Ginsburg, B. , year 2025 b. title Opencodereasoning: Advancing data distillation for competitive coding . https://arxiv.org/abs/2504.01943, http://arxiv.org/abs/2504.01943 arXiv:2504.01943

    arXiv 2025

  3. [3]

    , author Odena, A

    author Austin, J. , author Odena, A. , author Nye, M. , author Bosma, M. , author Michalewski, H. , author Dohan, D. , author Jiang, E. , author Cai, C. , author Terry, M. , author Le, Q. , author Sutton, C. , year 2021 . title Program synthesis with large language models . https://arxiv.org/abs/2108.07732, http://arxiv.org/abs/2108.07732 arXiv:2108.07732

    arXiv 2021

  4. [4]

    , author Levy, I

    author Bercovich, A. , author Levy, I. , author Golan, I. , author Dabbah, M. , author El-Yaniv, R. , author Puny, O. , author Galil, I. , author Moshe, Z. , author Ronen, T. , author Nabwani, N. , author Shahaf, I. , author Tropp, O. , author Karpas, E. , author Zilberstein, R. , author Zeng, J. , author Singhal, S. , author Bukharin, A. , author Zhang, ...

    arXiv 2025

  5. [5]

    title Bespoke-stratos: The unreasonable effectiveness of reasoning distillation

    author BespokeLabs , year 2025 . title Bespoke-stratos: The unreasonable effectiveness of reasoning distillation . howpublished www.bespokelabs.ai/blog/bespoke-stratos-the-unreasonable-effectiveness-of-reasoning-distillation . note Accessed: 2025-01-22

    work page 2025

  6. [6]

    , author Juravsky, J

    author Brown, B. , author Juravsky, J. , author Ehrlich, R. , author Clark, R. , author Le, Q.V. , author Ré, C. , author Mirhoseini, A. , year 2024 . title Large language monkeys: Scaling inference compute with repeated sampling . https://arxiv.org/abs/2407.21787, http://arxiv.org/abs/2407.21787 arXiv:2407.21787

    arXiv 2024

  7. [7]

    , author Davis, J.Q

    author Chen, L. , author Davis, J.Q. , author Hanin, B. , author Bailis, P. , author Stoica, I. , author Zaharia, M. , author Zou, J. , year 2024 . title Are more llm calls all you need? towards scaling laws of compound inference systems . https://arxiv.org/abs/2403.02419, http://arxiv.org/abs/2403.02419 arXiv:2403.02419

    arXiv 2024

  8. [8]

    , author Tworek, J

    author Chen, M. , author Tworek, J. , author Jun, H. , author Yuan, Q. , author Pinto, H.P.D.O. , author Kaplan, J. , author Edwards, H. , author Burda, Y. , author Joseph, N. , author Brockman, G. , et al., year 2021 . title Evaluating large language models trained on code . journal arXiv preprint arXiv:2107.03374

    arXiv 2021

Show all 66 references
  1. [9]

    , author Li, G

    author Chen, X. , author Li, G. , author Wang, Z. , author Jin, B. , author Qian, C. , author Wang, Y. , author Wang, H. , author Zhang, Y. , author Zhang, D. , author Zhang, T. , author Tong, H. , author Ji, H. , year 2025 . title Rm-r1: Reward modeling as reasoning . https:/...

    2025

  2. [10]

    , author Yang, D

    author DeepSeek-AI , author Guo, D. , author Yang, D. , author Zhang, H. , author Song, J. , author Zhang, R. , author Xu, R. , author Zhu, Q. , author Ma, S. , author Wang, P. , author Bi, X. , author Zhang, X. , author Yu, X. , author Wu, Y. , author Wu, Z.F. , author Gou, Z...

    2025 arXiv

  3. [11]

    , author Majumdar, S

    author Ficek, A. , author Majumdar, S. , author Noroozi, V. , author Ginsburg, B. , year 2025 . title Scoring verifiers: Evaluating synthetic verification in code and reasoning . https://arxiv.org/abs/2502.13820, http://arxiv.org/abs/2502.13820 arXiv:2502.13820

    2025 arXiv

  4. [12]

    , author Dubey, A

    author Grattafiori, A. , author Dubey, A. , author Jauhri, A. , author Pandey, A. , author Kadian, A. , author Al-Dahle, A. , author Letman, A. , author Mathur, A. , author Schelten, A. , author Vaughan, A. , et al., year 2024 . title The llama 3 herd of models . journal arXiv...

    2024 arXiv

  5. [13]

    , author Basart, S

    author Hendrycks, D. , author Basart, S. , author Kadavath, S. , author Mazeika, M. , author Arora, A. , author Guo, E. , author Burns, C. , author Puranik, S. , author He, H. , author Song, D. , author Steinhardt, J. , year 2021 . title Measuring coding challenge competence w...

    2021

  6. [14]

    , author Buys, J

    author Holtzman, A. , author Buys, J. , author Du, L. , author Forbes, M. , author Choi, Y. , year 2020 . title The curious case of neural text degeneration , in: booktitle International Conference on Learning Representations . https://openreview.net/forum?id=rygGQyrFvH

    2020

  7. [15]

    , author Han, K

    author Jain, N. , author Han, K. , author Gu, A. , author Li, W.D. , author Yan, F. , author Zhang, T. , author Wang, S. , author Solar-Lezama, A. , author Sen, K. , author Stoica, I. , year 2025 . title Livecodebench: Holistic and contamination free evaluation of large langua...

    2025

  8. [16]

    , author Ba, J

    author Kingma, D.P. , author Ba, J. , year 2015 . title Adam: A method for stochastic optimization , in: editor Bengio, Y. , editor LeCun, Y. (Eds.), booktitle 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Tr...

    2015 arXiv

  9. [17]

    , author Li, Z

    author Kwon, W. , author Li, Z. , author Zhuang, S. , author Sheng, Y. , author Zheng, L. , author Yu, C.H. , author Gonzalez, J.E. , author Zhang, H. , author Stoica, I. , year 2023 . title Efficient memory management for large language model serving with pagedattention , in:...

    2023

  10. [18]

    , author Cao, S

    author Li, D. , author Cao, S. , author Griggs, T. , author Liu, S. , author Mo, X. , author Tang, E. , author Hegde, S. , author Hakhamaneshi, K. , author Patil, S.G. , author Zaharia, M. , author Gonzalez, J.E. , author Stoica, I. , year 2025 a. title Llms can easily learn t...

    2025 arXiv

  11. [19]

    , author Fu, J

    author Li, R. , author Fu, J. , author Zhang, B.W. , author Huang, T. , author Sun, Z. , author Lyu, C. , author Liu, G. , author Jin, Z. , author Li, G. , year 2023 . title Taco: Topics in algorithmic code generation dataset . journal arXiv preprint arXiv:2312.14852

    2023 arXiv

  12. [20]

    , author Choi, D

    author Li, Y. , author Choi, D. , author Chung, J. , author Kushman, N. , author Schrittwieser, J. , author Leblond, R. , author Eccles, T. , author Keeling, J. , author Gimeno, F. , author Dal Lago, A. , author Hubert, T. , author Choy, P. , author de Masson d'Autume, C. , au...

    2022 arXiv

  13. [21]

    , author Zhang, D

    author Li, Z.Z. , author Zhang, D. , author Zhang, M.L. , author Zhang, J. , author Liu, Z. , author Yao, Y. , author Xu, H. , author Zheng, J. , author Wang, P.J. , author Chen, X. , author Zhang, Y. , author Yin, F. , author Dong, J. , author Li, Z. , author Bi, B.L. , autho...

    2025 arXiv

  14. [22]

    , author Zeng, L

    author Liu, C.Y. , author Zeng, L. , author Liu, J. , author Yan, R. , author He, J. , author Wang, C. , author Yan, S. , author Liu, Y. , author Zhou, Y. , year 2024 . title Skywork-reward: Bag of tricks for reward modeling in llms . journal arXiv preprint arXiv:2410.18451

    2024 arXiv

  15. [23]

    , author Chen, Y

    author Liu, Z. , author Chen, Y. , author Shoeybi, M. , author Catanzaro, B. , author Ping, W. , year 2025 a. title Acemath: Advancing frontier math reasoning with post-training and reward modeling . https://arxiv.org/abs/2412.15084, http://arxiv.org/abs/2412.15084 arXiv:2412.15084

    2025 arXiv

  16. [24]

    , author Wang, P

    author Liu, Z. , author Wang, P. , author Xu, R. , author Ma, S. , author Ruan, C. , author Li, P. , author Liu, Y. , author Wu, Y. , year 2025 b. title Inference-time scaling for generalist reward modeling . https://arxiv.org/abs/2504.02495, http://arxiv.org/abs/2504.02495 ar...

    2025

  17. [25]

    , author Tan, S

    author Luo, M. , author Tan, S. , author Huang, R. , author Patel, A. , author Ariyak, A. , author Wu, Q. , author Shi, X. , author Xin, R. , author Cai, C. , author Weber, M. , author Zhang, C. , author Li, L.E. , author Popa, R.A. , author Stoica, I. , year 2025 . title Deep...

    2025

  18. [26]

    , author Xu, C

    author Luo, Z. , author Xu, C. , author Zhao, P. , author Sun, Q. , author Geng, X. , author Hu, W. , author Tao, C. , author Ma, J. , author Lin, Q. , author Jiang, D. , year 2024 . title Wizardcoder: Empowering code large language models with evol-instruct , in: booktitle Th...

    2024

  19. [27]

    , author Phung, D.V

    author Mahan, D. , author Phung, D.V. , author Rafailov, R. , author Blagden, C. , author Lile, N. , author Castricato, L. , author Fränken, J.P. , author Finn, C. , author Albalak, A. , year 2024 . title Generative reward models . https://arxiv.org/abs/2410.12832, http://arxi...

    2024 arXiv

  20. [28]

    , author Noroozi, V

    author Majumdar, S. , author Noroozi, V. , author Narenthiran, S. , author Ficek, A. , author Balam, J. , author Ginsburg, B. , year 2024 . title Genetic instruct: Scaling up synthetic generation of coding instructions for large language models . journal arXiv preprint arXiv:2...

    2024 arXiv

  21. [29]

    , author Hanley, D

    author Moshkov, I. , author Hanley, D. , author Sorokin, I. , author Toshniwal, S. , author Henkel, C. , author Schifferer, B. , author Du, W. , author Gitman, I. , year 2025 . title Aimo-2 winning solution: Building state-of-the-art mathematical reasoning models with openmath...

    2025 arXiv

  22. [30]

    , author Agarwal, N

    author Nvidia , author : , author Adler, B. , author Agarwal, N. , author Aithal, A. , author Anh, D.H. , author Bhattacharya, P. , author Brundyn, A. , author Casper, J. , author Catanzaro, B. , author Clay, S. , author Cohen, J. , author Das, S. , author Dattagupta, A. , aut...

    2024 arXiv

  23. [31]

    title Open Thoughts

    author OpenThoughts , year 2025 . title Open Thoughts . howpublished https://open-thoughts.ai

    2025

  24. [32]

    , author Lozhkov, A

    author Penedo, G. , author Lozhkov, A. , author Kydlíček, H. , author Allal, L.B. , author Beeching, E. , author Lajarín, A.P. , author Gallouédec, Q. , author Habib, N. , author Tunstall, L. , author von Werra, L. , year 2025 a. title Codeforces . howpublished https://hugging...

    2025

  25. [33]

    , author Lozhkov, A

    author Penedo, G. , author Lozhkov, A. , author Kydlíček, H. , author Allal, L.B. , author Beeching, E. , author Lajarín, A.P. , author Gallouédec, Q. , author Habib, N. , author Tunstall, L. , author von Werra, L. , year 2025 b. title Codeforces cots . howpublished https://hu...

    2025

  26. [34]

    , author Yang, M.Y.R

    author Qu, Y. , author Yang, M.Y.R. , author Setlur, A. , author Tunstall, L. , author Beeching, E.E. , author Salakhutdinov, R. , author Kumar, A. , year 2025 . title Optimizing test-time compute via meta reinforcement fine-tuning . https://arxiv.org/abs/2503.07572, http://ar...

    2025 arXiv

  27. [35]

    , author Rajaraman, N

    author Setlur, A. , author Rajaraman, N. , author Levine, S. , author Kumar, A. , year 2025 . title Scaling test-time compute without verification or rl is suboptimal . https://arxiv.org/abs/2502.12118, http://arxiv.org/abs/2502.12118 arXiv:2502.12118

    2025 arXiv

  28. [36]

    , author Wang, Z

    author Shen, G. , author Wang, Z. , author Delalleau, O. , author Zeng, J. , author Dong, Y. , author Egert, D. , author Sun, S. , author Zhang, J.J. , author Jain, S. , author Taghibakhshi, A. , author Ausin, M.S. , author Aithal, A. , author Kuchaiev, O. , year 2024 . title ...

    2024

  29. [37]

    , author Li, J

    author Sun, S. , author Li, J. , author Yuan, W. , author Yuan, R. , author Li, W. , author Liu, P. , year 2024 . title The critique of critique , in: editor Ku, L.W. , editor Martins, A. , editor Srikumar, V. (Eds.), booktitle Findings of the Association for Computational Lin...

    2024 doi

  30. [38]

    , year 2025 a

    author Team, Q. , year 2025 a. title Qwen3 . https://qwenlm.github.io/blog/qwen3/

    2025

  31. [39]

    , year 2025 b

    author Team, Q. , year 2025 b. title Qwq-32b: Embracing the power of reinforcement learning . https://qwenlm.github.io/blog/qwq-32b/

    2025

  32. [40]

    title Tree sitter

    author TreeSitter , year 2013 . title Tree sitter . howpublished https://github.com/tree-sitter/tree-sitter

    2013

  33. [41]

    , author Li, L

    author Wang, P. , author Li, L. , author Shao, Z. , author Xu, R.X. , author Dai, D. , author Li, Y. , author Chen, D. , author Wu, Y. , author Sui, Z. , year 2024 a. title Math-shepherd: Verify and reinforce llms step-by-step without human annotations . https://arxiv.org/abs/...

    2024 arXiv

  34. [42]

    , author Wei, J

    author Wang, X. , author Wei, J. , author Schuurmans, D. , author Le, Q. , author Chi, E. , author Narang, S. , author Chowdhery, A. , author Zhou, D. , year 2023 a. title Self-consistency improves chain of thought reasoning in language models . https://arxiv.org/abs/2203.1117...

    2023 arXiv

  35. [43]

    , author Kordi, Y

    author Wang, Y. , author Kordi, Y. , author Mishra, S. , author Liu, A. , author Smith, N.A. , author Khashabi, D. , author Hajishirzi, H. , year 2023 b. title Self-instruct: Aligning language models with self-generated instructions , in: editor Rogers, A. , editor Boyd-Graber...

    2023

  36. [44]

    , author Liu, Q

    author Wang, Y. , author Liu, Q. , author Xu, J. , author Liang, T. , author Chen, X. , author He, Z. , author Song, L. , author Yu, D. , author Li, J. , author Zhang, Z. , author Wang, R. , author Tu, Z. , author Mi, H. , author Yu, D. , year 2025 a. title Thoughts are all ov...

    2025 arXiv

  37. [45]

    , author Yue, X

    author Wang, Y. , author Yue, X. , author Chen, W. , year 2025 b. title Critique fine-tuning: Learning to critique is more effective than learning to imitate . https://arxiv.org/abs/2501.17703, http://arxiv.org/abs/2501.17703 arXiv:2501.17703

    2025 arXiv

  38. [46]

    , author Bukharin, A

    author Wang, Z. , author Bukharin, A. , author Delalleau, O. , author Egert, D. , author Shen, G. , author Zeng, J. , author Kuchaiev, O. , author Dong, Y. , year 2024 b. title Helpsteer2-preference: Complementing ratings with preferences . https://arxiv.org/abs/2410.01257, ht...

    2024 arXiv

  39. [47]

    , author Dong, Y

    author Wang, Z. , author Dong, Y. , author Delalleau, O. , author Zeng, J. , author Shen, G. , author Egert, D. , author Zhang, J.J. , author Sreedhar, M.N. , author Kuchaiev, O. , year 2024 c. title Helpsteer2: Open-source dataset for training top-performing reward models . h...

    2024 arXiv

  40. [48]

    , author Dong, Y

    author Wang, Z. , author Dong, Y. , author Zeng, J. , author Adams, V. , author Sreedhar, M.N. , author Egert, D. , author Delalleau, O. , author Scowcroft, J. , author Kant, N. , author Swope, A. , author Kuchaiev, O. , year 2024 d. title H elp S teer: Multi-attribute helpful...

    2024

  41. [49]

    , author Wang, X

    author Wei, J. , author Wang, X. , author Schuurmans, D. , author Bosma, M. , author brian ichter , author Xia, F. , author Chi, E.H. , author Le, Q.V. , author Zhou, D. , year 2022 . title Chain of thought prompting elicits reasoning in large language models , in: editor Oh, ...

    2022

  42. [50]

    , author Cassano, F

    author Wei, Y. , author Cassano, F. , author Liu, J. , author Ding, Y. , author Jain, N. , author Mueller, Z. , author de Vries, H. , author Werra, L.V. , author Guha, A. , author ZHANG, L. , year 2024 a. title Selfcodealign: Self-alignment for code generation , in: booktitle ...

    2024

  43. [51]

    , author Wang, Z

    author Wei, Y. , author Wang, Z. , author Liu, J. , author Ding, Y. , author Zhang, L. , year 2024 b. title Magicoder: empowering code generation with oss-instruct , in: booktitle Proceedings of the 41st International Conference on Machine Learning , publisher JMLR.org

    2024

  44. [52]

    , author Huang, D

    author Wu, Y. , author Huang, D. , author Shi, W. , author Wang, W. , author Gao, L. , author Liu, S. , author Nan, Z. , author Yuan, K. , author Zhang, R. , author Zhang, X. , et al., year 2024 . title Inversecoder: Unleashing the power of instruction-tuned code llms with inv...

    2024 arXiv

  45. [53]

    , author Sun, Z

    author Wu, Y. , author Sun, Z. , author Li, S. , author Welleck, S. , author Yang, Y. , year 2025 . title Inference scaling laws: An empirical analysis of compute-optimal inference for problem-solving with language models . https://arxiv.org/abs/2408.00724, http://arxiv.org/ab...

    2025 arXiv

  46. [54]

    , author Sun, Q

    author Xu, C. , author Sun, Q. , author Zheng, K. , author Geng, X. , author Zhao, P. , author Feng, J. , author Tao, C. , author Lin, Q. , author Jiang, D. , year 2024 . title Wizard LM : Empowering large pre-trained language models to follow complex instructions , in: bookti...

    2024

  47. [55]

    , author Liu, Y

    author Xu, Z. , author Liu, Y. , author Yin, Y. , author Zhou, M. , author Poovendran, R. , year 2025 . title Kodcode: A diverse, challenging, and verifiable synthetic dataset for coding . https://arxiv.org/abs/2503.02951, http://arxiv.org/abs/2503.02951 arXiv:2503.02951

    2025 arXiv

  48. [56]

    , author Zhang, B

    author Yang, A. , author Zhang, B. , author Hui, B. , author Gao, B. , author Yu, B. , author Li, C. , author Liu, D. , author Tu, J. , author Zhou, J. , author Lin, J. , author Lu, K. , author Xue, M. , author Lin, R. , author Liu, T. , author Ren, X. , author Zhang, Z. , yea...

    2024 arXiv

  49. [57]

    , author Chiang, W.L

    author Yang, S. , author Chiang, W.L. , author Zheng, L. , author Gonzalez, J.E. , author Stoica, I. , year 2023 . title Rethinking benchmark and contamination for language models with rephrased samples . http://arxiv.org/abs/2311.04850 arXiv:2311.04850

    2023 arXiv

  50. [58]

    , author Zhang, Z

    author Yu, Q. , author Zhang, Z. , author Zhu, R. , author Yuan, Y. , author Zuo, X. , author Yue, Y. , author Fan, T. , author Liu, G. , author Liu, L. , author Liu, X. , author Lin, H. , author Lin, Z. , author Ma, B. , author Sheng, G. , author Tong, Y. , author Zhang, C. ,...

    2025 arXiv

  51. [59]

    , author Chen, Z

    author Yu, Y. , author Chen, Z. , author Zhang, A. , author Tan, L. , author Zhu, C. , author Pang, R.Y. , author Qian, Y. , author Wang, X. , author Gururangan, S. , author Zhang, C. , author Kambadur, M. , author Mahajan, D. , author Hou, R. , year 2025 b. title Self-generat...

    2025

  52. [60]

    , author Jiang, D

    author Zeng, H. , author Jiang, D. , author Wang, H. , author Nie, P. , author Chen, X. , author Chen, W. , year 2025 a. title Acecoder: Acing coder rl via automated test-case synthesis . https://arxiv.org/abs/2502.01718, http://arxiv.org/abs/2502.01718 arXiv:2502.01718

    2025 arXiv

  53. [61]

    , author Cheng, Q

    author Zeng, Z. , author Cheng, Q. , author Yin, Z. , author Zhou, Y. , author Qiu, X. , year 2025 b. title Revisiting the test-time scaling of o1-like models: Do they truly possess test-time scaling capabilities? https://arxiv.org/abs/2502.12215, http://arxiv.org/abs/2502.122...

    2025 arXiv

  54. [62]

    , author Hosseini, A

    author Zhang, L. , author Hosseini, A. , author Bansal, H. , author Kazemi, M. , author Kumar, A. , author Agarwal, R. , year 2025 a. title Generative verifiers: Reward modeling as next-token prediction . https://arxiv.org/abs/2408.15240, http://arxiv.org/abs/2408.15240 arXiv:...

    2025 arXiv

  55. [63]

    , author Lyu, F

    author Zhang, Q. , author Lyu, F. , author Sun, Z. , author Wang, L. , author Zhang, W. , author Hua, W. , author Wu, H. , author Guo, Z. , author Wang, Y. , author Muennighoff, N. , author King, I. , author Liu, X. , author Ma, C. , year 2025 b. title A survey on test-time sc...

    2025 arXiv

  56. [64]

    , author Zheng, C

    author Zhang, Z. , author Zheng, C. , author Wu, Y. , author Zhang, B. , author Lin, R. , author Yu, B. , author Liu, D. , author Zhou, J. , author Lin, J. , year 2025 c. title The lessons of developing process reward models in mathematical reasoning . https://arxiv.org/abs/25...

    2025 arXiv

  57. [65]

    , author Yin, L

    author Zheng, L. , author Yin, L. , author Xie, Z. , author Sun, C. , author Huang, J. , author Yu, C.H. , author Cao, S. , author Kozyrakis, C. , author Stoica, I. , author Gonzalez, J.E. , author Barrett, C. , author Sheng, Y. , year 2024 . title SGL ang: Efficient execution...

    2024

  58. [66]

    , author Zhang, X

    author Zhou, C. , author Zhang, X. , author Song, D. , author Chen, X. , author Gu, W. , author Ma, H. , author Tian, Y. , author Zhang, M. , author Hu, L. , year 2025 . title Refinecoder: Iterative improving of large language models via adaptive critique refinement for code g...

    2025 arXiv

Pith tools

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