Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Confucius3-Math: A Lightweight High-Performance Reasoning LLM for Chinese K-12 Mathematics Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A 14B model post-trained with pure reinforcement learning for about $26K reaches state-of-the-art accuracy on Chinese K-12 math benchmarks, matching or beating much larger models and serving about 15x the throughput of a 671B model.

desk verdict A credible low-cost RL recipe that is undermined by an internal, likely-contaminated benchmark; the public numbers and ablations are worth a look, the headline K-12 claim is not. read the letter →

arxiv 2506.18330 v2 pith:MF3U2EAQ submitted 2025-06-23 cs.LG cs.AIcs.CL

Lixin Wu , Na Cai , Qiao Cheng , Jiachen Wang , Yitao Duan This is my paper · ORCID
classification cs.LGcs.AIcs.CL
keywords reinforcementlearningmathematicalreasoningChineseK-12educationlargelanguagemodelsentropyregularizationgrouprelativepolicyoptimizationdataefficiencyLLMpost-training
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

This paper tries to establish that a 14-billion-parameter model, post-trained with pure reinforcement learning and no supervised fine-tuning, can match or beat much larger models on Chinese K-12 mathematics and competition benchmarks, at a training cost of roughly $26,000 and with about 15 times the serving throughput of a 671-billion-parameter model. The authors report that their model reaches 96.24 on an internal Chinese K-12 benchmark, 98.44 on MATH500, and 81.15 on AIME24, outperforming its own base model by up to 26.98 points. Three proposed mechanisms—targeted entropy regularization, recent sample recovery, and policy-specific hardness weighting—are what stabilize the RL run, cut wasted data, and raise accuracy. If the claims hold, they demonstrate that domain-focused, low-cost RL post-training is a practical route to deployable AI tutoring models, with consequences for educational access and affordability.

What carries the argument

The load-bearing machinery is a modified RL objective built on Group Relative Policy Optimization (GRPO) and its dynamic-sampling variant DAPO. Targeted Entropy Regularization adds $|L_{\text{entropy}} - 0.55| \times 0.001$ to the loss, constraining policy entropy to a narrow band; the paper finds this suppresses the mixed-language and repetition failure modes that otherwise destabilize the run. Policy-Specific Hardness Weighting changes the group-relative advantage estimator to $D(q) = \alpha \mu + 1.256$ with $\alpha = -0.256$, where $\mu$ is the mean reward on a question; this deliberately gives stronger advantage to questions the current policy finds hard, making difficulty relative to the evolving policy rather than a fixed curriculum. Recent Sample Recovery keeps the overflow samples that DAPO's dynamic sampling would discard and feeds them into the next batch, without importance sampling; the paper reports this not only reduces raw data consumption but also improves final accuracy at equal effective data, possibly through a smoothing effect. The paper also removes length normalization from the objective, following a critique that normalization biases the policy toward short correct answers and long incorrect ones.

What would settle it

A reader could settle the main claim by checking whether the 500 CK12-MATH questions overlap the 330,000 proprietary training questions, and then by re-grading Confucius3-Math and DeepSeek-R1 on a freshly sampled, independently verified set of Chinese K-12 problems. If the 96% accuracy and the 3.5-point lead over DeepSeek-R1 persist under model-blind human grading on the fresh sample, the central claim stands; if they vanish, the benchmark is the load-bearing artifact.

Watch

Extended reading notes

Core claim

The paper's central claim is that pure RL post-training of a strong 14B base model is sufficient to elicit state-of-the-art mathematical reasoning in Chinese K-12 and competition settings, without supervised fine-tuning or distillation from a stronger teacher. Starting from DeepSeek-R1-Distill-Qwen-14B, the authors train Confucius3-Math in three stages with gradually expanding context windows, from 4K to 8K to 16K, using GRPO in the first stage and a modified DAPO in the later stages. They report that the final model beats DeepSeek-R1 on CK12-MATH (96.24 vs 92.74), GAOKAO-Bench Math (98.46 vs 93.27), CMATH (96.13 vs 95.81), MATH500 (98.44 vs 97.30), and AIME24 (81.15 vs 79.80), while remaining comparable on MathBench K12 and AIME25. The largest lift over its base model is 26.98 points on AIME25, and the entire training run costs $26K in rented H800 GPU time.

Load-bearing premise

The headline K-12 result depends on CK12-MATH, an internal 500-question benchmark drawn from the same tutoring business that supplied the proprietary training data, and the paper does not show that those questions are cleanly held out or that the language-model judge used to verify answers and steps is unbiased toward the trained model.

Editorial extensions

If this is right

  • A 14B domain model can beat a 671B general reasoning model on the domain's own benchmarks, so model scale is not the only path to top-tier reasoning accuracy.
  • The training recipe is reusable on other verifiable reasoning domains at roughly $26K per run, making state-of-the-art RL post-training accessible to teams with a single 8-GPU server.
  • At $0.15 per million tokens and over 15x the throughput of DeepSeek-R1, the model can serve realistic tutoring traffic on consumer-grade hardware, which is the deployment condition the paper argues matters for educational equity.
  • The largest gains over the base model appear on competition-level sets, up to +26.98 on AIME25, suggesting pure RL extracts latent reasoning already present in a distilled base rather than only memorizing school problems.
  • Because the techniques modify the advantage estimator and data flow rather than the architecture, the paper argues they can be combined with other GRPO/DAPO-family algorithms and other base models.

Reading between the lines

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

  • An implication the paper does not draw: the entropy-target mechanism is a generic treatment for mixed-language outputs in multilingual RL, so it could be tested directly on code generation or multilingual reasoning tasks where language mixing also appears.
  • A testable extension: vary the RSR buffer size and the policy distance between reused samples in controlled runs; if quality improves with distance, the effect is a form of implicit off-policy smoothing, a mechanism the paper only speculates about.
  • The reliance on an LLM-as-a-judge for proprietary K-12 data means the method's reward signal and its main benchmark share a judgment style; a natural extension is to train on judge-verified answers and then re-evaluate with a different judge or human graders to confirm the gains are not judge-specific.
  • The paper computes training cost at a fixed $2 per GPU-hour rental; at different rates the dollar figure scales, but the structural advantage over on-policy distillation, which requires a separate teacher-serving server, should survive because the RL pipeline uses only one server.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 manuscript introduces Confucius3-Math, a 14B-parameter reasoning LLM for Chinese K-12 mathematics, built by pure RL post-training from DeepSeek-R1-Distill-Qwen-14B. Three training innovations are proposed: Targeted Entropy Regularization (TER), Recent Sample Recovery (RSR), and Policy-Specific Hardness Weighting (PSHW). The paper reports large gains over the base model on public benchmarks, SOTA or near-SOTA scores on several competition-level benchmarks, a 3.5-point advantage over DeepSeek-R1 on an internal CK12-MATH benchmark, a training cost of $26K, and roughly 15x the serving throughput of DeepSeek-R1. The central K-12 claim rests on CK12-MATH, an unreleased internal benchmark sampled from the authors' business data, and the evaluation protocol applies an LLM-as-a-judge step check to Confucius3-Math only.

Significance. If the claims hold, the paper is a useful existence proof that domain-specific RL post-training of a 14B model can produce strong mathematical reasoning at low training and inference cost. The ablations in Figures 3 and 4 give credible internal support for each of the three proposed training components, the cost figures are concrete, and open-sourcing the model and code is a plus for reproducibility. The high scores on public benchmarks (MATH500 98.44, AIME24 81.15) are impressive and potentially meaningful. However, the unique K-12 SOTA claim is not independently verifiable: CK12-MATH is an internal, unreleased benchmark drawn from the same business stream that supplied the proprietary RL training data, and no overlap analysis between the 500 evaluation items and the 330k training samples is reported. The LLM-as-a-judge step verification used only for Confucius3-Math is an additional uncontrolled variable. These audit gaps are load-bearing for the paper's positioning and must be addressed before the central claim can be accepted.

major comments (4)
  1. [Section 7 (CK12-MATH) and Section 3.1] The headline K-12 result, Confucius3-Math leading DeepSeek-R1 by 3.5 points on CK12-MATH, depends entirely on an internal 500-question benchmark sampled from actual user queries, while the proprietary training data (Section 3.1) are math questions accumulated during the same business's operations. The open-source data pipeline includes exact, fuzzy, and semantic deduplication (Section 3.2), but no comparable overlap, contamination, or deduplication analysis is reported between the CK12-MATH items and the 330k proprietary training samples. Because identical or near-identical questions plausibly recur in a tutoring business, the observed 3.5-point gap could be a training-set artifact. To support the central claim, the authors should release the CK12-MATH items, or at minimum provide a detailed overlap analysis (exact and fuzzy matching against the training pool), along with a clear statement of how the set was held out from training.
  2. [Section 7 (LLM-as-a-judge protocol)] The evaluation section states that CK12-MATH includes intermediate solution steps and that for Confucius3-Math a problem is solved only when both the final answer and the intermediate steps are judged correct by LLM-as-a-judge, whereas the baselines are presumably evaluated on final answers only. This asymmetric protocol makes the comparison against DeepSeek-R1 and other baselines difficult to interpret. The paper does not report which judge model was used, how the judge was prompted or calibrated, what its agreement with human graders is, or whether the same step-checking protocol was applied to all baselines. The authors should report these details and, ideally, run a version of the evaluation with an identical protocol for all models, including a human-annotated subset.
  3. [Table 3 and Abstract/Introduction SOTA claim] The abstract and introduction claim SOTA performance across a range of benchmarks, but Table 3 shows that Confucius3-Math is below Qwen3-14B on MathBench(K12) (95.10 vs. 96.51) and on AIME25 (69.95 vs. 70.40). The statement that the model outperforms all models on all but two benchmarks is technically accurate, but those two exceptions are central to the K-12 and current-year competition claims. The authors should temper the SOTA wording to 'comparable or leading on specific benchmarks' and should provide confidence intervals or error bars for the pass@1 estimates, especially for CK12-MATH where k=4 per question and the reported advantage is only 3.5 points.
  4. [Section 4 and Section 6 (base model and 'pure RL' claim)] The manuscript uses DeepSeek-R1-Distill-Qwen-14B, a model that already acquired reasoning capabilities through distillation from DeepSeek-R1, as the base. The phrase 'pure RL' in the introduction and the framing in Section 6 should be clarified to mean pure RL post-training of an already distilled reasoning model, not RL from a base pretrained model. This does not invalidate the approach, but it is material to the claim that RL alone elicits the observed capabilities.
minor comments (5)
  1. [Section 6.2 and Table 2] The three hyperparameters entropy_target (0.55), entropy_coeff (0.001), and the PSHW coefficient alpha (-0.256) are all set empirically, but the paper reports no sensitivity analysis or ablations for them. Since these parameters are integral to the claimed innovations, reporting at least a small sweep or noting the range of stable values would strengthen the contribution.
  2. [Table 3 and Section 7] Several baseline numbers are marked with an asterisk as taken from the respective publications, while others were produced with different sampling parameters and k values. The paper should state for each benchmark which settings were used for each model so that the cross-model comparisons are directly interpretable.
  3. [Section 8] The throughput comparison scales Confucius3-Math's single-GPU throughput by a factor of 8 to compare with DeepSeek-R1 on 8 GPUs. This assumes perfectly linear scaling across GPUs and should be stated explicitly as an idealized comparability assumption; reporting raw per-configuration numbers would be more informative.
  4. [Throughout] There are several typos and formatting issues, including 'summerarized' in Section 3.1, 'reword model' in Section 5, 'rigorious' in Section 6.3, and the mislabeled 'summarized' in the Table 1 caption. These should be corrected in a final revision.
  5. [Abstract and Section 1] The claim that the model runs 'efficiently on a single consumer-grade GPU' is supported by the 4090D experiment, but the paper should clarify whether the model is open-sourced with a specific license and whether the released GitHub repository contains the model weights or only code and training recipes.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: CK12-MATH is an internal-benchmark audit risk, not a derivation that reduces to the training inputs.

full rationale

I walked the claimed derivation chain: base-model selection (Sec 4), reward modeling (Sec 5), the three RL techniques (Sec 6.1-6.3), and the benchmark evaluation (Sec 7). None of the reported results is obtained by defining the output in terms of the input. TER adds |entropy_loss - entropy_target| * entropy_coeff to the loss; this is an explicit regularizer with an empirically set target, not a hidden restatement of the evaluation metrics. RSR reuses overflow samples from the dynamic-sampling buffer; it changes data scheduling, but no benchmark accuracy is defined by the reuse. PSHW modifies the DAPO advantage estimator with D(q)=alpha*mu+1.256; this is a policy-gradient weighting scheme, not a fitted version of CK12-MATH or MATH500 scores. The public benchmarks (GAOKAO-Bench, MathBench, CMATH, MATH500, AIME24/25) are external, and the paper reports pass@1 following DeepSeek's k=64 protocol for the competition sets, so those results have independent content. There is no load-bearing self-citation: the only in-house model mentioned, Confucius-o1-14B, is examined and rejected as a base. The genuine concern is external validity, not circularity: CK12-MATH (Sec 7) is sampled from actual user queries while the proprietary training data (Sec 3.1) are accumulated during operation of the same business, and no overlap analysis is reported. Contamination there would undermine the headline K-12 advantage, but it would do so as a data-audit failure, not because the CK12 number is an algebraic consequence of the training set. Under the hard rule that circularity requires a specific reduction such as Eq X = Eq Y by construction or a fitted parameter renamed as a prediction, I find no circular step.

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

The model's central result depends on unverified assumptions about data cleanliness, judge reliability, and cross-paper comparability. The three named training innovations introduce at least three hand-set scalar parameters. No new physical or ontological entities are proposed.

free parameters (3)
  • entropy_target = 0.55
    Set empirically in Section 6.1 to stabilize entropy and suppress mixed-language outputs; no principled derivation.
  • entropy_coeff = 0.001
    Set as a fixed coefficient in Section 6.1 for the Targeted Entropy Regularization term.
  • PSHW advantage coefficient alpha = -0.256
    Set empirically in Section 6.2: D(q) = alpha * mu + 1.256. The offset 1.256 is also hand-chosen; no derivation or sensitivity analysis is provided.
assumptions (4)
  • domain assumption DeepSeek-R1-Distill-Qwen-14B is a suitable base model and its early policy entropy predicts trainability under RL.
    Base model selection in Section 4 uses 150 steps of RL and entropy as a proxy; full-training transfer is assumed.
  • domain assumption Math-Verify and the proprietary LLM-as-a-judge reward correctly determine correctness of final answers and solving steps.
    Reward modeling in Section 5; no human agreement or adversarial analysis is reported.
  • domain assumption The deduplication and test-split removal in Section 3.2 prevent contamination between training data and public benchmarks.
    The paper states removal but does not provide a contamination audit or hash-level overlap check.
  • domain assumption Benchmark scores taken from other technical reports are comparable to the authors' own evaluation protocol.
    Section 7 uses asterisked published numbers with different sampling parameters and system prompts; this is not a controlled comparison.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Confucius3-Math: A Lightweight High-Performance Reasoning LLM for Chinese K-12 Mathematics Learning." pith.science (2026). https://pith.science/paper/MF3U2EAQ

@misc{pith2026250618330,
  author       = {Pith},
  title        = {Pith review of: Confucius3-Math: A Lightweight High-Performance Reasoning LLM for Chinese K-12 Mathematics Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MF3U2EAQ}},
  note         = {Machine review of arXiv:2506.18330}
}
read the original abstract

We introduce Confucius3-Math, an open-source large language model with 14B parameters that (1) runs efficiently on a single consumer-grade GPU; (2) achieves SOTA performances on a range of mathematical reasoning tasks, outperforming many models with significantly larger sizes. In particular, as part of our mission to enhancing education and knowledge dissemination with AI, Confucius3-Math is specifically committed to mathematics learning for Chinese K-12 students and educators. Built via post-training with large-scale reinforcement learning (RL), Confucius3-Math aligns with national curriculum and excels at solving main-stream Chinese K-12 mathematical problems with low cost. In this report we share our development recipe, the challenges we encounter and the techniques we develop to overcome them. In particular, we introduce three technical innovations: Targeted Entropy Regularization, Recent Sample Recovery and Policy-Specific Hardness Weighting. These innovations encompass a new entropy regularization, a novel data scheduling policy, and an improved group-relative advantage estimator. Collectively, they significantly stabilize the RL training, improve data efficiency, and boost performance. Our work demonstrates the feasibility of building strong reasoning models in a particular domain at low cost. We open-source our model and code at https://github.com/netease-youdao/Confucius3-Math.

Figures

Figures reproduced from arXiv: 2506.18330 by the authors.

Figure 1
Figure 1. Benchmark performance of Confucius3-Math [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Multi-Stage Length Expansion to substitute tokens with the same meaning from different languages. We tried adding linguistic constraints to the reward, but it was ineffective. Therefore we introduce “Targeted Entropy Regularization” by adding the following term to the loss function |entropy_loss − entropy_target| ∗ entropy_coeff where entropy_target is desired entropy level set empirically to 0.55. This is different… view at source ↗
Figure 3
Figure 3. Effects of Target Entropy Regularization on Policy Entropy [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: RSR+DAPO vs. DAPO Baselines. We compare against DeepSeek-R1, the SOTA open-source reasoning model with 671B parameters, our base model DeepSeek-R1-Distill-Qwen-14B, as well as two other strong open-source models: Qwen3-14B (Yang et al., 2025) and QwQ-32B (QwenTeam). Qw…
Figure 5
Figure 5. Figure 5: Deployment Performance Result Results are shown in figure 5. For each run, solid line represents total throughput, in tokens per second, while dashed line is output throughput, also in tokens per second. DeepSeek-R1 and Confucius3-Math’s performances on H800 are shown …

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Credit Cards, Confusion, Computation, and Consequences: What Can We Uncover About Language Model Reasoning?

    cs.CL 2026-07 conditional novelty 6.5 of 10

    CreditCardQA shows LLMs err mainly on credit-card contractual conditions and comparisons, not arithmetic, with Program-of-Thought narrowing open–closed model gaps.

  2. UloRL:An Ultra-Long Output Reinforcement Learning Approach for Advancing Large Language Models' Reasoning Abilities

    cs.CL 2025-07 conditional novelty 6.0 of 10

    A segment rollout plus dynamic masking of confident positive tokens lets a 30B-A3B reasoning model beat a 235B-A22B model on AIME2025 and BeyondAIME after 128k-token RL training.

Reference graph

Works this paper leans on

25 extracted references · 1 canonical work pages · cited by 2 Pith papers

  1. [3]

    Shaping AI's Impact on Billions of Lives

    URL https: //arxiv.org/abs/2412.02730. DeepSeek-AI, Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., Zhang, X., Yu, X., Wu, Y ., Wu, Z. F., Gou, Z., Shao, Z., Li, Z., Gao, Z., Liu, A., Xue, B., Wang, B., Wu, B., Feng, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., Dai, D., Chen, D., Ji, D., Li, E., Lin, F...

    work page Pith review arXiv

  2. [4]

    URL https://arxiv.org/abs/2501.12948. Guha, E., Marten, R., Keh, S., Raoof, N., Smyrnis, G., Bansal, H., Nezhurina, M., Mercat, J., Vu, T., Sprague, Z., Suvarna, A., Feuer, B., Chen, L., Khan, Z., Frankel, E., Grover, S., Choi, C., Muennighoff, N., Su, S., Zhao, W., Yang, J., Pimpalgaonkar, S., Sharma, K., Ji, C. C.-J., Deng, Y ., Pratt, S., Ramanujan, V ...

    arXiv

  3. [5]

    He, J., Liu, J., Liu, C

    URL https://arxiv.org/abs/2506.04178. He, J., Liu, J., Liu, C. Y ., Yan, R., Wang, C., Cheng, P., Zhang, X., Zhang, F., Xu, J., Shen, W., Li, S., Zeng, L., Wei, T., Cheng, C., An, B., Liu, Y ., and Zhou, Y . Skywork open reasoner 1 technical report.arXiv preprint arXiv:2505.22312,

    arXiv

  4. [6]

    HuggingFace

    URL https://arxiv.org/abs/2503.24290. HuggingFace. Math-verify.https://github.com/huggingface/Math-Verify. 14 Jones, A. L. Scaling scaling laws with board games,

    arXiv

  5. [8]

    URL http://dx.doi.org/10.21203/rs.3

    doi: 10.21203/rs.3.rs-4243877/v1. URL http://dx.doi.org/10.21203/rs.3. rs-4243877/v1. LI, J., Beeching, E., Tunstall, L., Lipkin, B., Soletskyi, R., Huang, S. C., Rasul, K., Yu, L., Jiang, A., Shen, Z., Qin, Z., Dong, B., Zhou, L., Fleureau, Y ., Lample, G., and Polu, S. Numi- namath. [https://huggingface.co/AI-MO/NuminaMath-1.5](https://github.com/ proje...

    doi:10.21203/rs.3.rs-4243877/v1

  6. [10]

    Metropolis, N

    URLhttps://arxiv.org/abs/2503.20783. Metropolis, N. and Ulam, S. The Monte Carlo method.j-J-AM-STAT-ASSOC, 44(247):335–341, September

    arXiv

  7. [12]

    URL https://arxiv.org/ abs/2501.19393. NVIDIA. Nemo curator: The gpu-accelerated open source framework for efficient generative ai model data curation.https://github.com/NVIDIA/NeMo-Curator. OpenAI. Learning to reason with llms. Open AI blog,

    arXiv

  8. [14]

    QwenTeam

    URLhttps://arxiv.org/abs/2410.18982. QwenTeam. Qwq-32b: Embracing the power of reinforcement learning. https://qwenlm.github. io/blog/qwq-32b/. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

    arXiv

Show all 25 references
  1. [16]

    Silver, D., Huang, A., Maddison, C

    URLhttps://arxiv.org/abs/2402.03300. Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., van den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V ., Lanctot, M., Dieleman, S., Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T., Leac...

    arXiv

  2. [17]

    Tesauro, G

    URL https://arxiv.org/ abs/2501.12599. Tesauro, G. Td-gammon, a self-teaching backgammon program, achieves master-level play.Neural Comput., 6(2):215–219, March

    arXiv

  3. [19]

    URL https: //doi.org/10.1057/s41599-025-04787-y

    doi: 10.1057/s41599-025-04787-y. URL https: //doi.org/10.1057/s41599-025-04787-y. Wang, W., Li, Z., Lian, D., Ma, C., Song, L., and Wei, Y . Mitigating the language mismatch and repetition issues in llm-based machine translation via model editing,

    doi

  4. [20]

    Wang, Z., Cui, G., Wan, K., and Zhao, W

    URL https: //arxiv.org/abs/2410.07054. Wang, Z., Cui, G., Wan, K., and Zhao, W. Dump: Automated distribution-level curriculum learning for rl-based llm post-training,

    arXiv

  5. [21]

    Wei, T., Luan, J., Liu, W., Dong, S., and Wang, B

    URLhttps://arxiv.org/abs/2504.09710. Wei, T., Luan, J., Liu, W., Dong, S., and Wang, B. Cmath: Can your language model pass chinese elementary school math test?,

  6. [23]

    URLhttps://arxiv.org/abs/2502.03387. Yu, Q., Zhang, Z., Zhu, R., Yuan, Y ., Zuo, X., Yue, Y ., Dai, W., Fan, T., Liu, G., Liu, L., Liu, X., Lin, H., Lin, Z., Ma, B., Sheng, G., Tong, Y ., Zhang, C., Zhang, M., Zhang, W., Zhu, H., Zhu, J., Chen, J., Chen, J., Wang, C., Yu, H., ...

    arXiv

  7. [24]

    Zhang, X., Li, C., Zong, Y ., Ying, Z., He, L., and Qiu, X

    URLhttps://arxiv.org/abs/2503.14476. Zhang, X., Li, C., Zong, Y ., Ying, Z., He, L., and Qiu, X. Evaluating the performance of large language models on gaokao benchmark.arXiv preprint arXiv:2305.12474,

    arXiv

  8. [25]

    Zhu, T., Zhang, K., and Wang, W

    URL https://arxiv.org/ abs/2411.14405. Zhu, T., Zhang, K., and Wang, W. Y . Embracing ai in education: Understanding the surge in large language model use by secondary students,

    arXiv

  9. [26]

    URL https://arxiv.org/abs/2411.18708. 16

    arXiv

  10. [1949]

    doi: https://doi.org/10

    ISSN 0162-1459 (print), 1537-274X (electronic). doi: https://doi.org/10. 2307/2280232. URL http://links.jstor.org/sici?sici=0162-1459%28194909%2944% 3A247%3C335%3ATMCM%3E2.0.CO%3B2-3;http://www.jstor.org/stable/2280232. Muennighoff, N., Yang, Z., Shi, W., Li, X. L., Fei-Fei, L...

  11. [1992]

    URL https://doi.org/ 10.1007/BF00992699

    doi: 10.1007/BF00992699. URL https://doi.org/ 10.1007/BF00992699. Liu, H., Zheng, Z., Qiao, Y ., Duan, H., Fei, Z., Zhou, F., Zhang, W., Zhang, S., Lin, D., and Chen, K. Mathbench: Evaluating the theory and application proficiency of llms with a hierarchical mathematics benchm...

    doi

  12. [1994]

    doi: 10.1162/neco.1994.6.2.215

    ISSN 0899-7667. doi: 10.1162/neco.1994.6.2.215. URL https://doi.org/10.1162/neco.1994.6.2.215. Wang, J. and Fan, W. The effect of chatgpt on students’learning performance, learning percep- tion, and higher-order thinking: insights from a meta-analysis.Humanities and Social Sci...

    1994 doi

  13. [2010]

    URL https: //onlinelibrary.wiley.com/doi/abs/10.1002/asmb.874

    doi: https://doi.org/10.1002/asmb.874. URL https: //onlinelibrary.wiley.com/doi/abs/10.1002/asmb.874. Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y . K., Wu, Y ., and Guo, D. Deepseekmath: Pushing the limits of mathematical reasoning in ope...

    doi

  14. [2021]

    URLhttps://doi.org/10.1038/s41586-021-03819-2

    doi: 10.1038/s41586-021-03819-2. URLhttps://doi.org/10.1038/s41586-021-03819-2. Kestin, G., Miller, K., Klales, A., Milbourne, T., and Ponti, G. Ai tutoring outperforms active learning. May

    doi

  15. [2023]

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al

    Curran Associates Inc. Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    arXiv

  16. [2024]

    Brown, N

    URL https: //arxiv.org/abs/2407.21787. Brown, N. and Sandholm, T. Safe and nested subgame solving for imperfect-information games. Advances in neural information processing systems, 30,

    arXiv

  17. [2025]

    Cobbe, K., Kosaraju, V ., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., and Schulman, J

    URL https://arxiv.org/abs/2505.14970. Cobbe, K., Kosaraju, V ., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., and Schulman, J. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168,

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.