Pith. sign in

REVIEW 5 major objections 6 minor 6 cited by

Reinforcement Pre-Training

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that treating next-token prediction as a reinforcement-learned reasoning task improves language modeling accuracy and scales with compute.

desk verdict A clean RL recipe for next-token reasoning, but the missing same-corpus NTP control leaves the core effect unidentified. read the letter →

arxiv 2506.08007 v1 pith:AOCPRYJ4 submitted 2025-06-09 cs.CL

Qingxiu Dong , Li Dong , Yao Tang , Tianzhu Ye , Yutao Sun , Zhifang Sui , Furu Wei This is my paper · ORCID
classification cs.CL
keywords reinforcementpre-trainingnext-tokenpredictionreasoninglearningverifiablerewardslanguagemodelwithscalinglaws
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 proposes casting ordinary next-token prediction as a reasoning problem: before predicting the next token, the model writes a chain of thought, and it is rewarded only when the predicted continuation is exactly the corpus's ground-truth continuation. The authors train a 14B reasoning model this way on 4,428 competition-math problems and report that next-token prediction accuracy improves on all difficulty splits, matching a 32B model under standard prediction; they also find that accuracy follows a smooth power law in training compute. The paper frames RPT as a scalable way to apply RL to large unannotated text corpora, avoiding the annotation bottleneck of verifiable-reward RL while keeping rewards verifiable. If the effect transfers beyond math text, RPT would make RL a first-class pre-training objective rather than a post-training step.

What carries the argument

The central mechanism is the next-token reasoning task coupled with a prefix-matching reward: the model samples $G$ thinking trajectories per context, each ending in a predicted continuation, and a deterministic verifier returns reward 1 only if the predicted continuation is a byte-prefix of the ground-truth completion at a valid token boundary. Entropy-based filtering selects high-entropy token positions for training, and the group-relative policy optimization update maximizes the expected reward. This combination converts unannotated text into a corpus of verifiable reasoning problems, which is what lets the paper claim RL can scale to pre-training.

What would settle it

Train RPT on a general-domain text corpus, or evaluate the current model on held-out non-math web text, and compare next-token accuracy with standard next-token pretraining at matched compute; if the advantage disappears outside math-like, high-entropy positions, the general pre-training claim is not supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, RPT reframes next-token prediction as a next-token reasoning task: for each position in the corpus, the model generates a chain-of-thought and then a predicted continuation, and it receives a verifiable intrinsic reward based on whether that continuation matches the ground truth from the text itself. The reward is $r^i_t = 1$ when the byte sequence of the prediction is an exact prefix of the ground-truth continuation and ends at a valid token boundary, and $0$ otherwise. Trained this way on 4,428 competition-math problems with entropy-based filtering of easy tokens, the 14B model improves next-token accuracy on every difficulty split relative to its starting point and matches a 32B model under standard next-token prediction. The paper also shows that it provides a stronger starting point for later reinforcement fine-tuning and improves zero-shot scores on SuperGPQA and MMLU-Pro.

Load-bearing premise

The gains seen on competition-math text will transfer to ordinary web text; the paper frames RPT as general-purpose but only trains on math documents.

Editorial extensions

If this is right

  • If RPT's gains hold, RL no longer needs annotated question-answer pairs; the next-token signal in any text corpus becomes the reward, so RL can be applied at pre-training scale.
  • On OmniMATH, RPT raises next-token accuracy at every difficulty level, and the 14B model matches a 32B model trained with standard next-token prediction.
  • RPT makes subsequent RL fine-tuning more sample-efficient: after 15 epochs on 256 examples, the RPT-initialized model reaches 58.3 percent versus 52.7 percent for the standard start.
  • The scaling curves fit a power law with R-squared above 0.98, so next-token accuracy under RPT should keep improving as training compute grows.
  • Zero-shot reasoning on general benchmarks improves, e.g. SuperGPQA from 36.1 to 39.0 and MMLU-Pro from 68.9 to 71.1 relative to the base reasoning model.

Reading between the lines

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

  • The paper leaves open whether the benefit transfers from competition-math text to ordinary web text; if the gain is concentrated at high-entropy positions like equations and formal symbols, the general pre-training claim would need narrowing.
  • An ablation without the entropy-filtering step would clarify how much of the gain comes from the RL objective itself rather than from concentrating training on difficult positions.
  • Because the reward only checks that the predicted continuation is a prefix of the ground truth at a token boundary, the same setup could extend to larger spans or to other exact-match completion tasks.
  • If the compute-for-accuracy tradeoff seen at 14B persists at larger scales, RPT could change the preferred allocation of compute between model size and RL training.
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

5 major / 6 minor

Summary. The paper introduces Reinforcement Pre-Training (RPT), which reframes standard next-token prediction as a reasoning task trained with reinforcement learning. For each prefix in a pre-training corpus, the model generates a chain-of-thought and then predicts the next token, receiving a binary prefix-matching reward against the ground-truth continuation. Experiments use DeepSeek-R1-Distill-Qwen-14B as the base model and train on entropy-filtered tokens from the OmniMATH dataset. The paper reports that RPT-14B improves next-token prediction accuracy on OmniMATH validation splits relative to the base model's standard next-token prediction, exhibits power-law scaling with training compute, provides a better starting point for RLVR fine-tuning on Skywork-OR1, and improves zero-shot accuracy on SuperGPQA and MMLU-Pro. The authors position RPT as a new general-purpose pre-training paradigm that scales RL to unannotated text.

Significance. If the central claim were established, RPT would be a noteworthy contribution: it replaces the standard likelihood objective during pre-training with a verifiable-reward RL objective over the intrinsic next-token signal, potentially avoiding reward hacking and annotation bottlenecks while adding a new compute-scaling axis. The paper's strengths include a clean problem formulation, a well-specified prefix-matching reward, hyperparameter detail, prompt-template ablations, and an explicit limitations section. However, the current evidence is narrower than the framing: experiments are confined to 4,428 OmniMATH competition problems on a 14B reasoning model, and the headline comparison lacks a matched standard-NTP control on the same filtered corpus. As a result, the paper supports a conditional claim about RPT on math-like text but does not yet substantiate the general 'web-text corpus' scaling paradigm advertised in the abstract and Figure 1.

major comments (5)
  1. [Section 4.1, Table 1] The headline next-token accuracy comparison is missing a matched standard-NTP control trained on the same entropy-filtered OmniMATH data. The baselines Qwen2.5-14B and R1-Distill-Qwen-14B have never been trained on OmniMATH, so the observed gains could in principle come from continued in-domain training on math text rather than from the RL/next-token-reasoning objective. To support the causal claim that 'incentivizing next-token reasoning' drives the improvement, the paper should add a control that continues standard next-token training on the identical filtered corpus with comparable compute and hyperparameters, evaluated on the same next-token accuracy metric used in Table 1.
  2. [Section 4.3, Table 2] The only same-corpus NTP control, '+ Continual NTP training', is evaluated on a different downstream RLVR metric and collapses from 51.2 to 10.7 before any RL fine-tuning. This 40-point drop is unexplained and suggests either severe catastrophic forgetting, an unstable training setup, or an evaluation artifact. As reported, this control cannot serve as a reliable benchmark for comparing NTP and RPT, and it does not provide the needed next-token accuracy measurement. The paper should either repair this control or present a separate stable NTP baseline.
  3. [Section 4.4, Table 3] The claim that RPT-14B 'surpasses the significantly larger R1-Distill-Qwen-32B' is based on an asymmetric comparison: RPT-14B is evaluated in reasoning mode while R1-Distill-Qwen-32B is evaluated in standard next-token prediction mode. This conflates the method with added test-time reasoning. A fair comparison would evaluate R1-Distill-Qwen-32B in the same reasoning mode or evaluate RPT-14B in standard mode. The comparison against R1-Distill-Qwen-14B in reasoning mode is more appropriate and should be the primary evidence.
  4. [Section 4.2, Eq. (5), Figure 5] The scaling-law claim is weaker than presented. Equation (5) has three free parameters (A, alpha, P*) fitted to only six training-compute checkpoints spanning roughly one order of magnitude on the horizontal axis. The reported R^2 values describe in-sample fit, not predictive power; no extrapolation, confidence intervals, or held-out compute points are provided. This is acceptable as a descriptive trend statement, but the paper should not call it a scaling law without additional checkpoints or a predictive validation.
  5. [Abstract, Section 3.3, Section 6] The abstract and Figure 1 frame RPT as 'scaling RL to the web-text corpus' and as a general pre-training paradigm, but all experiments train on 4,428 OmniMATH competition math problems, and Section 6 explicitly defers broader-domain evaluation to future work. This is an acknowledged limitation, but it is load-bearing for the title claim of 'pre-training' as a general-purpose method. Either the claims should be narrowed to math-domain pre-training, or the paper should include at least a small-scale demonstration on a non-math corpus to support the general claim.
minor comments (6)
  1. [Table 5 caption] The word 'Hyperparamters' is misspelled; it should be 'Hyperparameters'.
  2. [Section 4.4, first paragraph] The phrase 'as well as the the reasoning performance' contains a duplicated 'the'.
  3. [Table 3] The entry 'R1-Distill-Qwen14B' is missing a hyphen; it should read 'R1-Distill-Qwen-14B' for consistency with the rest of the paper.
  4. [Figure 4] The abbreviations 'R1-Qwen-14B' and 'R1-Qwen-32B' are used without definition in the figure; the caption should spell out 'R1-Distill-Qwen-14B/32B' or define the shorthand.
  5. [Section 4.1, paragraph on validation splits] The description of easy/medium/hard splits is ambiguous: a token position with entropy, say, 2.0 would satisfy all three thresholds. Please clarify whether the splits are cumulative overlapping sets or disjoint bins, since Table 1 reports three separate accuracy values.
  6. [Section D, Table 8] The metrics 'Random@1' and 'Pass@8' are not defined in the main text or the appendix; a one-sentence definition would help the reader interpret the prompt-template comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RPT's central accuracy gains are reported on held-out validation and external benchmarks, not derived from the training objective by construction.

full rationale

The paper's derivation chain is empirical rather than circular. RPT's training objective (Eq. 4) maximizes a prefix-matching reward (Eq. 3) for next-token prediction, and the headline results in Table 1 are next-token prediction accuracies on a held-out 200-sample OmniMATH validation split, while Tables 2 and 3 add downstream RLVR and zero-shot MMLU-Pro/SuperGPQA evaluations. Although the evaluation metric is closely aligned with the training reward, this alignment is the method being tested, not a concealed identity: the held-out split and external benchmarks are not used to fit any parameter in Eq. (4). The scaling curves in Figure 5 are descriptive power-law fits (Eq. 5) to observed checkpoint accuracies, not out-of-sample predictions; reporting high R-squared values is a goodness-of-fit statement, not an inference derived from the model's own assumptions. The paper even defers establishing scaling laws to future work, further indicating the curves are illustrative rather than load-bearing derivations. Self-citations in the manuscript concern infrastructure (verl), exact on-policy RL settings, and reasoning-pattern analysis references; none of these carries the central claim that RPT improves next-token accuracy. The absence of a matched standard-NTP control on the same corpus for the headline metric is an experimental-design or correctness concern, not circularity, because the claimed improvement is still measured on data and tasks independent of the training objective's parameters.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim depends on an unreported entropy threshold, a scaling-law parametric form fitted to six points, and the untested assumption that math-corpus results generalize to web text; no new physical or conceptual entities are introduced.

free parameters (3)
  • Entropy filtering threshold for training data = not reported
    Section 3.3 filters out low-entropy positions with a threshold that is never specified; the threshold shapes the training distribution and the easy/medium/hard validation splits, so the headline accuracy numbers depend on it.
  • Validation entropy thresholds (0.5, 1.0, 1.5) = 0.5, 1.0, 1.5
    These thresholds define the easy/medium/hard splits in Section 4.1; they are chosen by the authors and the same proxy model family is used to partition validation data, coupling data selection to the model family being tested.
  • Scaling law parameters A, alpha, P_star = fitted separately for easy/medium/hard, values not reported
    Equation (5) is fitted to six accuracy points per split in Section 4.2; the reported high R2 is a goodness-of-fit to the same data, not a predictive test.
assumptions (5)
  • domain assumption Ground-truth next token is a valid verifiable reward for general-purpose pre-training
    Section 3.2 defines reward as correctness against the corpus's own next token; the paper assumes optimizing this reward yields better representations, but the primary metric optimized is the reward itself.
  • domain assumption Results on OmniMATH transfer to general web text
    Abstract and Figure 1 frame RPT as scaling RL to web-text corpus, while Section 6 concedes experiments are limited to math documents; transfer is untested.
  • domain assumption A reasoning model initialization (R1-Distill-Qwen-14B) is a suitable base for RPT
    Section 3.3 starts from a reasoning-distilled 14B model; the authors list as future work applying RPT to a standard base LM, so the observed gains may be specific to this initialization.
  • ad hoc to paper Power-law scaling form P(C) = A C^{-alpha} + P*
    Equation (5) is asserted in Section 4.2 without derivation; it is fitted to six points per difficulty, so the scaling conclusion depends on this assumed form.
  • domain assumption Entropy of next-token distribution is a proxy for token difficulty
    The entropy filter in Section 3.3 and the validation splits in Section 4.1 assume that higher entropy tokens are harder and that the proxy model's entropy reliably ranks difficulty.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reinforcement Pre-Training." pith.science (2026). https://pith.science/paper/AOCPRYJ4

@misc{pith2026250608007,
  author       = {Pith},
  title        = {Pith review of: Reinforcement Pre-Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AOCPRYJ4}},
  note         = {Machine review of arXiv:2506.08007}
}
read the original abstract

In this work, we introduce Reinforcement Pre-Training (RPT) as a new scaling paradigm for large language models and reinforcement learning (RL). Specifically, we reframe next-token prediction as a reasoning task trained using RL, where it receives verifiable rewards for correctly predicting the next token for a given context. RPT offers a scalable method to leverage vast amounts of text data for general-purpose RL, rather than relying on domain-specific annotated answers. By incentivizing the capability of next-token reasoning, RPT significantly improves the language modeling accuracy of predicting the next tokens. Moreover, RPT provides a strong pre-trained foundation for further reinforcement fine-tuning. The scaling curves show that increased training compute consistently improves the next-token prediction accuracy. The results position RPT as an effective and promising scaling paradigm to advance language model pre-training.

Figures

Figures reproduced from arXiv: 2506.08007 by the authors.

Figure 1
Figure 1. Reinforcement pre-training (RPT) reframes next-token prediction as a reasoning task, where [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of standard next-token prediction and next-token reasoning. Standard next-token prediction estimates the next token in the pre-training corpus directly, while next-token reasoning performs reasoning over multiple tokens before making the prediction. 2 Preliminary Next-Token Prediction (NTP) Next-token prediction is the fundamental training objective for modern large language models [AAA+23]. Given an inpu… view at source ↗
Figure 3
Figure 3. An illustration of reinforcement pre-training. Given a context with a missing continuation, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Average next-token prediction accuracy across data of different difficulty levels. R1-Qwen-14B/32B denote R1- Distill-Qwen-14B/32B, respectively. As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Next-token prediction accuracy of reinforcement pre-training improves consistently with [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Reasoning pattern statistics of R1-Distill-Qwen-14B employed for problem solving and [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 6 Pith papers

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

  1. Beyond Test-Time Memory: State-Space Optimal Control for LLM Reasoning

    cs.LG 2026-03 conditional novelty 7.0 of 10

    A symplectic LQR layer inserted as an adapter into pretrained LLMs yields large gains on MATH-500, AMC and AIME by solving a latent optimal-control problem at inference time.

  2. CPMobius: Iterative Coach-Player Reasoning for Data-Free Reinforcement Learning

    cs.CL 2026-02 conditional novelty 6.0 of 10

    CPMobius uses iterative coach-player reinforcement learning to improve mathematical reasoning in LLMs without external training data, yielding +4.9 average accuracy gains on Qwen2.5-Math-7B-Instruct.

  3. Advancing Event Forecasting through Massive Training of Large Language Models: Challenges, Solutions, and Broader Impacts

    cs.LG 2025-07 conditional novelty 5.0 of 10

    A position paper advocating large-scale training of event forecasting LLMs, with proposals for label selection, counterfactual training data, auxiliary rewards, and multi-source datasets.

  4. Refine Thought: A Test-Time Inference Method for Embedding Model Reasoning

    cs.CL 2025-10 conditional novelty 4.0 of 10

    RT improves embedding-model retrieval on some reasoning-heavy tasks by iteratively refining the query embedding at test time, but reported gains are mixed and the key implementation detail is unspecified.

  5. A Survey on Latent Reasoning

    cs.CL 2025-07 conditional novelty 4.0 of 10

    A survey that organizes latent reasoning methods into vertical recurrence, horizontal recurrence, and infinite-depth diffusion, arguing that silent reasoning can beat explicit chain-of-thought.

  6. Reinforcement Learning Meets Large Language Models: A Survey of Advancements and Applications Across the LLM Lifecycle

    cs.CL 2025-09 conditional novelty 3.0 of 10

    A survey that maps reinforcement learning methods, datasets, benchmarks, and open-source tools across the full training lifecycle of large language models, focusing on verifiable-reward reasoning.

Reference graph

Works this paper leans on

16 extracted references · 1 canonical work pages · cited by 6 Pith papers

  1. [1]

    GPT-4 technical report.arXiv preprint arXiv:2303.08774,

    [AAA+23] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Flo- rencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report.arXiv preprint arXiv:2303.08774,

    arXiv

  2. [4]

    Notably, the performance of R1-Distill-Qwen-14B is evaluated in two different manner: standard next-token prediction and reasoning-based answer prediction (indicated as ‘+ think’)

    C Detailed Results on End Tasks Table 6 and Table 7 present a detailed per-category performance across the general end task bench- marks. Notably, the performance of R1-Distill-Qwen-14B is evaluated in two different manner: standard next-token prediction and reasoning-based answer prediction (indicated as ‘+ think’). The RPT-14B model demonstrates superio...

    work page 1990

  3. [5]

    Measuring massive multitask language understanding.arXiv preprint arXiv:2009.03300,

    [HBB+20] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding.arXiv preprint arXiv:2009.03300,

    arXiv 2009

  4. [8]

    Skywork open reasoner 1 technical report.arXiv preprint arXiv:2505.22312,

    [HLL+25] Jujie He, Jiacai Liu, Chris Yuhao Liu, Rui Yan, Chaojie Wang, Peng Cheng, Xiaoyu Zhang, Fuxiang Zhang, Jiacheng Xu, Wei Shen, Siyuan Li, Liang Zeng, Tianwen Wei, Cheng Cheng, Bo An, Yang Liu, and Yahui Zhou. Skywork open reasoner 1 technical report.arXiv preprint arXiv:2505.22312,

    arXiv

  5. [9]

    Openai o1 system card.arXiv preprint arXiv:2412.16720,

    [JKL+24] Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720,

    arXiv

  6. [10]

    Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei

    [KMH+20] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.CoRR, abs/2001.08361,

    arXiv 2001

  7. [11]

    General-reasoner: Advancing llm reasoning across all domains.arXiv preprint arXiv:2505.14652,

    [MLJ+25] Xueguang Ma, Qian Liu, Dongfu Jiang, Ge Zhang, Zejun Ma, and Wenhu Chen. General-reasoner: Advancing llm reasoning across all domains.arXiv preprint arXiv:2505.14652,

    arXiv

  8. [12]

    HybridFlow: A flexible and efficient RLHF framework.arXiv preprint arXiv: 2409.19256,

    [SZY+24] Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. HybridFlow: A flexible and efficient RLHF framework.arXiv preprint arXiv: 2409.19256,

    arXiv

Show all 16 references
  1. [13]

    DAPO: An open-source LLM reinforcement learning system at scale.ArXiv, abs/2503.14476,

    [YZZ+25] Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, et al. DAPO: An open-source LLM reinforcement lear...

    arXiv

  2. [14]

    Quiet-star: Language models can teach themselves to think before speaking

    [ZHS+24] Eric Zelikman, Georges Harik, Yijia Shao, Varuna Jayasiri, Nick Haber, and Noah D Goodman. Quiet-star: Language models can teach themselves to think before speaking. arXiv preprint arXiv:2403.09629,

    arXiv

  3. [15]

    A survey on test-time scaling in large language models: What, how, where, and how well?arXiv preprint arXiv:2503.24235,

    [ZLS+25a] Qiyuan Zhang, Fuyuan Lyu, Zexu Sun, Lei Wang, Weixu Zhang, Wenyue Hua, Haolun Wu, Zhihan Guo, Yufei Wang, Niklas Muennighoff, et al. A survey on test-time scaling in large language models: What, how, where, and how well?arXiv preprint arXiv:2503.24235,

    arXiv

  4. [16]

    Reinforcing general reasoning without verifiers

    [ZLS+25b] Xiangxin Zhou, Zichen Liu, Anya Sims, Haonan Wang, Tianyu Pang, Chongxuan Li, Liang Wang, Min Lin, and Chao Du. Reinforcing general reasoning without verifiers. arXiv preprint arXiv:2505.21493,

    arXiv

  5. [2022]

    Rae, Oriol Vinyals, and L

    [HBM+22b] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, ...

    arXiv

  6. [2023]

    SuperGPQA: Scaling LLM evaluation across 285 graduate disciplines.arXiv preprint arXiv:2502.14739,

    [DYM+25] Xinrun Du, Yifan Yao, Kaijing Ma, Bingli Wang, Tianyu Zheng, King Zhu, Minghao Liu, Yiming Liang, Xiaolong Jin, Zhenlin Wei, et al. SuperGPQA: Scaling LLM evaluation across 285 graduate disciplines.arXiv preprint arXiv:2502.14739,

    arXiv

  7. [2024]

    Deepseek-r1: Incentivizing rea- soning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    [GYZ+25] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing rea- soning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    arXiv

  8. [2025]

    Omni-MATH: A universal Olympiad level mathematic benchmark for large language models.ArXiv, abs/2410.07985,

    [GSY+24] Bofei Gao, Feifan Song, Zhe Yang, Zefan Cai, Yibo Miao, Qingxiu Dong, Lei Li, Chenghao Ma, Liang Chen, Runxin Xu, Zhengyang Tang, Benyou Wang, Daoguang Zan, Shanghaoran Quan, Ge Zhang, Lei Sha, Yichang Zhang, Xuancheng Ren, Tianyu Liu, and Baobao Chang. Omni-MATH: A u...

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.