Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

PATS: Process-Level Adaptive Thinking Mode Switching

T0 review · 4 major / 4 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 a large language model can switch between simple, medium, and complex reasoning at every step, using the process reward model's score as a difficulty signal, and that this recovers almost all the accuracy of…

desk verdict Useful step-level adaptive beam-width heuristic; the headline accuracy-parity claim is statistically fragile but the method is clean and worth refereeing. read the letter →

arxiv 2505.19250 v1 pith:WWGG575B submitted 2025-05-25 cs.CL

Yi Wang , Junxiao Liu , Shimao Zhang , Jiajun Chen , Shujian Huang This is my paper
classification cs.CL
keywords LLMreasoningprocessrewardmodelbeamsearchadaptiveinferencefast-slowthinkingtest-timescalingmathematical
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 proposes PATS, a training-free method that lets a large language model change its reasoning strategy at every step rather than committing to one strategy for a whole solution. The model runs a PRM-guided beam search and uses the reward score of the chosen step to decide whether the next step should be 'simple' (2 candidates), 'medium' (4), or 'complex' (8). The central claim is that this process-level, difficulty-aware switching recovers nearly all the accuracy of always using the complex mode while spending a little over half the tokens. Across five math benchmarks with a 7B policy model, PATS averages 61.3% accuracy versus 61.6% for All-complex, using 55.4% of its output tokens, and beats solution-level switching by 4.4 accuracy points with about 7% fewer tokens. The authors argue this is evidence that reasoning difficulty varies within a solution and that resource allocation should adapt at the step level.

What carries the argument

The load-bearing mechanism is the thinking-mode ladder defined by beam-search width: simple (2 candidates per step), medium (4), and complex (8). At each step, after the best candidate is selected, its PRM score is the difficulty sensor. The switching policy is a set of thresholds: score ≥0.85 steps the mode down one rung; score <0.75 jumps to complex; score <0.4 penalizes the step and retries it once in complex mode. Progressive down-switching smooths mode changes, while the one-time penalty limits error propagation without allowing infinite retries. The argument that this works rests on the hypothesis that a higher PRM score at the current step indicates an easier next step.

What would settle it

Run PATS on a benchmark with step-level difficulty annotations and compare the PRM score of each selected step against the actual next-step error rate; if the correlation is near zero, the mode-switching policy should be no better than a random switch with the same mode distribution, and PATS's accuracy-token advantage should disappear.

Watch

Extended reading notes

Core claim

PATS establishes that the accuracy-efficiency tradeoff of test-time computation in LLM reasoning can be improved by switching thinking modes at the process level, i.e., at each reasoning step, rather than at the solution level. The method defines three thinking modes by beam-search width (simple=2, medium=4, complex=8 candidates per step). After selecting the highest-scoring candidate at a step, the PRM score of that candidate is read as a difficulty estimate: if it is high (≥0.85) the mode steps down; if low (<0.75) it jumps to complex; if very low (<0.4) the step is penalized and retried once in complex mode. The paper reports that this policy, starting in complex mode, matches the accuracy of running complex mode everywhere on five math benchmarks (61.3% vs 61.6%) while using 55.4% of the tokens, and that it outperforms coarse-grained solution-level switching by 4.4 accuracy points with about 7% fewer tokens. The authors also report that the benefit holds across policy models of 1.5B, 3B, and 7B parameters and across three PRMs, with the best relative gain when the PRM is weak.

Load-bearing premise

The entire switching policy assumes the PRM score of the chosen step is a dependable real-time measure of how hard the next reasoning step will be, so that fixed thresholds (0.85, 0.75, 0.4) trigger the right mode changes.

Editorial extensions

If this is right

  • PATS recovers almost all of the accuracy of always-complex reasoning while spending 55.4% of its output tokens, across GSM8K, MATH500, AMC23, MinervaMATH, and AIME24 with Qwen2.5-7B-Instruct.
  • Process-level switching beats solution-level switching: +4.4 average accuracy points and about 7% fewer tokens than Solution-verification Switch, showing that waiting for a full solution to decide is too coarse.
  • The one-time bad-step penalty is load-bearing: removing it drops average accuracy by 5.4 points, and making it infinite raises tokens by roughly half with lower accuracy.
  • The benefit generalizes across policy model scales (1.5B, 3B, 7B) and across three PRMs, with PATS beating All-complex in accuracy when the PRM is Math-Shepherd.
  • Harder tasks (AMC23) show a higher proportion of complex modes in mid-to-late reasoning stages than easier tasks (MATH500), which the paper reads as evidence that the mode distribution tracks task difficulty.

Reading between the lines

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

  • Beyond the paper: if PRM scores are a reliable step-difficulty oracle, the same switching policy could be applied to other process-level search methods such as MCTS or speculative decoding, where a per-step confidence signal is available.
  • Beyond the paper: the threshold values (0.85/0.75/0.4) are tuned to the Qwen2.5-Math-PRM-7B score distribution; a calibration step or score-normalization scheme would be needed before the method transfers to a new PRM without re-tuning.
  • Beyond the paper: the mode ladder (2/4/8) is a coarse discretization of width; a continuous width schedule derived from the same PRM scores might extract further token savings, though the paper does not test this.
  • Beyond the paper: the fact that PATS gains the most over baselines when the PRM is weak (Math-Shepherd) suggests the step-difficulty signal is most useful precisely when the reward model is noisy, since the switching policy can recover from bad steps; this is testable by degrading a strong PRM with noise and observing whether PATS's margin grows.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes PATS (Process-Level Adaptive Thinking Mode Switching), a training-free inference-time method that adaptively changes the reasoning strategy of an LLM at each step of PRM-guided beam search. Thinking modes are defined by the number of candidate steps generated per step (simple: 2, medium: 4, complex: 8), and switching is driven by the PRM score of the selected top candidate: scores above 0.85 downgrade the mode, scores below 0.75 upgrade it to complex, and scores below 0.4 trigger a one-time penalty that regenerates the step in complex mode. Experiments on five math benchmarks with Qwen2.5-7B-Instruct and Qwen2.5-Math-PRM-7B report an average accuracy of 61.3% with 2808 output tokens, compared with 61.6% accuracy and 5070.6 tokens for the fixed All-complex baseline, i.e. roughly 55.4% of the token budget. Additional experiments vary the policy model scale (1.5B, 3B, 7B) and the PRM (Math-Shepherd, Qwen2.5-Math-7B-PRM800K), and the paper claims "strong generalization" across these settings. The central claim is that process-level switching preserves accuracy while substantially reducing inference cost, and that this is superior to coarse-grained solution-level switching.

Significance. If the accuracy-parity result is statistically robust, the contribution is meaningful: the paper introduces a plausible and simple mechanism for step-level adaptive allocation of test-time compute, and it is one of few works that treat the switching decision at the granularity of individual reasoning steps. The paper is also honest about its scope, noting the limitation to 1.5B-7B policy models and the reliance on PRMs. The method is conceptually clean and the open-source code link is a practical strength. However, the current empirical support for the headline claim is not yet convincing because the reported differences in accuracy are comparable to the noise from one or two answers on the smallest benchmarks, and the switching thresholds are tuned without a reported sensitivity analysis. The contribution is therefore promising but requires substantially stronger empirical validation before the central claims can be accepted.

major comments (4)
  1. [Table 1 / §5.1] The central accuracy-parity claim is not statistically supported. On AIME24 there are only 30 problems, so one additional correct answer changes that dataset's accuracy by 3.3 points and the reported five-benchmark average by 0.67 points. The reported gap between PATS (61.3%) and All-complex (61.6%) is 0.3 points, i.e. within a single-question flip, and the 4.4-point advantage over Solution-verification Switch is also sensitive to a handful of answers on AIME24 and AMC23. Since generation uses temperature 0.6 and is therefore stochastic, the paper should report confidence intervals, standard errors, or at least multiple-seed runs. Without these, the claim that PATS "maintains high accuracy" while using 55.4% of the tokens cannot be distinguished from sampling noise.
  2. [§4.1 / §3.3] The three switching thresholds (value_good=0.85, value_low=0.75, value_bad=0.4) are set by "empirical configuration" with no sensitivity analysis and no hold-out validation. The entire behavior of PATS depends on these thresholds, and the paper gives no evidence that they are robust across PRMs, policy scales, or task difficulties. The authors should report accuracy and token usage for a small grid of threshold values (e.g. varying each threshold by ±0.05), and ideally select thresholds on a separate validation set rather than on the test benchmarks. As written, the main result could be an artifact of threshold overfitting.
  3. [§5.2 / Table 4] The claim of "strong generalization" across policy model scales is overstated. For Qwen2.5-1.5B-Instruct, PATS achieves 42.6% average accuracy versus 46.7% for All-complex, a 4.1-point shortfall, and on AIME24 PATS obtains only 3.3% accuracy, which is worse than All-simple (6.7%) and much worse than All-complex (10.0%). Thus the accuracy-parity result does not hold for the smallest policy model, and the generalization claim should be revised to acknowledge this degradation. The same issue affects the use of "consistently" in the text: Table 6 shows PATS below All-complex on three of five benchmarks for the 1.5B model.
  4. [§3.3] The paper's core hypothesis is that the PRM score of the top candidate is a reliable real-time indicator of step difficulty, so that the fixed thresholds correctly identify when to switch modes. The paper provides no direct evidence for this assumption, such as calibration plots of PRM scores versus empirical step difficulty or versus the probability of eventually reaching a correct answer. Since the switching policy is entirely driven by this signal, the authors should at least report a simple diagnostic (e.g. the distribution of PRM scores on correct versus incorrect trajectories, or the effect of replacing the PRM with a noisier proxy) to justify the mechanism.
minor comments (4)
  1. [§1] There is a grammatical error in the introduction: "we performs finer-grained" should be "we perform finer-grained".
  2. [Table 5] "Math-Shpherd" is a typo for "Math-Shepherd".
  3. [§3.1] The text refers to "best-first search (BFS) with K=1" but the method is elsewhere called beam search; the relationship between the two is not defined, and the notation K is never introduced.
  4. [Limitations] The Limitations section acknowledges the restriction to small policy models and the reliance on a PRM, but it does not mention the lack of variance estimation or the threshold-selection issue, both of which are more pressing threats to the reported results.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PATS's accuracy-efficiency result is an empirical system comparison, not a derivation that reduces to its inputs.

full rationale

The central claim—PATS reaches 61.3% average accuracy vs 61.6% for All-complex while using 55.4% of its tokens—is an experimental outcome, not a definitional consequence. The switching rule in Section 3.3 maps the PRM score of the selected step to beam widths (2/4/8) via thresholds value_good=0.85, value_low=0.75, value_bad=0.4; the rule does not encode the target accuracy or token budget, and accuracy is measured on external benchmarks with the same PRM used for the fixed-width baselines. The paper explicitly labels the PRM-score/difficulty link as a hypothesis ('We hypothesize that a higher PRM score at the current step indicates a lower reasoning difficulty'), and the thresholds are described as 'empirical configuration' rather than fitted to reproduce the reported numbers. Self-citations appear only in supporting roles (e.g., Zhang et al. 2024 for temperature 0.6; She et al. 2025 and Zhang et al. 2025 among PRM-guided-search references) and are not load-bearing; the method and evaluation do not depend on an unverified same-author result. The Limitations section acknowledges the small model scales and reliance on PRMs, which are scope caveats rather than circular reasoning. The main residual concern is statistical fragility of the five-benchmark averages (especially AIME24/AMC23) and the absence of a held-out threshold selection, but that is a correctness/robustness issue, not circularity.

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

The method introduces no new entities. It relies on two free-parameter families: the three reward thresholds and the candidate-width discretization, both set empirically without sensitivity analysis. The key domain assumption is that PRM scores serve as reliable step-level difficulty signals. The generalization results suggest this assumption breaks down for smaller policy models.

free parameters (5)
  • value_good threshold = 0.85
    Threshold above which the mode switches to a simpler one; set in Section 4.1 by 'empirical configuration' with no sensitivity analysis.
  • value_low threshold = 0.75
    Threshold below which the mode switches to complex; set in Section 4.1 empirically.
  • value_bad threshold = 0.4
    Threshold below which the step is penalized and regenerated once in complex mode; set in Section 4.1 empirically.
  • mode candidate counts = 2, 4, 8
    The discretization of thinking modes into simple, medium, and complex with widths 2, 4, 8 is chosen by hand in Section 3.2; no justification that these widths are optimal.
  • temperature = 0.6
    Sampling temperature used for candidate generation, set in Section 4.1 following prior work; no sensitivity reported.
assumptions (3)
  • domain assumption PRM score of the selected step is a reliable indicator of step difficulty and correctness.
    Stated as a hypothesis in Section 3.3 and used to drive all mode switches. If PRM scores are miscalibrated, the switching policy fails, as appears to happen for the 1.5B model.
  • standard math Generating more candidates at a step improves the chance of selecting a correct next step.
    This is the standard rationale for PRM-guided beam search, cited from prior work in Section 2.2; the paper does not re-derive it.
  • ad hoc to paper The three-level discretization (2, 4, 8) captures meaningful differences in reasoning effort.
    The widths are chosen without an optimization or theoretical motivation in Section 3.2; results may depend on this specific choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PATS: Process-Level Adaptive Thinking Mode Switching." pith.science (2026). https://pith.science/paper/WWGG575B

@misc{pith2026250519250,
  author       = {Pith},
  title        = {Pith review of: PATS: Process-Level Adaptive Thinking Mode Switching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WWGG575B}},
  note         = {Machine review of arXiv:2505.19250}
}
read the original abstract

Current large-language models (LLMs) typically adopt a fixed reasoning strategy, either simple or complex, for all questions, regardless of their difficulty. This neglect of variation in task and reasoning process complexity leads to an imbalance between performance and efficiency. Existing methods attempt to implement training-free fast-slow thinking system switching to handle problems of varying difficulty, but are limited by coarse-grained solution-level strategy adjustments. To address this issue, we propose a novel reasoning paradigm: Process-Level Adaptive Thinking Mode Switching (PATS), which enables LLMs to dynamically adjust their reasoning strategy based on the difficulty of each step, optimizing the balance between accuracy and computational efficiency. Our approach integrates Process Reward Models (PRMs) with Beam Search, incorporating progressive mode switching and bad-step penalty mechanisms. Experiments on diverse mathematical benchmarks demonstrate that our methodology achieves high accuracy while maintaining moderate token usage. This study emphasizes the significance of process-level, difficulty-aware reasoning strategy adaptation, offering valuable insights into efficient inference for LLMs.

Figures

Figures reproduced from arXiv: 2505.19250 by the authors.

Figure 1
Figure 1. Illustration of our Process-Level Adaptive Thinking Mode Switching (PATS) paradigm. For math reasoning tasks, the model performs PRM-guided beam search, where the number of candidate steps (2/4/8) serves as a proxy for thinking mode complexity. At each step, the policy model selects the top-scoring candidate , infers reasoning difficulty, and dynamically switches to the corresponding mode. The framework further inco… view at source ↗
Figure 2
Figure 2. Comparison of thinking mode distributions over the reasoning process for MATH500 (easier) and AMC23 [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

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. Hierarchical Budget Policy Optimization for Adaptive Reasoning

    cs.AI 2025-07 conditional novelty 6.0 of 10

    Training reasoning models with hierarchical token budgets and budget-aware rewards produces up to 60.6% token reduction with no accuracy loss, and benchmark-level evidence of length adaptation.

Reference graph

Works this paper leans on

33 extracted references · 3 canonical work pages · cited by 1 Pith paper

  1. [1]

    Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, and 1 others. 2024. Do not think that much for 2+ 3=? on the overthinking of o1-like llms. arXiv preprint arXiv:2412.21187

    arXiv 2024

  2. [2]

    Xiaoxue Cheng, Junyi Li, Wayne Xin Zhao, and Ji-Rong Wen. 2025. Think more, hallucinate less: Mitigating hallucinations via dual process of fast and slow thinking. arXiv preprint arXiv:2501.01306

    arXiv 2025

  3. [3]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, and 1 others. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168

    arXiv 2021

  4. [4]

    Yichao Fu, Junda Chen, Siqi Zhu, Zheyu Fu, Zhongdongming Dai, Aurick Qiao, and Hao Zhang. 2024. Efficiently serving llm reasoning programs with certaindex. arXiv preprint arXiv:2412.20993

    arXiv 2024

  5. [5]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    arXiv 2025

  6. [6]

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720

    arXiv 2024

  7. [7]

    Daniel Kahneman. 2011. Thinking, fast and slow. macmillan

    2011

  8. [8]

    Jikun Kang, Xin Zhe Li, Xi Chen, Amirreza Kazemi, Qianyi Sun, Boxing Chen, Dong Li, Xu He, Quan He, Feng Wen, and 1 others. 2024. Mindstar: Enhancing math reasoning in pre-trained llms at inference time. arXiv preprint arXiv:2405.16265

    arXiv 2024

Show all 33 references
  1. [9]

    Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, and 1 others. 2022. Solving quantitative reasoning problems with language models. Advances in Neural Information Processin...

    2022

  2. [10]

    Zhong-Zhi Li, Duzhen Zhang, Ming-Liang Zhang, Jiaxin Zhang, Zengyan Liu, Yuxuan Yao, Haotian Xu, Junhao Zheng, Pei-Jie Wang, Xiuyi Chen, Yingying Zhang, Fei Yin, Jiahua Dong, Zhiwei Li, Bao-Long Bi, Ling-Rui Mei, Junfeng Fang, Zhijiang Guo, Le Song, and Cheng-Lin Liu. 2025. ht...

    2025 arXiv

  3. [11]

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let's verify step by step. In The Twelfth International Conference on Learning Representations

    2023

  4. [12]

    Xinyin Ma, Guangnian Wan, Runpeng Yu, Gongfan Fang, and Xinchao Wang. 2025. Cot-valve: Length-compressible chain-of-thought tuning. arXiv preprint arXiv:2502.09601

    2025 arXiv

  5. [13]

    OpenAI . 2024. https://openai.com/index/learning-to-reason-with-llms/ Learning to reason with LLMs . Accessed: 2025-04-29

    2024

  6. [14]

    Swarnadeep Saha, Archiki Prasad, Justin Chih-Yao Chen, Peter Hase, Elias Stengel-Eskin, and Mohit Bansal. 2024. System-1. x: Learning to balance fast and slow planning with language models. arXiv preprint arXiv:2407.14414

    2024 arXiv

  7. [15]

    Shuaijie She, Junxiao Liu, Yifeng Liu, Jiajun Chen, Xin Huang, and Shujian Huang. 2025. R-prm: Reasoning-driven process reward modeling. arXiv preprint arXiv:2503.21295

    2025 arXiv

  8. [16]

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling llm test-time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314

    2024 arXiv

  9. [17]

    DiJia Su, Sainbayar Sukhbaatar, Michael Rabbat, Yuandong Tian, and Qinqing Zheng. 2024. Dualformer: Controllable fast and slow thinking by learning with randomized reasoning traces. In The Thirteenth International Conference on Learning Representations

    2024

  10. [18]

    Hanshi Sun, Momin Haider, Ruiqi Zhang, Huitao Yang, Jiahao Qiu, Ming Yin, Mengdi Wang, Peter Bartlett, and Andrea Zanette. 2024. Fast best-of-n decoding via speculative rejection. arXiv preprint arXiv:2410.20290

    2024 arXiv

  11. [19]

    Yiyou Sun, Georgia Zhou, Hao Wang, Dacheng Li, Nouha Dziri, and Dawn Song. 2025. Climbing the ladder of reasoning: What llms can-and still can't-solve after sft? arXiv preprint arXiv:2504.11741

    2025

  12. [20]

    Jonathan Uesato, Nate Kushman, Ramana Kumar, Francis Song, Noah Siegel, Lisa Wang, Antonia Creswell, Geoffrey Irving, and Irina Higgins. 2022. Solving math word problems with process-and outcome-based feedback. arXiv preprint arXiv:2211.14275

    2022 arXiv

  13. [21]

    Ante Wang, Linfeng Song, Ye Tian, Baolin Peng, Dian Yu, Haitao Mi, Jinsong Su, and Dong Yu. 2024. Litesearch: Efficacious tree search for llm. arXiv preprint arXiv:2407.00320

    2024 arXiv

  14. [22]

    Ante Wang, Linfeng Song, Ye Tian, Dian Yu, Haitao Mi, Xiangyu Duan, Zhaopeng Tu, Jinsong Su, and Dong Yu. 2025. Don't get lost in the trees: Streamlining llm reasoning by overcoming tree search exploration pitfalls. arXiv preprint arXiv:2502.11183

    2025 arXiv

  15. [23]

    Peiyi Wang, Lei Li, Zhihong Shao, RX Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. 2023. Math-shepherd: Verify and reinforce llms step-by-step without human annotations. arXiv preprint arXiv:2312.08935

    2023 arXiv

  16. [24]

    Peter C Wason and J St BT Evans. 1974. Dual processes in reasoning? Cognition, 3(2):141--154

    1974

  17. [25]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824--24837

    2022

  18. [26]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, and 1 others. 2024 a . Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115

    2024 arXiv

  19. [27]

    An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, and 1 others. 2024 b . Qwen2. 5-math technical report: Toward mathematical expert model via self-improvement. arXiv preprint arXiv:2409.12122

    2024 arXiv

  20. [28]

    Wenlin Yao, Haitao Mi, and Dong Yu. 2024. Hdflow: Enhancing llm complex problem-solving with hybrid thinking and dynamic workflows. arXiv preprint arXiv:2409.17433

    2024 arXiv

  21. [29]

    Shimao Zhang, Yu Bao, and Shujian Huang. 2024. Edt: Improving large language models' generation by entropy-based dynamic temperature sampling. arXiv preprint arXiv:2403.14541

    2024 arXiv

  22. [30]

    Shimao Zhang, Xiao Liu, Xin Zhang, Junxiao Liu, Zheheng Luo, Shujian Huang, and Yeyun Gong. 2025. Process-based self-rewarding language models. arXiv preprint arXiv:2503.03746

    2025 arXiv

  23. [31]

    Chujie Zheng, Zhenru Zhang, Beichen Zhang, Runji Lin, Keming Lu, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. 2024. Processbench: Identifying process errors in mathematical reasoning. arXiv preprint arXiv:2412.06559

    2024 arXiv

  24. [32]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  25. [33]

    write newline

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

Pith tools

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