Pith. sign in

REVIEW 5 major objections 4 minor 4 cited by

The Majority is not always right: RL training for solution aggregation

T0 review · 5 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Majority voting can be wrong; this paper trains an LLM to reconcile candidate solutions instead, and shows gains on four math competitions.

desk verdict The RL-trained aggregator idea is solid and the easy/hard mixture finding is real, but the evaluation has a load-bearing gap: no contamination check between DeepScaler and the four MathArena test sets, on which the reported margins are 1–2 questions. read the letter →

arxiv 2509.06870 v1 pith:JUNNZY3E submitted 2025-09-08 cs.CL

Wenting Zhao , Pranjal Aggarwal , Swarnadeep Saha , Asli Celikyilmaz , Jason Weston , Ilia Kulikov This is my paper · ORCID
classification cs.CL
keywords solutionaggregationreinforcementlearningfromverifiablerewardsmajorityvotingmathreasoningbenchmarksGRPOtest-timecomputescalingLLMAIME/HMMT
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

Majority voting throws away correct answers whenever they are in the minority. This paper tries to fix that by training a small model, AggLM-1.7B, to treat aggregation as a reasoning task: given a problem and eight candidate solutions, it reviews, corrects, and combines them into a final answer, trained by reinforcement learning with a binary reward for matching the known correct answer. The key training choice is a balanced mix of 'hard' solution sets, where the majority answer is wrong, and 'easy' sets, where the majority is right. On four math-competition benchmarks the trained aggregator outperforms majority voting, prompted aggregation, and reward-model selection, and it keeps winning when the candidate solutions come from a stronger or non-thinking generator. A careful reader would care because this points to learned aggregation as a cheaper alternative to scaling the number of samples or the model size.

What carries the argument

The load-bearing object is AggLM, a generative aggregator trained with Group-relative Policy Optimization (GRPO) under a verifiable binary reward, namely whether the final answer matches the ground truth. Its inputs are the problem plus m candidate solutions; its job is to select a correct candidate when one exists and to synthesize a new correct solution when none does. The crucial data mechanism is the hard/easy split: a solution set is hard when the majority answer is wrong, and the training mixture keeps all hard sets while adding a moderate fraction of easy sets; the paper shows this balance is necessary, with gains concentrated at small majority-answer size.

What would settle it

Run an n-gram or embedding-based near-duplicate search between the roughly 40k training problems and the 120 evaluation problems, and re-evaluate AggLM on a fresh set of unseen math-competition problems; if training overlaps the test problems or fresh-set gains vanish, the aggregation claim is not as general.

Watch

Extended reading notes

Core claim

The paper's central claim is that aggregation can be learned as an explicit reasoning skill, instead of being left to a fixed heuristic. Concretely, it constructs training examples by sampling 128 solutions per problem from Qwen3-1.7B, grouping them into 16 sets of 8, labeling a set hard if its majority answer is wrong, and training the aggregator with GRPO on all hard sets plus 5-50% easy sets. The resulting AggLM-1.7B is claimed to outperform majority voting, best-of-N and weighted-majority with AceMath 7B/72B reward models, and prompt-only aggregation on AIME24, AIME25, HMMT24, and HMMT25; to transfer to solutions sampled from Qwen3-8B and from non-thinking mode; and to use roughly a thir

Load-bearing premise

The paper assumes the DeepScaler training corpus does not contain the AIME24/25 or HMMT24/25 evaluation problems; it reports no contamination check, and if those problems appeared during RL training, the reported gains would be inflated.

Editorial extensions

If this is right

  • Aggregating eight solutions with the learned aggregator can beat majority voting over sixteen solutions, so aggregation offers a token-efficient route to higher accuracy.
  • The trained aggregator transfers to candidate solutions from a stronger model (Qwen3-8B) and from non-thinking mode, suggesting the skill is not tied to the training generator's distribution.
  • Fine-tuning the solution model on the same data does not close the gap, so the improvement comes from learning to aggregate, not just from seeing more data.
  • A single model can be multitasked to both solve and aggregate with near-matching performance, so aggregation could be folded into normal post-training.
  • Reward-model selection often underperforms majority voting in this setup, whereas the RL-trained aggregator reliably improves on it.

Reading between the lines

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

  • If the training data is clean, this suggests a new scaling axis for reasoning: invest in a trained synthesizer over parallel samples rather than only in more samples or larger generators.
  • The hard/easy mixture result points toward active or curriculum balancing for other reinforcement-learning-with-verifiable-rewards tasks, where rewards are sparse and the model must learn when to trust the majority.
  • Because AggLM also helps on weak non-thinking solutions, it may offer a way to amplify cheap base models without requiring expensive long chain-of-thought at test time.
  • The token-efficiency result implies that for a fixed inference budget, a learned aggregator might dominate naive sample scaling, though this depends on the cost of the aggregator's own reasoning.
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 / 4 minor

Summary. The paper introduces AggLM, a 1.7B aggregation model trained with GRPO on a dataset derived from DeepScaler and Qwen3-1.7B solutions. Given a math problem and m=8 candidate solutions, AggLM is trained to output a final solution; training examples are labeled easy/hard depending on whether the majority solution is correct, and a balanced mixture (50% easy relative to hard) is used. The main experiments aggregate solutions from Qwen3-1.7B thinking mode, Qwen3-8B thinking mode, and Qwen3-1.7B non-thinking mode, evaluated on AIME24, AIME25, HMMT24, HMMT25. The paper reports consistent gains over majority voting, AceMath reward-model selection, and prompted aggregation; additional experiments analyze the easy/hard ablation, number of solution sets, scaling with candidate count, multitask training, and token usage.

Significance. Learned solution aggregation is a timely and plausible test-time scaling approach. The paper's key strengths are (1) isolating RL-trained aggregation from prompted aggregation, (2) showing generalization to a stronger solution model, and (3) the easy/hard mixture ablation. If the empirical claims hold, the method is practically useful. However, the current empirical case is under-supported: the training data could overlap the evaluation sets, the easy/hard mixture selection appears to use evaluation-set feedback, and the 30-problem benchmarks lack uncertainty estimates. These issues are fixable and do not invalidate the proposed method, but they need to be addressed for the central comparison to be trustworthy.

major comments (5)
  1. [§4.1–§4.2, Tables 1–3] No contamination check is reported between the DeepScaler training set and the MathArena evaluation problems (AIME24/25, HMMT24/25). Since AggLM receives the full problem statement during RL training and receives a verifiable reward for the final answer, even a small number of memorized problems could inflate the results. The margins in Table 1 are roughly 3–7 points over prompted aggregation and 3–5 points over majority voting, i.e., about 1–2 questions on a 30-problem set. MathArena's 'uncontaminated' status applies to pretraining corpora, not to a subsequent RL fine-tuning corpus. The authors should report a decontamination analysis (e.g., exact or fuzzy overlap of problem statements), and either exclude overlapping problems from evaluation or add a separately curated held-out benchmark.
  2. [§6, Table 4] The easy/hard mixture percentage p is selected using the same four evaluation benchmarks on which the main results are reported. Table 4 is an ablation of Easy% on AIME24/25 and HMMT24/25, and the paper then uses 50% for the main experiments. If this value was chosen after observing these results, the main tables are optimistic. Moreover, at 0% easy the aggregator does not consistently beat majority voting (AIME24: 64.22 vs 67.92; HMMT24: 27.80 vs 29.01), so the claim that a balanced mixture is critical depends on this choice. The authors should pre-register a p value or select it on a separate validation split that is not part of the reported evaluation.
  3. [Tables 1–3] Each evaluation dataset contains 30 problems, and most of the reported differences between AggLM and the baselines are only a few points, i.e., roughly one or two problems. The paper reports no confidence intervals, bootstrap estimates, or paired significance tests. This makes it difficult to assess whether the 'consistent outperformance' is beyond sampling noise, especially for the smaller-margin comparisons such as HMMT25 in Tables 1 and 3. The authors should provide per-problem scores and standard errors or bootstrap intervals, and where possible use paired tests across the same solution sets.
  4. [§6, Table 7] The token-efficiency claim is based on output tokens per generation for solution vs aggregator models. This ignores the input cost of the candidate solutions and the number of aggregator rollouts used in the protocol. A fair comparison should report total inference tokens per final answer, including (i) generating the m=8 candidate solutions, (ii) the concatenated-solution prompt consumed by the aggregator, and (iii) the four aggregation samples per solution set used in the main protocol. With those costs included, the conclusion may still hold, but the current evidence is incomplete.
  5. [§6, Table 6] The 'Is aggregation or extra data responsible for gains?' experiment trains a solution model on the same DeepScaler data, but no training details are given (SFT vs RL, number of steps, compute budget). If the solution model is only fine-tuned with SFT, the comparison does not isolate the effect of aggregation versus additional data, because the aggregator is trained with RLVR on verifiable rewards. The authors should specify the exact training setup and ideally train the solution model with the same RLVR objective and comparable compute before concluding that gains are not due to extra data.
minor comments (4)
  1. [§3, Figure 3] The number of candidate solutions is denoted m in the formulation but k in Figure 3 and in the scaling analysis. Define both symbols explicitly and use them consistently.
  2. [§4.2] The phrase 'Pass@1 for each set is computed as the success rate over these four answers' conflates pass@1 with an average over four aggregator samples. I would call this 'estimated per-sample accuracy' or define it as the mean over four independent aggregator rollouts.
  3. [§4.1, References] DeepScaler is cited to a Notion blog post rather than an archival dataset or technical report. Please provide a stable reference, version, and download link, since the contamination analysis and reproducibility depend on knowing exactly which problem statements appear in the training data.
  4. [Figure 2] The caption reads 'train by RL'; this should be 'trained by RL.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are driven by external training/evaluation comparisons and ablations, not by equations that reduce to their own inputs.

full rationale

The paper makes an empirical claim: an RL-trained aggregator (AggLM-1.7B) outperforms majority voting, reward-model selection, and prompted aggregation on four MathArena datasets. The core comparison is not circular: AggLM is trained on DeepScaler problems with ground-truth answers and evaluated on held-out MathArena problems, and the reported gains are measured against baselines on the same solution sets. The only tunable 'knob' is the easy/hard mixture percentage p (Section 4.1, Table 4), and it is presented as an ablation, not as a fitted parameter that is then renamed a prediction. Table 6 explicitly tests whether extra data, rather than aggregation, explains the gains by training the solution model on the same DeepScaler data and showing it does not close the gap. The model's behavior is evaluated against external math benchmarks and verifiable rewards, so there is no self-definitional reduction. The paper does cite prior work that shares an author (Stahlberg et al. 2022; Welleck et al. 2024), but those citations are used for motivation or for defining standard baselines, not as load-bearing justifications for the central claim. The skeptical concern about possible DeepScaler/MathArena contamination is a legitimate external-validity risk, but it is not a circularity: contamination would be an alternative explanation for the empirical results, not a case of the derivation being equivalent to its inputs by construction. No uniqueness theorem, ansatz-smuggling citation, or fitted-input-called-prediction pattern is present. The derivation chain is self-contained in the sense required by this analysis; correctness risks such as data leakage are outside the circularity score.

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

The method introduces no new theoretical entities. The main 'free' choices are hyperparameters of the RL training pipeline, most importantly the easy/hard mixture ratio, which was tuned on the evaluation benchmarks.

free parameters (3)
  • easy mixture percentage p = 50%
    The proportion of easy examples in the training mixture was chosen based on ablation results reported on the evaluation benchmarks (Table 4, Section 6).
  • number of candidate solutions per set m = 8
    Training sets of 8 solutions; the paper shows generalization to other values of k, so this is a design choice, not a fitted parameter.
  • number of solution sets per problem s = 16
    Diversity parameter; ablations show modest gains as s grows (Table 5).
assumptions (5)
  • domain assumption RLVR with a binary exact-match reward is a sufficient learning signal for aggregation.
    The method assumes that optimizing the probability of matching the ground-truth answer trains a useful aggregator policy.
  • domain assumption math_verify correctly determines whether two solutions are equivalent.
    The reward and evaluation both rely on `math_verify` for answer matching; errors here would corrupt training and metrics.
  • domain assumption The solution model's samples are independent and cover the true answer distribution.
    The aggregation protocol draws independent samples from p_theta and assumes the set is representative enough to train an aggregator.
  • domain assumption DeepScaler and MathArena evaluation datasets do not overlap.
    No contamination check is reported; if the aggregator saw AIME/HMMT problems during RL training, the results are inflated. This is the load-bearing assumption.
  • standard math GRPO loss gradient estimates are valid for policy optimization.
    The paper relies on the standard GRPO update, assuming its estimator is unbiased and appropriate for this RL setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Majority is not always right: RL training for solution aggregation." pith.science (2026). https://pith.science/paper/JUNNZY3E

@misc{pith2026250906870,
  author       = {Pith},
  title        = {Pith review of: The Majority is not always right: RL training for solution aggregation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JUNNZY3E}},
  note         = {Machine review of arXiv:2509.06870}
}
read the original abstract

Scaling up test-time compute, by generating multiple independent solutions and selecting or aggregating among them, has become a central paradigm for improving large language models (LLMs) on challenging reasoning tasks. While most prior work relies on simple majority voting or reward model ranking to aggregate solutions, these approaches may only yield limited benefits. In this work, we propose to learn aggregation as an explicit reasoning skill: given a set of candidate solutions, we train an aggregator model to review, reconcile, and synthesize a final, correct answer using reinforcement learning from verifiable rewards. A key ingredient is careful balancing of easy and hard training examples, allowing the model to learn both to recover minority-but-correct answers as well as easy majority-correct answers. Empirically, we find our method, AggLM, outperforms both strong rule-based and reward-model baselines, across multiple benchmarks. Furthermore, it generalizes effectively to solutions from differing models, including stronger ones than contained in the training data, all while requiring substantially fewer tokens than majority voting with larger numbers of solutions.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

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

  1. When Many Answers Are Valid, Voting Fails: Symbolic Verification for Best-of-K Causal Reasoning in LLMs

    cs.AI 2026-08 conditional novelty 7.0 of 10

    CALVER is a training-free symbolic verifier that selects the best-of-K LLM causal answer by checking each trace against graph-validity criteria, outperforming plurality when multiple valid answers exist.

  2. Autodata: An agentic data scientist to create high quality synthetic data

    cs.AI 2026-06 unverdicted novelty 6.0 of 10

    Autodata trains meta-optimized AI agents to generate superior synthetic datasets, yielding performance gains over classical methods on CS research, legal, and math reasoning tasks.

  3. Efficient Reasoning on the Edge

    cs.LG 2026-03 accept novelty 5.5 of 10

    LoRA adapters, budget-forced GRPO, dynamic switching, parallel verification and FPTQuant enable practical chain-of-thought reasoning on quantized Qwen2.5-7B for edge devices.

  4. Demystifying Multi-Agent Debate: The Role of Confidence and Diversity

    cs.CL 2026-01 conditional novelty 5.0 of 10

    Starting debates with diverse answers and having agents share calibrated confidence turns multi-agent LLM debate from a no-gain martingale into a process that drifts toward correct answers and beats majority vote.

Reference graph

Works this paper leans on

27 extracted references · 7 canonical work pages · cited by 4 Pith papers

  1. [1]

    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 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    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 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  3. [3]

    Let ' s sample step by step: Adaptive-consistency for efficient reasoning and coding with LLM s

    Pranjal Aggarwal, Aman Madaan, Yiming Yang, and Mausam. Let ' s sample step by step: Adaptive-consistency for efficient reasoning and coding with LLM s. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp.\ 12375--12396, Singapore, December 2023. Association for Co...

    doi:10.18653/v1/2023.emnlp-main.761 2023

  4. [4]

    Matharena: Evaluating llms on uncontaminated math competitions

    Mislav Balunovi \'c , Jasper Dekoninck, Ivo Petrov, Nikola Jovanovi \'c , and Martin Vechev. Matharena: Evaluating llms on uncontaminated math competitions. arXiv preprint arXiv:2505.23281, 2025

    arXiv 2025

  5. [5]

    Large language monkeys: Scaling inference compute with repeated sampling

    Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V Le, Christopher R \'e , and Azalia Mirhoseini. Large language monkeys: Scaling inference compute with repeated sampling. arXiv preprint arXiv:2407.21787, 2024

    arXiv 2024

  6. [6]

    Universal self-consistency for large language models

    Xinyun Chen, Renat Aksitov, Uri Alon, Jie Ren, Kefan Xiao, Pengcheng Yin, Sushant Prakash, Charles Sutton, Xuezhi Wang, and Denny Zhou. Universal self-consistency for large language models. In ICML 2024 Workshop on In-Context Learning, 2024. URL https://openreview.net/forum?id=LjsjHF7nAN

    work page 2024

  7. [7]

    Deep think with confidence

    Yichao Fu, Xuewei Wang, Yuandong Tian, and Jiawei Zhao. Deep think with confidence. arXiv preprint arXiv:2508.15260, 2025

    arXiv 2025

  8. [8]

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

    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 reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    arXiv 2025

Show all 27 references
  1. [9]

    Mirror-consistency: Harnessing inconsistency in majority voting

    Siyuan Huang, Zhiyuan Ma, Jintao Du, Changhua Meng, Weiqiang Wang, and Zhouhan Lin. Mirror-consistency: Harnessing inconsistency in majority voting. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (eds.), Findings of the Association for Computational Linguistics: EMNLP 20...

    2024 doi

  2. [10]

    Openai o1 system card

    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, 2024

    2024 arXiv

  3. [11]

    Enhancing language model reasoning via weighted reasoning in self-consistency

    Tim Knappe, Ryan Luo Li, Ayush Chauhan, Kaylee Chhua, Kevin Zhu, and Sean O'Brien. Enhancing language model reasoning via weighted reasoning in self-consistency. In The 4th Workshop on Mathematical Reasoning and AI at NeurIPS'24, 2024. URL https://openreview.net/forum?id=2w0CIzWlle

    2024

  4. [12]

    Acemath: Advancing frontier math reasoning with post-training and reward modeling

    Zihan Liu, Yang Chen, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Acemath: Advancing frontier math reasoning with post-training and reward modeling. arXiv preprint arXiv:2412.15084, 2024

    2024 arXiv

  5. [13]

    Tang, Manan Roongta, Colin Cai, Jeffrey Luo, Li Erran Li, Raluca Ada Popa, and Ion Stoica

    Michael Luo, Sijun Tan, Justin Wong, Xiaoxiang Shi, William Y. Tang, Manan Roongta, Colin Cai, Jeffrey Luo, Li Erran Li, Raluca Ada Popa, and Ion Stoica. Deepscaler: Surpassing o1-preview with a 1.5b model by scaling rl. https://pretty-radio-b75.notion.site/DeepScaleR-Surpassi...

    2025

  6. [14]

    Learning to reason across parallel samples for llm reasoning

    Jianing Qi, Xi Ye, Hao Tang, Zhigang Zhu, and Eunsol Choi. Learning to reason across parallel samples for llm reasoning. arXiv preprint arXiv:2506.09014, 2025

    2025

  7. [15]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  8. [16]

    Felix Stahlberg and Bill Byrne. On NMT search errors and model errors: Cat got your tongue? In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan (eds.), Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint C...

    2019 doi

  9. [17]

    Uncertainty determines the adequacy of the mode and the tractability of decoding in sequence-to-sequence models

    Felix Stahlberg, Ilia Kulikov, and Shankar Kumar. Uncertainty determines the adequacy of the mode and the tractability of decoding in sequence-to-sequence models. In Smaranda Muresan, Preslav Nakov, and Aline Villavicencio (eds.), Proceedings of the 60th Annual Meeting of the ...

    2022 doi

  10. [18]

    Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations, 2023. URL https://o...

    2023

  11. [19]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 0 24824--24837, 2022

    2022

  12. [20]

    From decoding to meta-generation: Inference-time algorithms for large language models

    Sean Welleck, Amanda Bertsch, Matthew Finlayson, Hailey Schoelkopf, Alex Xie, Graham Neubig, Ilia Kulikov, and Zaid Harchaoui. From decoding to meta-generation: Inference-time algorithms for large language models. Transactions on Machine Learning Research, 2024. ISSN 2835-8856...

    2024

  13. [21]

    Inference scaling laws: An empirical analysis of compute-optimal inference for LLM problem-solving

    Yangzhen Wu, Zhiqing Sun, Shanda Li, Sean Welleck, and Yiming Yang. Inference scaling laws: An empirical analysis of compute-optimal inference for LLM problem-solving. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/foru...

    2025

  14. [22]

    Dynamic voting for efficient reasoning in large language models

    Mingfeng Xue, Dayiheng Liu, Wenqiang Lei, Xingzhang Ren, Baosong Yang, Jun Xie, Yidan Zhang, Dezhong Peng, and Jiancheng Lv. Dynamic voting for efficient reasoning in large language models. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.), Findings of the Association for Co...

    2023 doi

  15. [23]

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

    2024 arXiv

  16. [24]

    Qwen3 technical report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025

    2025 arXiv

  17. [25]

    @esa (Ref

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

  18. [26]

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

  19. [27]

    MIdaId: b5VȮBd)G̶ Rૉ,l

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

    1976

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.