Pith. sign in

REVIEW 4 major objections 4 minor 39 references

Verbosity-Aware Rationale Reduction: Effective Reduction of Redundant Rationale via Principled Criteria

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that fine-tuning with verbosity-based sentence pruning yields both higher accuracy and shorter rationales than training on complete chain-of-thought.

desk verdict Genuinely new sentence-level rationale pruning with a likelihood criterion, but the headline gains are not backed by the reported tables or variance evidence. read the letter →

arxiv 2412.21006 v3 pith:Z724SXFY submitted 2024-12-30 cs.CL cs.AI

Joonwon Jang , Jaehee Kim , Wonbin Kweon , Seonghyeon Lee , Hwanjo Yu This is my paper · ORCID
classification cs.CLcs.AI
keywords rationalereductionchain-of-thoughtverbositysentencepruninglikelihoodratiofine-tuningefficiencyreasoningtoken
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

Large language models trained to reason out loud write many intermediate sentences, and prior attempts to shorten those sentences during training have typically sacrificed accuracy. This paper proposes a sentence-level pruning criterion called verbosity: at each training step, the model compares the log-likelihood of the answer with and without a candidate early sentence, and drops the sentence only if its removal makes the correct answer more likely, with an added wrong-answer contrast in the VARR+ variant. The paper claims that models trained this way outperform models trained on complete rationales by an average of 7.71% in accuracy while reducing generated tokens by 19.87% (the main-results section reports 19.37%). If correct, this means many chain-of-thought sentences are not merely unnecessary but actively harmful, and that principled sentence-level pruning can improve both cost and quality at the same time.

What carries the argument

The central object is the verbosity score, a likelihood ratio that compares the model's probability of an answer when a candidate rationale sentence is removed versus when the full rationale is present: $\mathrm{verbosity}(y)=\log[p_\theta(y|R\setminus\{r_i\},x)/p_\theta(y|R,x)]$. In VARR+ this is paired with a wrong-answer contrast, so a sentence is removed only when the ground-truth answer becomes more likely and sampled wrong answers become relatively less likely. The training loop scans sentences from the front of the rationale, respects a linear removal cap $r(t)=\lfloor N_t\,t/T\rfloor$, and begins with a short warm-up phase before pruning starts, so the criterion decides which sentences vanish while the schedule decides how many can vanish.

What would settle it

At the final checkpoint, recompute $\mathrm{verbosity}(y_g)$ for every sentence the method removed during training: if most recomputed values are negative, the early-training decisions were not stable and the reported gains would not transfer to a one-shot pruning pipeline. A complementary check is to replace all training-time pruning decisions with scores from the final checkpoint and compare accuracy and token savings.

Watch

Extended reading notes

Core claim

The central claim is that redundancy in reasoning chains is concentrated in early sentences and can be identified by a likelihood-based score rather than by heuristics. Given an input $x$, a full rationale $R$, and a candidate sentence $r_i$, the verbosity of $r_i$ with respect to answer $y$ is $\mathrm{verbosity}(y)=\log[p_\theta(y|R\setminus\{r_i\},x)/p_\theta(y|R,x)]$, the log-likelihood ratio between the reduced and complete rationales. VARR removes $r_i$ during chain-of-thought fine-tuning when $\mathrm{verbosity}(y_g)\ge 0$; VARR+ additionally requires $\mathrm{verbosity}(y_w)-\mathrm{verbosity}(y_g)\le 0$ for sampled wrong answers. The paper argues, from a pilot study and extensive ablations, that this selective sentence removal preserves and even improves reasoning while cutting generated tokens, and that enforced removal without such criteria, or removal from random or back positions, degrades accuracy.

Load-bearing premise

The load-bearing premise is that the verbosity score computed at one training step remains a stable, reliable guide to which sentences are truly redundant for the rest of training; the pilot study supporting this rests on a single model and reports no error bars.

Editorial extensions

If this is right

  • If the central claim holds, existing chain-of-thought fine-tuning data can be shortened without an accuracy penalty, because many rationale sentences are not just redundant but harmful.
  • Sentence-level pruning with the verbosity criterion dominates token-level pruning: applying the same criterion to tokens gains 24.74% over ICoT-SI, and moving from tokens to sentences adds another 15.98% on the studied tasks.
  • Removal position matters: enforced removal of the first two front sentences without the criterion causes substantial accuracy drops, so the gain comes from criterion-guided selection rather than simply from shorter rationales.
  • The reported average token reduction of about one fifth comes with improved accuracy, meaning the efficiency gain does not trade away reasoning quality.
  • The wrong-answer contrast in VARR+ improves robustness and stability over VARR alone on most datasets, suggesting the extra constraint protects sentences needed to disambiguate correct from incorrect answers.

Reading between the lines

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

  • The verbosity criterion could likely be run offline before training, using a frozen reference model to produce a pruned dataset once; if that matches VARR's interleaved training results, the method becomes a cheap data-preprocessing step instead of a custom training loop.
  • The front-heavy redundancy finding suggests a testable hypothesis for long-context or multi-hop datasets: if critical premises appear early, front pruning should be less effective, and a position-agnostic version of the criterion would need to select later sentences instead.
  • A natural extension is to apply the same likelihood-ratio criterion at inference time to decide when to stop generating, effectively letting the model choose online which sentences to emit; the paper's training-time results make this plausible but do not demonstrate it.
  • Because the criterion is computed from the model itself, it may be sensitive to model scale and calibration; a cross-scale test would reveal whether the pruning signal transfers across model families beyond the 1B-to-7B range studied.
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 VARR/VARR+, a sentence-level rationale reduction framework for chain-of-thought fine-tuning. It defines a likelihood-based criterion called verbosity, derived from the KL divergence between a model with complete rationales and one with a candidate sentence removed, and removes early rationale sentences when the model's likelihood of the ground-truth answer does not decrease (Eq. 5) and, for VARR+, when the correct-answer likelihood improves relative to sampled wrong answers (Eq. 9). Training follows a warm-up stage and a linear removal schedule adopted from ICoT-SI. The authors evaluate on MathQA, GSM8K, CommonsenseQA, TriviaQA, and StrategyQA with Mistral 7B and Llama3.2 1B/3B, and claim an average accuracy improvement of 7.71% and a token-generation reduction of 19.87% (abstract) or 19.37% (Section 5.4) relative to Explicit-CoT.

Significance. If the empirical claims hold, the work is a useful contribution to efficient reasoning: it provides a principled, likelihood-based criterion for rationale pruning, grounds sentence-level reduction in a redundancy analysis (Section 3.2), and includes ablations that isolate the reduction unit, sentence position, warm-up ratio, and optimizer reinitialization. The use of open models and datasets, together with the detailed Algorithm 1, supports reproducibility. The main weaknesses are in verification of the headline numbers and in variance reporting, not in the coherence of the method's formulation.

major comments (4)
  1. [Section 5.4, Table 1] The headline claim of a 7.71% average accuracy improvement cannot be verified from the numerical results presented. Table 1, the only per-dataset accuracy table, reports VARR-Sent versus Exp-CoT on Mistral 7B; the relative differences are about +1.99%, -0.51%, +6.20%, +0.61%, and +4.67% for MathQA, GSM8K, CommonsenseQA, TriviaQA, and StrategyQA, averaging approximately 2.6%, not 7.71%. No corresponding per-dataset table is supplied for Llama3.2 1B and 3B, so the aggregate 7.71% figure is not reproducible from the manuscript. Please report the full per-model, per-dataset accuracy and token counts and state explicitly how the average is computed.
  2. [Appendix C; Section 5.4] The experimental evaluation is based on a single run per configuration (Appendix C), with no confidence intervals, standard deviations, or significance tests. Given the small training sets for CommonsenseQA (609 examples) and StrategyQA (1,832 examples) and the 1B-7B model scale, seed-to-seed variation in accuracy is typically several points; the negative GSM8K difference (-0.28) and the small TriviaQA gain (+0.51) in Table 1 are consistent with noise. The statement that VARR+ 'significantly increases' performance is therefore not supported by the evidence as reported. At minimum, report multiple seeds with means and variances for the main comparisons.
  3. [Abstract; Section 5.4] The abstract states a token-generation reduction of 19.87%, while Section 5.4 reports 19.37% for the same claim. This unexplained discrepancy affects a headline quantitative result; the correct average and its definition (over which datasets and models) should be stated consistently.
  4. [Section 3.2, Figure 1] The pilot study motivating early-sentence removal uses a single model (Mistral 7B) and reports no error bars or repeated sampling statistics. Because this observation directly justifies the sequential front-to-back removal order in Algorithm 1, it is load-bearing; the authors should show that the NLL-difference pattern is stable across seeds, models, or at least multiple rationale samples.
minor comments (4)
  1. [Throughout] Please correct typographical errors, including 'Quanitifying' (Section 3.1), 'Intutively' (Section 4.1), 'COT' in Figure 3, 'CommomsenseQA' in Figure 4, 'RAND-SENT REMOV AL' in Table 3, and 'enforcely' in Section 4.4.
  2. [Appendices F-H] Appendix F refers to 'as shown in Table' without a table number, and Appendix H's text refers to 'Table 8' for the MATH experiment while the displayed table is numbered Table 7; please correct the cross-references and table numbering throughout the appendices.
  3. [Section 4.3, Eq. (7)] The number K of negative samples in Eq. (7) is never specified in the main text or Appendix C for the natural-language TriviaQA setting; please state how K is chosen for each answer format.
  4. [Section 5.4] The baseline degradation percentages for ICoT-SI (21.98%) and Coconut (25.20%) are given without a corresponding numerical table; please either reference the relevant figure/table explicitly or include the underlying per-dataset numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the verbosity criterion is a model-dependent decision rule, but the central claims are evaluated on held-out test sets against external baselines and no fitted parameter or imported self-citation reduces the result to its inputs.

full rationale

The paper defines verbosity(yg) as a log-likelihood ratio (Eqs. 2-3) and uses non-negativity as a removal criterion (Eqs. 5 and 9). This is a self-referential decision rule because the trained model computes its own pruning signal, but it is not a fitted parameter renamed as a prediction. The load-bearing empirical claims, that VARR/VARR+ improve accuracy by 7.71% and reduce generated tokens by 19.37-19.87% relative to Explicit-CoT, are measured on held-out test splits against external baselines (Explicit-CoT, ICoT-SI, Coconut). Therefore these claims do not reduce by construction to the pruning criterion. The paper contains no self-citations by the present authors, no imported uniqueness theorem, and no ansatz smuggled in via citation. The pilot study in Section 3 is single-model and lacks error bars, and Appendix C discloses single run evaluation; these are statistical robustness limitations, not circularity. The training-time dependence of the pruning decision on the evolving model is a legitimate methodological risk, but it is not a circular reduction because the test evaluation is independent of the pruning criterion. Consequently, no circular step can be exhibited under the required evidentiary standard, and the score is 0.

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

The central claim rests on several modeling choices: the sentence as the reduction unit, the likelihood-ratio criterion computed from the model being trained, the front-position prior from a single-model pilot study, and hand-set hyperparameters (warm-up ratio, removal schedule, number of negatives). None of these are derived from first principles, but they are individually reasonable and partially ablated.

free parameters (4)
  • Warm-up ratio = 0.1
    Fraction of training steps before verbosity-based removal begins; chosen via ablation on TriviaQA and CommonsenseQA (Sec. 5.5.3, Fig. 4), and performance varies with it.
  • Removal schedule r(t) = linearly increasing from 0 to N_t
    Maximum number of sentences removable at step t, adopted from ICoT-SI (Deng et al., 2024b), Eq. (10). It directly controls how much reduction occurs and is not derived from first principles.
  • Number of negative samples K = in-batch negatives (task-dependent)
    Used in Eq. (7) to approximate expectation over wrong answers; set to the full set of non-correct labels for multiple-choice/TF tasks and in-batch negatives for others (Appendix C).
  • Threshold 0 in Eqs. (5) and (9) = 0
    The removal criteria use zero as the cutoff for verbosity differences; this is a hand-set threshold, not a fitted constant.
assumptions (4)
  • domain assumption Sentences are linguistically natural and effective units for rationale reduction.
    The paper argues tokens can distort semantics (Sec. 2.2) and shows empirically that sentence-level removal beats token-level (Table 1), but this is a modeling choice, not a derived fact.
  • domain assumption The model's own likelihood p_theta is a reliable measure of whether a sentence is redundant for future training.
    Eqs. (2)-(5) and (9) use the in-training model's log-likelihood ratio to decide removal; this assumes the current model's probabilities predict which sentences will be unneeded by the final model.
  • domain assumption Redundancy is concentrated in early rationale sentences.
    Section 3.2 and Figure 1 support this for Mistral 7B without error bars; the method only removes from the front, so this generalization is load-bearing.
  • standard math The ground-truth answer distribution q(y|x) is one-hot.
    Used in the derivation of Eq. (3) from Eq. (2); standard in supervised learning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Verbosity-Aware Rationale Reduction: Effective Reduction of Redundant Rationale via Principled Criteria." pith.science (2026). https://pith.science/paper/Z724SXFY

@misc{pith2026241221006,
  author       = {Pith},
  title        = {Pith review of: Verbosity-Aware Rationale Reduction: Effective Reduction of Redundant Rationale via Principled Criteria},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z724SXFY}},
  note         = {Machine review of arXiv:2412.21006}
}
read the original abstract

Large Language Models (LLMs) rely on generating extensive intermediate reasoning units (e.g., tokens, sentences) to enhance final answer quality across a wide range of complex tasks. While this approach has proven effective, it inevitably increases substantial inference costs. Previous methods adopting token-level reduction without clear criteria result in poor performance compared to models trained with complete rationale. To address this challenge, we propose a novel sentence-level rationale reduction framework leveraging likelihood-based criteria, verbosity, to identify and remove redundant reasoning sentences. Unlike previous approaches, our method leverages verbosity to selectively remove redundant reasoning sentences while preserving reasoning capabilities. Our experimental results across various reasoning tasks demonstrate that our method improves performance by an average of 7.71% while reducing token generation by 19.87% compared to model trained with complete reasoning paths.

Figures

Figures reproduced from arXiv: 2412.21006 by the authors.

Figure 1
Figure 1. NLL differences across varying sizes of {ri}. The ‘original’ represents the NLL with the complete rationale, while ‘front’, ‘random’, and ‘back’ indicate that {ri} is sampled from the front, random, and back indices of the full index set, respectively. 3.1 Quanitifying the Redundancy Before delving into the method, we first investi￾gate which positions within the rationale sentences should be selected for reduction.… view at source ↗
Figure 2
Figure 2. Overview of our VARR/VARR+ framework. Initially, we select a candidate sentence from the beginning [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Pareto plot of accuracy versus the number of generated tokens. The gray dotted lines connect the Pareto [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Accuracy (barplot) and the average generated [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Removal ratio of redundant sentences during [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: NLL differences across varying sizes of {ri}. The ‘original’ represents the NLL for the full rationale, while ‘front’, ‘random’, and ‘back’ indicate that ri are sampled from the front, random, and back indices of the full index set, respectively. B VARR Implementation …
Figure 7
Figure 7. Figure 7: Additional analysis of the Removal ratio on [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 6 canonical work pages

  1. [1]

    LlamaTeam AI@Meta. 2024. https://arxiv.org/abs/2407.21783 The llama 3 herd of models . Preprint, arXiv:2407.21783

    arXiv 2024

  2. [2]

    Aida Amini, Saadia Gabriel, Shanchuan Lin, Rik Koncel-Kedziorski, Yejin Choi, and Hannaneh Hajishirzi. 2019 a . Mathqa: Towards interpretable math word problem solving with operation-based formalisms. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 ...

    work page 2019

  3. [3]

    Aida Amini, Saadia Gabriel, Shanchuan Lin, Rik Koncel-Kedziorski, Yejin Choi, and Hannaneh Hajishirzi. 2019 b . https://doi.org/10.18653/v1/N19-1245 M ath QA : Towards interpretable math word problem solving with operation-based formalisms . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguisti...

    doi:10.18653/v1/n19-1245 2019

  4. [4]

    Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. 2024. A general theoretical paradigm to understand learning from human preferences. In International Conference on Artificial Intelligence and Statistics, pages 4447--4455. PMLR

    2024

  5. [6]

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

    arXiv 2021

  6. [7]

    Yihe Deng, Weitong Zhang, Zixiang Chen, and Quanquan Gu. 2024 a . https://arxiv.org/abs/2311.04205 Rephrase and respond: Let large language models ask better questions for themselves . Preprint, arXiv:2311.04205

    arXiv 2024

  7. [8]

    Yuntian Deng, Yejin Choi, and Stuart Shieber. 2024 b . https://arxiv.org/abs/2405.14838 From explicit cot to implicit cot: Learning to internalize cot step by step . Preprint, arXiv:2405.14838

    arXiv 2024

  8. [9]

    Yuntian Deng, Kiran Prasad, Roland Fernandez, Paul Smolensky, Vishrav Chaudhary, and Stuart Shieber. 2023. https://arxiv.org/abs/2311.01460 Implicit chain of thought reasoning via knowledge distillation . Preprint, arXiv:2311.01460

    arXiv 2023

Show all 39 references
  1. [10]

    Qingxiu Dong, Jingjing Xu, Lingpeng Kong, Zhifang Sui, and Lei Li. 2023. https://openreview.net/forum?id=pNtG6NAmx0 Statistical knowledge assessment for large language models . In Thirty-seventh Conference on Neural Information Processing Systems

    2023

  2. [11]

    Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346--361

    2021

  3. [12]

    Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. 2024. Training large language models to reason in a continuous latent space. arXiv preprint arXiv:2412.06769

    2024 arXiv

  4. [13]

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset. NeurIPS

    2021

  5. [14]

    Jiwoo Hong, Noah Lee, and James Thorne. 2024. Orpo: Monolithic preference optimization without reference model. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 11170--11189

    2024

  6. [15]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas ...

    2023 arXiv

  7. [16]

    Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p...

    2017

  8. [17]

    Seungone Kim, Se June Joo, Doyoung Kim, Joel Jang, Seonghyeon Ye, Jamin Shin, and Minjoon Seo. 2023. The cot collection: Improving zero-shot and few-shot learning of language models via chain-of-thought fine-tuning. In The 2023 Conference on Empirical Methods in Natural Langua...

    2023

  9. [18]

    Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, John D Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, Lei M Zhang, Kay McKinney, Disha Shrivastava, Cosmin Paduraru, George Tucker, Doina Precup, Feryal Behbahani, and Aleksandra Faust. 2024...

    2024 arXiv

  10. [19]

    Tengxiao Liu, Qipeng Guo, Xiangkun Hu, Cheng Jiayang, Yue Zhang, Xipeng Qiu, and Zheng Zhang. 2024. https://openreview.net/forum?id=w4AnTVxAO9 Can language models learn to skip steps? In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  11. [20]

    Ilya Loshchilov and Frank Hutter. 2019. https://openreview.net/forum?id=Bkg6RiCqY7 Decoupled weight decay regularization . In International Conference on Learning Representations

    2019

  12. [21]

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. 2024. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36

    2024

  13. [22]

    Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, et al. 2021. Show your work: Scratchpads for intermediate computation with language models. arXiv preprint arXiv:2112.00114

    2021 arXiv

  14. [23]

    OpenAI . 2024. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/ Gpt-4o mini: advancing cost-efficient intelligence

    2024

  15. [24]

    Santosh Kumar Radha, Yasamin Nouri Jelyani, Ara Ghukasyan, and Oktay Goktas. 2024. https://arxiv.org/abs/2409.12618 Iteration of thought: Leveraging inner dialogue for autonomous large language model reasoning . Preprint, arXiv:2409.12618

    2024 arXiv

  16. [25]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36

    2024

  17. [26]

    Varshini Reddy, Rik Koncel-Kedziorski, Viet Dac Lai, Michael Krumdick, Charles Lovering, and Chris Tanner. 2024. Docfinqa: A long-context financial reasoning dataset. arXiv preprint arXiv:2401.06915

    2024

  18. [27]

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2024. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36

    2024

  19. [28]

    Damien Sileo. 2024. https://aclanthology.org/2024.lrec-main.1361 tasksource: A large collection of NLP tasks with a structured dataset preprocessing framework . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evalu...

    2024

  20. [29]

    Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2019. Commonsenseqa: A question answering challenge targeting commonsense knowledge. In Proceedings of the 2019 Conference of the North, page 4149. Association for Computational Linguistics

    2019

  21. [30]

    Shubham Toshniwal, Wei Du, Ivan Moshkov, Branislav Kisacanin, Alexan Ayrapetyan, and Igor Gitman. 2024. Openmathinstruct-2: Accelerating ai for math with massive open-source instruction data. arXiv preprint arXiv:2410.01560

    2024 arXiv

  22. [31]

    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. 2023. https://openreview.net/forum?id=1PL1NIMMrw Self-consistency improves chain of thought reasoning in language models . In The Eleventh International Conferenc...

    2023

  23. [32]

    Yu Wang, Shiwan Zhao, Zhihu Wang, Heyuan Huang, Ming Fan, Yubo Zhang, Zhixing Wang, Haijun Wang, and Ting Liu. 2024. https://arxiv.org/abs/2409.03271 Strategic chain-of-thought: Guiding accurate reasoning in llms through strategy elicitation . Preprint, arXiv:2409.03271

    2024 arXiv

  24. [33]

    Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, Rui Zheng, Xiaoran Fan, Xiao Wang, Limao Xiong, Yuhao Zhou, Weiran Wang, Changhao Jiang, Yicheng Zou, Xiangyang Liu, Zhangyue Yin, Shihan Dou, Rongxiang Weng, W...

    2023 arXiv

  25. [34]

    Gonzalez, and Bin CUI

    Ling Yang, Zhaochen Yu, Tianjun Zhang, Shiyi Cao, Minkai Xu, Wentao Zhang, Joseph E. Gonzalez, and Bin CUI. 2024. https://openreview.net/forum?id=ANO1i9JPtb Buffer of thoughts: Thought-augmented reasoning with large language models . In The Thirty-eighth Annual Conference on N...

    2024

  26. [35]

    Griffiths, Yuan Cao, and Karthik R Narasimhan

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik R Narasimhan. 2023. https://openreview.net/forum?id=5Xc1ecxO1h Tree of thoughts: Deliberate problem solving with large language models . In Thirty-seventh Conference on Neural Informat...

    2023

  27. [36]

    Zhangyue Yin, Qiushi Sun, Qipeng Guo, Zhiyuan Zeng, Xiaonan Li, Junqi Dai, Qinyuan Cheng, Xuan-Jing Huang, and Xipeng Qiu. 2024. Reasoning in flux: Enhancing large language models reasoning through uncertainty-aware adaptive guidance. In Proceedings of the 62nd Annual Meeting ...

    2024

  28. [37]

    Longhui Yu, Weisen Jiang, Han Shi, Jincheng YU, Zhengying Liu, Yu Zhang, James Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. 2024 a . https://openreview.net/forum?id=N8N0hgNDRt Metamath: Bootstrap your own mathematical questions for large language models . In The Twelfth I...

    2024

  29. [38]

    Ping Yu, Jing Xu, Jason Weston, and Ilia Kulikov. 2024 b . Distilling system 2 into system 1. arXiv preprint arXiv:2407.06023

    2024 arXiv

  30. [39]

    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...

  31. [40]

    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 10, 2026 · model on record in the stance chip above.