Pith. sign in

REVIEW 5 major objections 7 minor 33 references

Reasoning Strategies in Large Language Models: Can They Follow, Prefer, and Optimize?

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Prompting can steer a language model into one of four human-inspired reasoning strategies, and combining the resulting answers with simple vote or confidence rules outperforms any single prompt.

desk verdict Honest empirical study of strategy-conditioned prompting, but the conclusion overstates 'consistently outperform' when the paper's own Table 3 shows several selectors losing to the best single prompt. read the letter →

arxiv 2507.11423 v2 pith:DKY64QWK submitted 2025-07-15 cs.CL

Yanjian Zhang , Guillaume Wisniewski , Nadi Tomeh , Thierry Charnois This is my paper
classification cs.CL
keywords reasoningstrategiespromptengineeringlogicaldeductionensembleselectionchain-of-thoughtLLMTruthQuestZebraLogic
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 tests whether a language model's reasoning style can be deliberately controlled, and whether controlling it helps solve logical puzzles. The authors prompt off-the-shelf models with step-by-step descriptions of four human reasoning strategies (supposition following, chain construction, compound reasoning, and concatenation) and find that the models broadly adopt the requested style without fine-tuning. No single strategy beats the standard no-strategy prompt, but an oracle that always picked the best strategy per problem would gain up to 40 points, so the paper runs all strategies in parallel and selects among their answers using majority vote, probability or entropy confidence, or a verifier. These selectors consistently beat every individual prompt, lifting accuracy by 7–11 points on TruthQuest and 1–3 points on ZebraLogic. The central conclusion is that reasoning style is a controllable latent variable and that simple ensembles can make LLM reasoning more reliable.

What carries the argument

The load-bearing object is a strategy-conditioned prompt: a template that defines one of four human reasoning strategies and then instructs the model to follow a numbered sequence of steps for that strategy (for example, for supposition following, enumerate propositions, make a supposition, trace consequences, test the alternative on contradiction). This prompt converts a latent default preference into an externally invoked reasoning mode. The second mechanism is the ensemble selector: all four strategy prompts run in parallel and the final answer is chosen either by majority vote, by the highest combined probability of the reasoning and answer segments, by the lowest mean token entropy, by an external LLM verifier that scores chunks of the reasoning, or by hybrids that use a vote with probability or verifier as tie-breakers.

What would settle it

Have two independent annotators classify a sample of model responses on both TruthQuest and ZebraLogic into the four strategies using a pre-registered coding scheme, and report inter-annotator agreement: if agreement is near chance, or if responses under a given prompt are not classified as that strategy more often than chance, the claim that prompts control reasoning style fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that reasoning style is a controllable latent variable in large language models. Using prompts that pair each of four human-inspired strategies with a concrete step-by-step procedure, the authors find that Phi-4-14B, DeepSeek-R1-Distill-Qwen-7B, and Qwen3-8B follow the requested strategy on the TruthQuest benchmark, with adherence between 32% and 99% depending on model and strategy. On TruthQuest and ZebraLogic, no single strategy prompt outperforms the baseline prompt that specifies no strategy; accuracy across the five prompts is nearly flat. The gap to the oracle (always choosing the best strategy per problem, up to 82.9% versus 49.0% for Phi-4 on TruthQuest) shows that the information needed for a large gain is present in the model's own outputs. Combining all strategies with majority vote, maximum answer probability, minimum entropy, or an LLM verifier beats every individual prompt, with vote-plus-verifier and vote-plus-probability among the best, although all selectors fall short of the oracle.

Load-bearing premise

The whole comparison assumes the prompts actually elicit the intended strategies, but adherence was verified by manually annotating only 100 answers per prompt on one dataset, with annotation rules the authors call necessary but insufficient and with no annotator-agreement check, and some prompts were followed as rarely as 32% of the time.

Editorial extensions

If this is right

  • Because no single strategy wins, adaptive or ensembled selection across strategies is a practical path to more reliable reasoning without retraining or meta-prompting.
  • Oracle gaps of up to 40 points mean fixed prompts leave large headroom for logical-deduction tasks.
  • Verifier-based selection helps on short, simple reasoning chains and degrades as problems get harder, so chunk-wise verification is not a general fix.
  • The vote-plus-verifier hybrid matches the best accuracy while paying extra decoding cost only on ties, reducing the compute overhead of ensembling.
  • Because strategy control works zero-shot, it can be combined with self-consistency sampling, an extension the paper leaves to future work.

Reading between the lines

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

  • A testable extension: adding more distinct strategy prompts (for example, visual or symbolic styles) should widen the gap between the ensemble and any single prompt, and this widening would confirm that diversity, not sampling, drives the gains.
  • The oracle gap implies the correct answer is usually already among the generated outputs, so better selection (such as a trained reranker over strategy outputs) may be a cheaper route to higher accuracy than a stronger base model.
  • Adherence varies sharply by model (concatenation is followed only 32% of the time by one model), so strategy control is itself model-dependent; asking a model to name its own strategy before answering and checking whether self-reports predict accuracy is a natural next experiment.
  • The results reframe reliability as a property of the selection mechanism: whether a model can be made to reason the way a problem demands may matter more than how it reasons by default.
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 / 7 minor

Summary. The paper investigates whether LLMs can be prompted to follow four human-inspired reasoning strategies (supposition following, chain construction, compound, and concatenation) and whether ensembling the outputs of these strategies—via majority vote, probability-based confidence, minimum entropy, or a verifier—improves accuracy on the TruthQuest and ZebraLogic benchmarks. It reports that no single strategy consistently dominates, that the prompts elicit partial adherence to the intended strategy based on a small manual annotation, and that some ensemble methods improve over individual prompts, with an oracle indicating substantial potential headroom. The authors conclude that reasoning style is a controllable latent variable and that their selectors consistently outperform every individual strategy prompt.

Significance. If the claims were fully supported, the paper would make a useful contribution by extending prior work on LLM strategy preferences to strategy control and by proposing inference-time combination methods that require no additional training. The oracle analysis is a helpful way to quantify the potential gains from adaptive selection, and the experimental design covers two benchmarks and three models. The paper also gives a candid discussion of verifier limitations on hard problems. However, as detailed in the major comments, the central claim of consistent ensemble superiority is contradicted by the paper's own data, and the missing self-consistency control prevents the gains from being attributed to strategy diversity. The current version therefore does not yet establish the main conclusions.

major comments (5)
  1. [Section 4.2, Table 3, and Conclusion] The conclusion states that the proposed selectors achieve gains 'consistently outperforming every individual strategy prompt,' but Table 3 contradicts this. For Phi-4 on TruthQuest, min entropy@4 (48.8%), max prob@4 (48.4%), and verifier (46.9%) all fall below the best single strategy (49.0%, Chain Construction). For Phi-4 on ZebraLogic, min entropy@4 (27.1%) and max prob@4 (26.5%) fall below the best single strategy (27.3%). The claim should be revised to reflect that only some selectors improve some models on some datasets.
  2. [Section 4.1 and Appendix D] The paper does not include a self-consistency baseline: majority voting or confidence-based selection over multiple samples of a single strategy (e.g., four samples of Chain Construction or of the no-strategy prompt). Appendix D notes that the no-strategy prompt was sampled four times per question, so this control is readily available. Without it, the observed ensemble gains could be due to ensembling multiple stochastic samples—which is known to help from self-consistency—rather than to the diversity of reasoning strategies. This is load-bearing for the claim that strategy diversity is the source of the improvement.
  3. [Section 3.3, Table 1, and Appendix A] The evidence that prompts reliably elicit the intended strategies is thin. Table 1 is based on manual annotation of only 100 responses per prompt and per model on TruthQuest, with annotation rules described as 'necessary but insufficient' in Appendix A and no inter-annotator agreement reported. Adherence is as low as 32% for concatenation on R1-Distill, and no adherence measurement is reported for ZebraLogic. Consequently, the interpretation of the ensemble as combining four distinct reasoning modes rather than merely four different instruction variants is not strongly established.
  4. [Introduction, bullet points] The Introduction claims that an oracle 'would raise accuracy by up to 40 percentage points, exposing substantial untapped potential.' However, the largest gap between the oracle and the best single strategy in Table 2 is 82.9% - 49.0% = 33.9 points (Phi-4 on TruthQuest); the other gaps are 26.7, 10.4, and 4.4 points. The 'up to 40' figure is not supported by the reported data and should be corrected.
  5. [Section 4.2 and Table 3] No error bars, confidence intervals, or significance tests are reported for the ensemble gains. Many differences are small (1–3 points on ZebraLogic) and some are negative, so the point estimates do not substantiate the claim of 'consistent' improvement. A paired test (e.g., McNemar) or bootstrap confidence intervals should be provided for the key comparisons.
minor comments (7)
  1. [Table 2 header] The header 'Phi4-14B' should be 'Phi-4-14B' for consistency with the model name used elsewhere.
  2. [Tables 4–7] The row and column formatting in these tables appears misaligned in the manuscript; please ensure the layout clearly labels each strategy and complexity subset.
  3. [Section 3.2] The paper states that the third prompt formulation 'produced the best results' after manual review of 'a few randomly selected examples,' but no details are given. Specify the number of examples, the review criteria, and how 'best' was judged.
  4. [Appendix E] The main text does not mention that the reported max prob@4 and min entropy@4 results use λp = λe = 1/2, nor that the sensitivity analysis in Appendix E shows the optimal λ varies by dataset. Add a sentence in Section 4.1 clarifying the fixed setting and the potential sensitivity.
  5. [Section 4.1] The verifier uses R1-Distill for all generator models, including R1-Distill itself. Clarify whether this choice is intentional and note any possible bias when the verifier is the same as the generator.
  6. [Section 5] The claim that 'our study is the first to demonstrate that an LLM can be prompted to switch among multiple reasoning strategies' is strong; given existing work on adaptive reasoning and strategy prompting, this novelty claim should be softened or more carefully delimited.
  7. [Tables 8 and 9 captions] The captions contain typos ('Gree' should be 'Green', 'Answser' should be 'Answer'). Please proofread.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's central claims are empirical benchmark results, and its oracle and ensemble selectors are not constructed from the target conclusions.

full rationale

This is an empirical study with no derivation chain that reduces to its inputs. The four reasoning strategies are taken from prior work by Mondorf and Plank, not defined in terms of the paper's own results, and the prompts are hand-written instructions. The claim that prompts control reasoning strategy is supported by a manual annotation of 100 responses per model on TruthQuest; however weak or insufficiently validated that annotation may be, it is not circular, because the annotation rules are independent of the accuracy results and of the ensemble outcomes. The oracle accuracy is explicitly presented as an upper bound ('the proportion of problems for which at least one strategy-specific prompt yields the correct answer') and is used only to motivate the ensemble, not as a selector, so it does not force the ensemble numbers. The ensemble criteria are defined from independent signals: majority vote over strategy outputs, token-level probability products, mean entropy, and a separate verifier LLM. None of these are fitted to the accuracy numbers they are used to predict. The main results use a fixed combination weight lambda = 1/2, chosen a priori rather than selected to maximize test-set accuracy, and the appendix's sensitivity analysis does not rename a fitted parameter as a prediction. There are no load-bearing self-citations: the cited prior work is by other authors, and no uniqueness theorem or ansatz is smuggled in via citation. The strongest concern is an internal inconsistency: the conclusion that ensemble selectors 'consistently outperforming every individual strategy prompt' is contradicted by Table 3 for Phi-4 on TruthQuest, where min entropy@4 (48.8%), max prob@4 (48.4%), and verifier (46.9%) all fall below the best single strategy (49.0%). That is a correctness or validity problem, not a circularity problem, because the claim and the data are not equivalent by construction. Under the circularity rubric, the paper is self-contained against external benchmarks, and no targeted result is assumed in its own inputs.

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

The central claims rest on dataset validity, the accuracy of the regex-based evaluation, the reliability of manual strategy annotation, and the assumption that the hand-written prompts actually induce distinct strategies. There are no invented physical entities or formal derivation components. The only tuned numbers are the hyperparameters λp, λe, and the verifier chunk size, all set by hand without a principled selection procedure.

free parameters (3)
  • λp (probability combination weight) = 0.5
    In Appendix E, the combined response probability is defined as Prational^(2(1-λp)) × Panswer^(2λp), with λp set to 1/2 in the main article. The supplementary analysis sweeps λp across 100 values and shows the optimal value is dataset-specific, so the default is a hand-chosen hyperparameter, not fitted to the main results.
  • λe (entropy combination weight) = 0.5
    Similarly, the combined entropy is defined as (1-λe)Hrational + λeHanswer, with λe = 1/2 in the main text. The appendix shows different optimal λe for TruthQuest vs. ZebraLogic. This is a hand-chosen hyperparameter.
  • Verifier chunk size = approximately 100 words
    The verifier splits each answer into roughly 100-word chunks truncated by sentence boundaries. This threshold is chosen manually and is not justified with sensitivity analysis.
assumptions (4)
  • domain assumption The ground-truth labels in TruthQuest and ZebraLogic are correct, and the regular-expression evaluation that checks whether the generated text contains the correct answer is a valid accuracy measure.
    Section 2.2 states accuracy is defined as the percentage of responses containing the correct final answer. This ignores whether the reasoning itself is valid, and the regex could produce false positives if an answer line is mentioned only in passing.
  • domain assumption The manual annotation of strategy adherence is reliable and representative.
    In Section 3.3, adherence is measured on 100 randomly selected TruthQuest questions for two models, annotated by the authors with rules described as 'necessary but insufficient'. No inter-annotator agreement is reported, and adherence rates for some strategies are low (e.g., 32% for concatenation on R1-Distill).
  • domain assumption The four reasoning strategies are distinct, and the designed prompts faithfully elicit each strategy.
    The strategies are borrowed from Mondorf and Plank (2024a) and represented by manual prompt templates. The paper's own adherence table (Table 1) shows substantial variability, especially for chain construction and concatenation, so the intended distinctness is only partially achieved.
  • domain assumption The tested LLMs (Phi-4-14B, R1-Distill-Qwen-7B, Qwen3-8B) are representative of current models, and the sampling hyperparameters (temperature 0.6, top-p 0.9) from prior work are appropriate.
    Section 2.3 fixes these hyperparameters to match Mondorf and Plank. The paper does not test sensitivity to these choices or include greedy decoding comparisons.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reasoning Strategies in Large Language Models: Can They Follow, Prefer, and Optimize?." pith.science (2026). https://pith.science/paper/DKY64QWK

@misc{pith2026250711423,
  author       = {Pith},
  title        = {Pith review of: Reasoning Strategies in Large Language Models: Can They Follow, Prefer, and Optimize?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DKY64QWK}},
  note         = {Machine review of arXiv:2507.11423}
}
read the original abstract

Human reasoning involves different strategies, each suited to specific problems. Prior work shows that large language model (LLMs) tend to favor a single reasoning strategy, potentially limiting their effectiveness in diverse reasoning challenges. In this work, we investigate whether prompting can control LLMs reasoning strategies and assess its impact on logical problem-solving. While our experiments show that no single strategy consistently improves accuracy, performance could be enhanced if models could adaptively choose the optimal strategy. We propose methods to guide LLMs in strategy selection, highlighting new ways to refine their reasoning abilities.

Figures

Figures reproduced from arXiv: 2507.11423 by the authors.

Figure 1
Figure 1. The prompt guiding the LLM to adopt the Chain Construction strategy for TruthQuest. introduced in Section 2. We aim to evaluate both the model’s accuracy in providing the correct an￾swer and its adherence to the strategy specified in the prompt. First, we checked whether the model followed the strategy suggested by the prompt. For each prompt, we manually annotate answers produced by Phi-4-14B and R1-Distill to 100 … view at source ↗
Figure 4
Figure 4. The overall accuracy of Phi-4 model in Ze [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. The overall accuracy of Phi-4 model in Ze [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figures from the paper (2 more)
Figure 6
Figure 6. Figure 6: The accuracy distribution of different merging [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: The accuracy distribution of different merging [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 13 canonical work pages

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  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 word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Hewett, Mojan Javaheripi, Piero Kauffmann, James R

    Marah Abdin, Jyoti Aneja, Harkirat Singh Behl, S \'e bastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J. Hewett, Mojan Javaheripi, Piero Kauffmann, James R. Lee, Yin Tat Lee, Yuanzhi Li, Weishung Liu, Caio C'esar Teodoro Mendes, Anh Nguyen, Eric Price, Gustavo de Rosa, Olli Saarikivi, and 8 others. 2024. https://api.semanticscholar...

    work page 2024

  4. [4]

    Guangsheng Bao, Hongbo Zhang, Cunxiang Wang, Linyi Yang, and Yue Zhang. 2025. https://aclanthology.org/2025.coling-main.524/ How likely do LLM s with C o T mimic human reasoning? In Proceedings of the 31st International Conference on Computational Linguistics, pages 7831--7850, Abu Dhabi, UAE. Association for Computational Linguistics

    2025

  5. [5]

    Bradley C. A. Brown, Jordan Juravsky, Ryan Saul Ehrlich, Ronald Clark, Quoc V. Le, Christopher R \' e , and Azalia Mirhoseini. 2024. https://doi.org/10.48550/ARXIV.2407.21787 Large language monkeys: Scaling inference compute with repeated sampling . CoRR, abs/2407.21787

    arXiv doi:10.48550/arxiv.2407.21787 2024

  6. [6]

    Junhyuk Choi, Yeseon Hong, and Bugeun Kim. 2024. https://api.semanticscholar.org/CorpusID:271244746 People will agree what i think: Investigating llm's false consensus effect . ArXiv, abs/2407.12007

    arXiv 2024

  7. [7]

    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. https://api.semanticscholar.org/CorpusID:239998651 Training verifiers to solve math word problems . ArXiv, abs/2110.14168

    arXiv 2021

  8. [8]

    Tiwalayo Eisape, Mh Tessler, Ishita Dasgupta, Fei Sha, Sjoerd van Steenkiste, and Tal Linzen. 2023. https://api.semanticscholar.org/CorpusID:264832674 A systematic comparison of syllogistic reasoning in humans and language models . ArXiv, abs/2311.00445

    arXiv 2023

Show all 33 references
  1. [9]

    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

    2025 arXiv

  2. [10]

    Philip N Johnson-Laird. 2010. Mental models and human reasoning. Proceedings of the National Academy of Sciences, 107(43):18243--18250

    2010

  3. [11]

    Sangeet Khemlani and Phil Johnson-Laird. 2019. https://doi.org/10.1007/s13218-019-00599-w Why machines don’t (yet) reason like people . KI - Künstliche Intelligenz

    2019 doi

  4. [12]

    Takeshi Kojima, Shixiang (Shane) Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/file/8bb0d291acd4acf06ef112099c16f326-Paper-Conference.pdf Large language models are zero-shot reasoners . In Advances in Neural Inf...

    2022

  5. [13]

    Bill Yuchen Lin, Ronan Le Bras, and Yejin Choi. 2024. https://huggingface.co/spaces/allenai/ZebraLogic Zebralogic: Benchmarking the logical reasoning ability of language models

    2024

  6. [14]

    Bill Yuchen Lin, Ronan Le Bras, Kyle Richardson, Ashish Sabharwal, Radha Poovendran, Peter Clark, and Yejin Choi. 2025. https://arxiv.org/abs/2502.01100 Zebralogic: On the scaling limits of llms for logical reasoning . Preprint, arXiv:2502.01100

    2025 arXiv

  7. [15]

    Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blunsom. 2017. https://api.semanticscholar.org/CorpusID:12777818 Program induction by rationale generation: Learning to solve and explain algebraic word problems . In Annual Meeting of the Association for Computational Linguistics

    2017

  8. [16]

    Thomas McCoy, Shunyu Yao, Dan Friedman, Matthew Hardy, and Thomas L

    R. Thomas McCoy, Shunyu Yao, Dan Friedman, Matthew Hardy, and Thomas L. Griffiths. 2023. https://api.semanticscholar.org/CorpusID:262464572 Embers of autoregression: Understanding large language models through the problem they are trained to solve . ArXiv, abs/2309.13638

    2023 arXiv

  9. [17]

    Philipp Mondorf and Barbara Plank. 2024 a . https://doi.org/10.18653/v1/2024.acl-long.508 Comparing inferential strategies of humans and large language models in deductive reasoning . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (V...

    2024 doi

  10. [18]

    Philipp Mondorf and Barbara Plank. 2024 b . https://aclanthology.org/2024.emnlp-main.404 Liar, liar, logical mire: A benchmark for suppositional reasoning in large language models . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, page...

    2024

  11. [19]

    Elizabeth Newton and Maxwell Roberts. 2004. Methods of thought: Individual differences in reasoning strategies. Psychology Press

    2004

  12. [20]

    Andreas Opedal, Alessandro Stolfo, Haruki Shirakami, Ying Jiao, Ryan Cotterell, Bernhard Scholkopf, Abulhair Saparov, and Mrinmaya Sachan. 2024. https://api.semanticscholar.org/CorpusID:267334679 Do language models exhibit the same cognitive biases in problem solving as human ...

    2024 arXiv

  13. [21]

    Vaishnavi Shrivastava, Ananya Kumar, and Percy Liang. 2025. https://arxiv.org/abs/2502.01126 Language models prefer what they know: Relative confidence estimation via confidence preferences . Preprint, arXiv:2502.01126

    2025 arXiv

  14. [22]

    Amir Taubenfeld, Tom Sheffer, Eran Ofek, Amir Feder, Ariel Goldstein, Zorik Gekhman, and Gal Yona. 2025. https://arxiv.org/abs/2502.06233 Confidence improves self-consistency in llms . Preprint, arXiv:2502.06233

    2025

  15. [23]

    Qwen Team. 2025. https://qwenlm.github.io/blog/qwen3/ Qwen3

    2025

  16. [24]

    Johnson-Laird

    Jean-Baptiste Van der Henst, Yingrui Yang, and P.N. Johnson-Laird. 2002. https://doi.org/10.1207/s15516709cog2604\_2 Strategies in sentential reasoning . Cognitive Science, 26(4):425--468

    2002 doi

  17. [25]

    Le, Ed H

    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 Conferen...

    2023

  18. [26]

    Xuezhi Wang and Denny Zhou. 2024. https://api.semanticscholar.org/CorpusID:267681847 Chain-of-thought reasoning without prompting . ArXiv, abs/2402.10200

    2024 arXiv

  19. [27]

    Chi, Quoc V

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. http://papers.nips.cc/paper\_files/paper/2022/hash/9d5609613524ecf4f15af0f7b31abca4-Abstract-Conference.html Chain-of-thought prompting elicits reasoning...

    2022

  20. [28]

    Xin Xu, Yan Xu, Tianhao Chen, Yuchen Yan, Chengwu Liu, Zaoyu Chen, Yufei Wang, Yichun Yin, Yasheng Wang, Lifeng Shang, and Qun Liu. 2025. https://arxiv.org/abs/2502.12022 Teaching llms according to their aptitude: Adaptive reasoning for mathematical problem solving . Preprint,...

    2025

  21. [29]

    Shangzi Xue, Zhenya Huang, Jiayu Liu, Xin Lin, Yuting Ning, Binbin Jin, Xin Li, and Qi Liu. 2024. https://proceedings.neurips.cc/paper_files/paper/2024/file/01025a4e79355bb37a10ba39605944b5-Paper-Conference.pdf Decompose, analyze and rethink: Solving intricate problems with hu...

    2024

  22. [30]

    Zhangyue Yin, Qiushi Sun, Qipeng Guo, Zhiyuan Zeng, Xiaonan Li, Junqi Dai, Qinyuan Cheng, Xuanjing Huang, and Xipeng Qiu. 2024. https://doi.org/10.18653/v1/2024.acl-long.131 Reasoning in flux: Enhancing large language models reasoning through uncertainty-aware adaptive guidanc...

    2024 doi

  23. [31]

    Lunjun Zhang, Arian Hosseini, Hritik Bansal, Mehran Kazemi, Aviral Kumar, and Rishabh Agarwal. 2024. https://doi.org/10.48550/ARXIV.2408.15240 Generative verifiers: Reward modeling as next-token prediction . CoRR, abs/2408.15240

    2024 doi arXiv

  24. [32]

    Le, and Ed H

    Denny Zhou, Nathanael Sch \" a rli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc V. Le, and Ed H. Chi. 2023. https://openreview.net/forum?id=WZH7099tgfM Least-to-most prompting enables complex reasoning in large language mo...

    2023

  25. [33]

    Zhi Zhou, Tan Yuhao, Zenan Li, Yuan Yao, Lan-Zhe Guo, Xiaoxing Ma, and Yu-Feng Li. 2025. https://arxiv.org/abs/2502.00511 Bridging internal probability and self-consistency for effective and efficient llm reasoning . Preprint, arXiv:2502.00511

    2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.