Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

DynamicMind: A Tri-Mode Thinking System for Large Language Models

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 A cheap router chooses among fast, normal, and slow thinking per question, keeping accuracy high while cutting token use.

desk verdict A useful token-saving heuristic undermined by overclaimed accuracy and unreported filtering in the training data. read the letter →

arxiv 2506.05936 v1 pith:GP3ME7KH submitted 2025-06-06 cs.CL cs.AI

Wei Li , Yanbin Wei , Qiushi Huang , Jiangyue Yan , Yang Chen , James T. Kwok , Yu Zhang This is my paper
classification cs.CLcs.AI
keywords largelanguagemodelszero-shotquestionansweringthinkingmodeschain-of-thoughtpromptingadaptiveroutingtokenefficiencyDensityreasoningdepth
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 DynamicMind, a system that lets a frozen large language model answer zero-shot questions in one of three thinking modes: fast, normal, or slow. The central claim is that a lightweight router can predict which mode yields the best accuracy-per-token trade-off for each question, so the whole system stays close to the accuracy of the most careful mode while spending far fewer tokens. To build the router, the paper defines a Thinking Density score, labels roughly 39,000 training questions with their best mode, and proves that choosing the best mode per question is Pareto-optimal against any fixed mode. If the claim holds, adaptive reasoning depth can be added to an existing LLM with only three prompt templates and a small classifier, with no fine-tuning of the LLM itself. The paper also argues that the native normal mode carries much of the trade-off, since removing it either costs substantial accuracy or greatly increases token use.

What carries the argument

The load-bearing object is the Thinking Density metric $E_k^m(q)=\frac{\text{accuracy}_k^m(q)}{(\text{avg.tok}_k^m(q))^\alpha}$, which collapses the two goals of answer correctness and token cost into one scalar, with $\alpha$ setting the relative weight. The TMC dataset is built by evaluating each question in all three modes $k$ times, applying filters to keep only reliable labels, and assigning the mode with maximum Thinking Density; the Mind Router is then trained with cross-entropy to predict that mode from the question text. The router's decision rule is exactly $\arg\max_m E_k^m(q)$, and the paper's theorem shows that this rule Pareto-dominates any single fixed mode in expectation over questions.

What would settle it

Regenerate the TMC dataset without the 80-percent-accuracy and length-ordering filters, retrain the router, and compare DynamicMind's accuracy and token use against an always-slow policy on the hardest test questions; if the filtered router is no better than always-slow on those questions while using no fewer tokens, the central trade-off claim collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is that no single reasoning mode is best for every question, and that a per-question choice among three prompt-defined modes can beat any fixed mode on the accuracy/efficiency trade-off. The three modes are Fast (immediate, intuitive answers), Normal (the model's native instruction-following behavior), and Slow (decomposed, verified step-by-step reasoning). The paper labels each training question with the mode that maximizes Thinking Density, $E_k^m(q)=\frac{\text{accuracy}_k^m(q)}{(\text{avg.tok}_k^m(q))^\alpha}$, trains a lightweight classifier to predict that label, and proves that the ideal arg-max router is Pareto-optimal against every fixed mode. In experiments the routed system keeps accuracy close to the most accurate single mode while cutting token consumption sharply; for example, on one tested 8B model the average Thinking Density rises from 0.27 with chain-of-thought prompting to 1.33 with DynamicMind.

Load-bearing premise

The load-bearing premise is that the questions used to train the router, those surviving the filter that every mode score at least 80 percent and that response lengths come out in the order fast, normal, slow, are representative enough of all test questions, including hard and out-of-domain ones; if many questions were discarded, the learned mode preferences may not transfer.

Editorial extensions

If this is right

  • With DynamicMind, a frozen LLM can approach the accuracy of its slowest, most careful prompting mode on hard questions while using fast-mode-level tokens on easy ones, all without fine-tuning the LLM.
  • The Pareto optimality theorem implies that no fixed prompt style can improve both expected accuracy and expected token efficiency at once; changing $\alpha$ simply moves the system along that frontier.
  • On the reported benchmarks, the routed system's average Thinking Density is several times that of chain-of-thought prompting (for example, 1.33 vs 0.27 for one 8B model), quantifying the accuracy-per-token gain.
  • The normal mode is not a neutral third option: ablating it drops average accuracy by about 10 points for one of the tested models, so the native mode carries much of the system's robustness.
  • Mind Routers transfer between LLMs with only about 2 percent average accuracy loss, but token counts fluctuate, so the router is partly model-specific and benefits from retraining.

Reading between the lines

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

  • Because the TMC dataset discards questions where all modes score below 80 percent and where response lengths violate fast, then normal, then slow, the router may be optimized for questions the model can already solve; the paper does not report the size of the discarded fraction, so the reported gains may not extend to the hardest tail.
  • With $\alpha=1$, Thinking Density is simply accuracy per token, so the same training pipeline could generate a family of routers for different latency budgets by changing $\alpha$ only.
  • Since training labels come from one frozen model's own outputs, a stronger base model or a changed prompt set would require regenerating the TMC labels; the router does not learn a model-independent notion of difficulty.
  • The router could in principle be combined with other efficiency techniques, such as token-budget constraints or answer self-consistency, to push the Pareto frontier further than any single technique alone.
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 DynamicMind, a tri-mode thinking framework for zero-shot question answering. It adds a 'normal' thinking mode to the familiar fast/slow dichotomy, defines a Thinking Density metric that combines accuracy and token consumption, constructs a Thinking Mode Capacity (TMC) dataset by labeling questions with the mode that maximizes that metric, and trains a lightweight DeBERTa-based Mind Router to predict the preferred mode. The system is evaluated on mathematical, commonsense, and scientific QA benchmarks, with the authors claiming superior accuracy and a favorable accuracy-efficiency trade-off compared with CoT, PBC, and TALE-EP baselines.

Significance. If the central claim were supported, a lightweight router that selects among fast, normal, and slow reasoning prompts would be practically valuable, and the Thinking Density metric plus the TMC dataset could be useful resources for adaptive-reasoning research. The paper includes experiments across multiple benchmarks and two base LLMs, a transferability study, and a commitment to release code and checkpoints. However, as I detail below, the empirical results do not support the 'superior ZSQA' claim, the training/evaluation setup has a selection-bias problem, and the Pareto-optimality theorem is not proven. These issues are load-bearing rather than cosmetic.

major comments (4)
  1. [Abstract, §4.2, Tables 2 and 3] The abstract's claim of 'superior ZSQA capabilities' is contradicted by the paper's own results. In Table 2, DynamicMind with Llama-3.1 reaches an average accuracy of 62.35%, below PBC's 66.33%; with Qwen-2.5 it reaches 77.54%, below vanilla CoT's 79.11%. The §4.2 statement that DynamicMind delivers 'accuracy comparable to the most accurate mode (e.g., Slow Mode)' is also not supported: in Table 3, Llama DynamicMind is 5.26 points below Slow-Only (62.35 vs 67.61), and Qwen DynamicMind is 1.36 points below Slow-Only (77.54 vs 78.90). The paper can claim token savings and a particular accuracy/efficiency trade-off, but it cannot claim accuracy superiority relative to the baselines it reports.
  2. [Appendix B, Tables 2 and 3] The TMC construction filters out questions where all three thinking modes achieve below 80% accuracy and where the fast/normal/slow length ordering is violated, yet the paper reports no removal counts. The Mind Router is trained only on the surviving subset, while Tables 2 and 3 evaluate it on full test sets, including hard and out-of-domain questions. If a large fraction of questions is removed—likely for math, where mean accuracy is around 65% for Llama—the router never learns from hard questions during training, so its test-time choices on those questions are unlearned. The reported Thinking Density gains may therefore be an artifact of training on an easy, filtered subset. The authors should report per-dataset and per-model filter rates, evaluate on the filtered-out questions separately, and show that the router's decisions are not simply routing hard questions to fast or normal modes at the price of accuracy.
  3. [Eqs. (2), (3), (5), Tables 2 and 3] The evaluation is substantially circular with respect to the Thinking Density metric. The Mind Router is trained to predict the mode that maximizes Thinking Density (Eq. (3) and loss (5)), and the primary reported metric in Tables 2 and 3 is the same Thinking Density (Eq. (2)). It is therefore expected—almost by construction—that DynamicMind achieves higher TD than baselines that do not optimize this objective. The paper should treat accuracy and token consumption as the primary headline metrics and present TD as a secondary, derived quantity. Reporting accuracy and tokens separately is necessary but not sufficient; the discussion should acknowledge that TD gains are partly the router matching its own training labels, especially on in-domain test sets drawn from the same distribution as TMC.
  4. [Appendix C, Theorem 1] The proof of Pareto optimality is invalid. The theorem asserts that the per-question argmax of the weighted sum Acc + α·Eff dominates every fixed mode in both expected accuracy and expected efficiency (Eq. (7)). The proof shows only that E[Acc_R*] + α E[Eff_R*] ≥ E[Acc_m] + α E[Eff_m] pointwise, which does not imply both inequalities in Eq. (7). In the 'otherwise' branch, when E[Eff_R*] ≥ E[Eff_m], no argument is given for why E[Acc_R*] ≥ E[Acc_m]; indeed, a router that sometimes chooses a high-efficiency/low-accuracy mode can have lower expected accuracy than a fixed high-accuracy mode. The claimed Pareto dominance over all fixed modes is therefore unproven. A correct proof would need to establish the accuracy inequality separately, or the theorem should be weakened to a statement about the weighted objective.
minor comments (4)
  1. [Table 6, §4.4] The sensitivity analysis for α is internally inconsistent. The text says that setting α to 0 'improved accuracy but resulted in increased token consumption for both models,' but the table's Δ#Token↓ rows show large positive values for Llama at α=0 (e.g., +65.53 for Math), which conventionally means token consumption decreased. In addition, the Qwen α=0 rows appear to duplicate the Llama α=2 rows, which is likely a copy-paste error. Please clarify the sign convention and correct the table.
  2. [Table 2] The Qwen CoT row has missing spacing between reported values (e.g., '0.1980.64'), making the table difficult to read. Other rows in the same table and in Table 3 would also benefit from consistent alignment.
  3. [Throughout] There are several typographical and reference issues: 'distribustion' in Figure 5's caption, 'intristic' in §3.3, 'logrithm' in §3.5 and Appendix C, and the reference to 'Kahneman Daniel, 2017' should be Daniel Kahneman, Thinking, Fast and Slow (2011).
  4. [§4.3 and Table 4] The cross-model transferability results report only differences in accuracy and token counts without any variance or significance information. Given that several deltas are small (e.g., −0.26 on MMLU), it is unclear whether these differences are meaningful; please add confidence intervals or per-task breakdowns.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the router is trained on disjoint data and evaluated on held-out metrics; the main concern is a training-set filter, which is a generalization risk, not a circular reduction.

full rationale

DynamicMind's derivation chain is not circular. The Thinking Density metric (Eq. 2) is defined from measured accuracy and token counts, and the TMC labels (Eq. 3) are obtained by running the frozen LLM in the three modes; the Mind Router is then trained with cross-entropy on those labels (Eq. 5) and evaluated on held-out test splits and out-of-domain datasets (Appendix D). This is standard supervised learning, not a fitted parameter disguised as prediction: the test questions are disjoint from the TMC training subset, and the reported accuracy/token/TD numbers are measured outcomes of the router's choices, not values recovered from the training labels. The Pareto optimality theorem (Appendix C) is a scalarization argument about the ideal router R*; it does not smuggle the empirical results into the assumptions. The only significant concern is Appendix B's filtering rule (removing questions with <80% accuracy in all modes and questions whose response lengths violate fast<=normal<=slow) without reporting removal counts; this is a train/evaluation distribution-mismatch risk, not a circular reduction, and would be a correctness/generalization criticism rather than a circularity finding. No load-bearing self-citations, ansatz-by-citation, or uniqueness-by-self-citation steps were found. Score 0.

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

The central claim depends on the hand-chosen alpha, an 80% accuracy filter, a length-ordering constraint, and mode-specific token caps. It also assumes output token count is a valid cost proxy and that the normal mode prompt isolates native model behavior. The evaluation metric is the same objective used to create training labels.

free parameters (4)
  • alpha in Thinking Density (Eq. 2) = 1 for TMC; 0 and 2 in sensitivity
    Controls the accuracy-efficiency tradeoff in the metric that defines optimal mode labels; chosen by hand, not learned.
  • Accuracy filter threshold in TMC construction = 80%
    Questions where all three modes have accuracy below 80% are discarded; threshold selected by authors and not varied in sensitivity analysis.
  • Length-ordering constraint (fast <= normal <= slow) = binary constraint
    Discards questions not matching the assumed token-length ordering; this assumption enforces the tri-mode design but may bias the dataset.
  • Mode-specific token generation limits = 128 (fast), 2048 (normal), 4096 (slow)
    These caps constrain output length and thus directly affect token consumption, the efficiency component of Thinking Density.
assumptions (3)
  • domain assumption Output token count is a valid proxy for computational cost
    The paper's efficiency metric and all efficiency comparisons use response token count only; input prompt lengths differ across modes and are not counted, so the proxy may not reflect true cost.
  • domain assumption The 'normal mode' prompt preserves the model's native pretrained reasoning capability
    Section 3.3 asserts that a minimal instruction prompt captures the model's intrinsic behavior without interference; this is assumed, not verified against the model's training distribution.
  • domain assumption The three mode prompts induce distinct, stable reasoning behaviors independent of the question
    The framework relies on fast, normal, and slow prompts producing consistent accuracy and token profiles; the length-ordering filter in Appendix B is an attempt to enforce this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DynamicMind: A Tri-Mode Thinking System for Large Language Models." pith.science (2026). https://pith.science/paper/GP3ME7KH

@misc{pith2026250605936,
  author       = {Pith},
  title        = {Pith review of: DynamicMind: A Tri-Mode Thinking System for Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GP3ME7KH}},
  note         = {Machine review of arXiv:2506.05936}
}
read the original abstract

Modern large language models (LLMs) often struggle to dynamically adapt their reasoning depth to varying task complexities, leading to suboptimal performance or inefficient resource utilization. To address this, we introduce DynamicMind, a novel tri-mode thinking system. DynamicMind empowers LLMs to autonomously select between Fast, Normal, and Slow thinking modes for zero-shot question answering (ZSQA) tasks through cognitive-inspired prompt engineering. Our framework's core innovations include: (1) expanding the established dual-process framework of fast and slow thinking into a tri-mode thinking system involving a normal thinking mode to preserve the intrinsic capabilities of LLM; (2) proposing the Thinking Density metric, which aligns computational resource allocation with problem complexity; and (3) developing the Thinking Mode Capacity (TMC) dataset and a lightweight Mind Router to predict the optimal thinking mode. Extensive experiments across diverse mathematical, commonsense, and scientific QA benchmarks demonstrate that DynamicMind achieves superior ZSQA capabilities while establishing an effective trade-off between performance and computational efficiency.

Figures

Figures reproduced from arXiv: 2506.05936 by the authors.

Figure 1
Figure 1. Accuracy and efficiency trade-off for adopting [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the DynamicMind framework, where the Mind Router guides the LLM Thinker to use the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 5
Figure 5. The distribustion of TMC dataset. where Pm(y|q) is the probability distribution over responses y given q under mode m, Tm denotes the prompt template encoding the reasoning strategy for mode m, ⊕ denotes the concatenation opera￾tion, y<t = (y1, . . . , yt−1) are tokens generated before timestep t, and θ denotes the frozen parame￾ters of the LLM Thinker. To be specific, the three thinking modes are specifically desig… view at source ↗
Figures from the paper (1 more)
Figure 6
Figure 6. Figure 6: Illustrations of the cases where the Mind Router effectively selects the suitable thinking mode. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

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. Towards Concise and Adaptive Thinking in Large Reasoning Models: A Survey

    cs.AI 2025-07 conditional novelty 3.0 of 10

    A comprehensive review that categorizes methods for shortening and adaptively triggering chain-of-thought reasoning in large language models.

Reference graph

Works this paper leans on

44 extracted references · 12 canonical work pages · cited by 1 Pith paper

  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]

    Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, et al. 2024. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17682--17690

    2024

  4. [4]

    Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. 2020. Piqa: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 7432--7439

    2020

  5. [5]

    Jo Boaler. 2022. Mathematical mindsets: Unleashing students' potential through creative mathematics, inspiring messages and innovative teaching. John Wiley & Sons

    work page 2022

  6. [6]

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

    arXiv 2024

  7. [7]

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. Boolq: Exploring the surprising difficulty of natural yes/no questions. In NAACL

    2019

  8. [8]

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457

    arXiv 2018

Show all 44 references
  1. [9]

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

    2021 arXiv

  2. [10]

    Kahneman Daniel. 2017. Thinking, fast and slow

    2017

  3. [11]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  4. [12]

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

    1974

  5. [13]

    Shane Frederick. 2005. Cognitive reflection and decision making. Journal of Economic perspectives, 19(4):25--42

    2005

  6. [14]

    Xinyu Guan, Li Lyna Zhang, Yifei Liu, Ning Shang, Youran Sun, Yi Zhu, Fan Yang, and Mao Yang. 2025. https://arxiv.org/abs/2501.04519 rstar-math: Small llms can master math reasoning with self-evolved deep thinking . Preprint, arXiv:2501.04519

    2025 arXiv

  7. [15]

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

    2025 arXiv

  8. [16]

    Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. 2024. Token-budget-aware llm reasoning. arXiv preprint arXiv:2412.18547

    2024 arXiv

  9. [17]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021 a . Measuring massive multitask language understanding. In ICLR . OpenReview.net

    2021

  10. [18]

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

    2021

  11. [19]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199--22213

    2022

  12. [20]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437

    2024 arXiv

  13. [21]

    Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. 2022. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Sy...

    2022

  14. [22]

    Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. Can a suit of armor conduct electricity? a new dataset for open book question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2381--2391

    2018

  15. [23]

    Stephanie Nelli, Lukas Braun, Tsvetomira Dumbalska, Andrew Saxe, and Christopher Summerfield. 2023. Neural knowledge assembly in humans and neural networks. Neuron, 111(9):1504--1516

    2023

  16. [24]

    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

  17. [25]

    Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E Gonzalez, M Waleed Kadous, and Ion Stoica. 2024. Routellm: Learning to route llms from preference data. In The Thirteenth International Conference on Learning Representations

    2024

  18. [26]

    Jiabao Pan, Yan Zhang, Chen Zhang, Zuozhu Liu, Hongwei Wang, and Haizhou Li. 2024. Dynathink: Fast or slow? a dynamic decision-making framework for large language models. arXiv preprint arXiv:2407.01009

    2024 arXiv

  19. [27]

    Xiaoye Qu, Yafu Li, Zhaochen Su, Weigao Sun, Jianhao Yan, Dongrui Liu, Ganqu Cui, Daizong Liu, Shuxian Liang, Junxian He, et al. 2025. A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond. arXiv preprint arXiv:2503.21614

    2025

  20. [28]

    Nils Reimers and Iryna Gurevych. 2019. https://arxiv.org/abs/1908.10084 Sentence-bert: Sentence embeddings using siamese bert-networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics

    2019 arXiv

  21. [29]

    Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2021. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64(9):99--106

    2021

  22. [30]

    Maarten Sap, Hannah Rashkin, Derek Chen, Ronan LeBras, and Yejin Choi. 2019. Socialiqa: Commonsense reasoning about social interactions. arXiv preprint arXiv:1904.09728

    2019 arXiv

  23. [31]

    Yi Shen, Jian Zhang, Jieyun Huang, Shuming Shi, Wenjing Zhang, Jiangze Yan, Ning Wang, Kai Wang, and Shiguo Lian. 2025. Dast: Difficulty-adaptive slow-thinking for large reasoning models. arXiv preprint arXiv:2503.04472

    2025

  24. [32]

    DiJia Su, Sainbayar Sukhbaatar, Michael Rabbat, Yuandong Tian, and Qinqing Zheng. 2024. Dualformer: Controllable fast and slow thinking by learning with randomized reasoning traces. arXiv preprint arXiv:2410.09918

    2024 arXiv

  25. [33]

    Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Hanjie Chen, Xia Hu, et al. 2025. Stop overthinking: A survey on efficient reasoning for large language models. arXiv preprint arXiv:2503.16419

    2025 arXiv

  26. [34]

    Nicholas B Turk-Browne, Brian J Scholl, Marvin M Chun, and Marcia K Johnson. 2009. Neural evidence of statistical learning: Efficient detection of visual regularities without awareness. Journal of cognitive neuroscience, 21(10):1934--1945

    2009

  27. [35]

    Amos Tversky and Daniel Kahneman. 1974. Judgment under uncertainty: Heuristics and biases: Biases in judgments reveal some heuristics of thinking under uncertainty. science, 185(4157):1124--1131

    1974

  28. [36]

    Peifeng Wang, Zhengyang Wang, Zheng Li, Yifan Gao, Bing Yin, and Xiang Ren. 2023. Scott: Self-consistent chain-of-thought distillation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5546--5558

    2023

  29. [37]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171

    2022 arXiv

  30. [38]

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

    2022

  31. [39]

    Yuyang Wu, Yifei Wang, Tianqi Du, Stefanie Jegelka, and Yisen Wang. 2025. When more is less: Understanding chain-of-thought length in llms. arXiv preprint arXiv:2502.07266

    2025 arXiv

  32. [40]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Me...

    2024 arXiv

  33. [41]

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2024. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36

    2024

  34. [42]

    Ping Yu, Jing Xu, Jason E Weston, and Ilia Kulikov. 2024. https://openreview.net/forum?id=WUoC4BpJBC Distilling system 2 into system 1 . In The First Workshop on System-2 Reasoning at Scale, NeurIPS'24

    2024

  35. [43]

    Xiaosong Yuan, Chen Shen, Shaotian Yan, Xiaofeng Zhang, Liang Xie, Wenxiao Wang, Renchu Guan, Ying Wang, and Jieping Ye. 2024. https://proceedings.neurips.cc/paper_files/paper/2024/file/e304e04a6f455dd82f8a85a0a3679493-Paper-Conference.pdf Instance-adaptive zero-shot chain-of-...

    2024

  36. [44]

    Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. Hellaswag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791--4800

    2019

Pith tools

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