Pith. sign in

REVIEW 4 major objections 6 minor 5 cited by

ThinkSwitcher: When to Think Hard, When to Think Fast

T0 review · 4 major / 6 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 ThinkSwitcher claims that a single large reasoning model can cut inference token use by 20–30% by learning to choose between short and long chain-of-thought per question, with only a small accuracy drop.

desk verdict A clean, useful routing method with broad experiments, but the undertested query embedding and missing error bars keep the headline numbers from being fully reproducible. read the letter →

arxiv 2505.14183 v1 pith:I6D57CIT submitted 2025-05-20 cs.CL

Guosheng Liang , Longguang Zhong , Ziyi Yang , Xiaojun Quan This is my paper
classification cs.CL
keywords largereasoningmodelschain-of-thoughtadaptivedepthtokenefficiencypass-ratesupervisionroutingoverthinkingmathbenchmarks
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

The paper argues that large reasoning models already contain a latent ability to reason concisely, and that this ability can be switched on with a prompt—an empty block—without retraining the model. On top of that observation, it proposes ThinkSwitcher, a lightweight module that reads a question embedding and predicts the pass rate the model would achieve under short and long chain-of-thought prompting, then routes each question to the path expected to do better. The switcher is trained with self-supervised labels: for each training question, pass rates are estimated by sampling multiple responses in both modes. Across three sizes (1.5B, 7B, 14B) of the DeepSeek-R1-Distill-Qwen family and seven math benchmarks, ThinkSwitcher is reported to reduce average output tokens by 20–30% compared with always-long reasoning while losing only 1–2% average accuracy, with negligible switcher overhead. If this holds, a single deployed model could serve both easy and hard questions efficiently, replacing dual-model deployments that keep separate fast and slow reasoners.

What carries the argument

The switching mechanism is a five-layer multi-layer perceptron (MLP) that takes the reasoning model's query embedding $x_q$ and outputs two predicted pass rates $\hat{y}_{SC}$ and $\hat{y}_{LC}$; the routing rule sends a query to long CoT only when $\hat{y}_{LC} - \hat{y}_{SC}$ exceeds a tunable threshold $\tau$. Training uses a margin-aware objective $L_{\text{switch}} = L_{\text{MSE}} + \lambda_{\text{margin}} L_{\text{margin}}$, where the margin term supervises the predicted difference directly, and the supervision targets are empirical pass rates computed from $k=8$ sampled responses per mode. The other load-bearing piece is the prompt intervention: the empty <think></think> block that elicits short CoT without modifying backbone weights.

What would settle it

On a held-out benchmark, compute the rank correlation between the switcher's predicted margin $\hat{y}_{LC} - \hat{y}_{SC}$ and the true pass-rate gap $y_{LC} - y_{SC}$; if the correlation is near zero, the router is no better than random at choosing which questions deserve long thinking. Equivalently, compare accuracy on the questions ThinkSwitcher routes to long CoT against the same number of randomly selected questions at the same token budget: if the routed subset shows no accuracy advantage, the central claim collapses.

Watch

Extended reading notes

Core claim

The central discovery is that a single large reasoning model can adapt its reasoning depth per query without any weight change. Short chain-of-thought is induced reliably by appending an empty think block to the prompt, and a small regressor trained on the model's own pass rates can decide between the short and long paths from the query embedding alone. The paper reports that on GSM8K, MATH-500, AIME 2024 and 2025, LiveAoPSBench, Omni-MATH-500, and the math subset of OlympiadBench, ThinkSwitcher produces a better accuracy–cost frontier than always-long, always-short, random, and BERT-based routing, cutting token counts by roughly 20–30% with a 1–2% average accuracy drop and reducing decoding FLOPs by a similar margin.

Load-bearing premise

The load-bearing premise is that the question embedding fed to the switcher carries enough information about task difficulty and the relative benefit of long versus short reasoning that the regressor's predicted pass-rate margins are trustworthy, since the paper does not specify which hidden state or pooling produces the embedding or analyze prediction failures.

Editorial extensions

If this is right

  • A single model deployment can replace dual-model systems that keep one fast and one slow reasoner, removing extra memory and serving cost while keeping complex-task accuracy.
  • The reported 20–30% token reduction translates to a similar reduction in decoding FLOPs, while the switcher itself adds only millions of FLOPs against trillions for decoding.
  • Smaller and weaker models gain the most from switching because they are more prone to over-elaboration, so the benefit scales with deployment pressure.
  • The empty-think-block prompt is a zero-training intervention that works across several reasoning-model families and can be adopted immediately as a cheap efficiency baseline.
  • Because pass-rate labels come from the backbone model's own responses, the same self-supervised pipeline can regenerate training data for new benchmarks without human annotation.

Reading between the lines

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

  • A testable consequence of the paper's design is that the ceiling of the method is set by how much difficulty information the query embedding carries; on out-of-distribution questions where that embedding is uninformative, the router should degrade toward random selection.
  • The same switching idea could be carried over to code generation or agentic tasks, but the pass-rate gap between short and long reasoning will be harder to estimate there because correctness is less easily verifiable.
  • The margin-aware loss suggests a general design principle for learned routers: when the decision depends on a predicted difference, the training objective should supervise that difference directly, not just its components.
  • A natural extension is to use the predicted margin to interpolate reasoning effort continuously rather than choosing between two discrete modes, which could smooth the accuracy-cost frontier further.
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 / 6 minor

Summary. The paper proposes ThinkSwitcher, a framework that lets a single large reasoning model (DeepSeek-R1-Distill-Qwen-1.5B/7B/14B) dynamically choose between a long chain-of-thought (LC) and a prompt-induced short chain-of-thought (SC) mode. The core idea is to train a lightweight MLP regressor on self-supervised labels: for each training question, the pass rates of SC and LC are estimated by sampling k responses, and the regressor predicts those pass rates from a query embedding xq. At inference, LC is selected when the predicted LC pass rate exceeds the predicted SC pass rate by a tunable threshold τ. Experiments on seven math benchmarks (GSM8K, MATH500, AIME24, AIME25, LiveAoPS, Omni-MATH-500, OlympiadBench) report that ThinkSwitcher reduces average output tokens by roughly 20–30% relative to LC-Only, with a 1–2% average accuracy drop, and that it outperforms random mixing and a ModernBERT-based router on the accuracy–cost frontier. The authors also provide a prompt-design observation (empty <think> block suffices to induce short CoT), an ablation of the margin loss, an analysis of the pass-rate sampling count k, and an estimate of the switcher's negligible FLOP overhead.

Significance. If the central claims hold, ThinkSwitcher is a useful practical contribution: it addresses the overthinking problem of LRMs at inference time without backbone modification or large-scale post-training, and it introduces a margin-aware, self-supervised training objective for routing between reasoning modes. The empirical scope is real: three model sizes, seven math benchmarks, and a comparison against random and BERT-based routers. The paper is also honest about several limitations (math-only evaluation, up to 14B parameters). The strongest assets are the clean prompt-based short-CoT induction, the explicit formulation of pass-rate regression targets, and the cost-overhead table that shows the router's computational impact is negligible. However, the manuscript currently leaves a key component underspecified—the query embedding xq—and the headline numbers rest on test-set-selected thresholds without error bars, which weakens confidence in the claimed accuracy–efficiency frontier.

major comments (4)
  1. [§4.2] The query embedding xq is never defined: the paper says only that it is 'extracted from the reasoning model,' without specifying the hidden layer, the token position, or the pooling operation. This is the load-bearing input to the router, and the entire per-query routing claim depends on it. The authors should specify exactly how xq is computed, and provide an analysis of router prediction quality (e.g., correlation between predicted pass-rate margin and realized margin, calibration plots, and failure cases). Without this, the 20–30% token reduction at 1–2% accuracy loss cannot be attributed to learned per-query selection.
  2. [§5.2 and Table 1] The main results use decision thresholds τ selected on the test benchmarks themselves ('selected to achieve a favorable balance,' τ = 0.04/0.05/0.03 for 1.5B/7B/14B), while the Random baselines are fixed-probability policies that are not tuned on the test set. This asymmetry can inflate ThinkSwitcher's apparent advantage. The authors should either select τ on a held-out validation set and report the resulting test performance, or report the full τ-sweep curves with a clear statement of which operating points are chosen and how. The current presentation mixes model selection with evaluation.
  3. [Table 1 and §6.1] No variance or confidence intervals are reported for any accuracy or token count, and several test sets are small (e.g., AIME24/AIME25 have only 30 questions each per year). Single-run numbers on such sets can easily move by several points, and the claimed 1–2% accuracy drop and the per-benchmark token savings are not supported without error bars or multiple seeds. The authors should provide at least three independent generations per setting (or bootstrap confidence intervals) for the headline accuracy and token metrics.
  4. [Abstract and §7] The paper claims a general solution for 'complex tasks' and positions ThinkSwitcher as a unified LRM deployment framework, but all experiments are on mathematics benchmarks. The limitations section acknowledges this, but the abstract and conclusion do not qualify the claim. Either the scope should be explicitly narrowed in the abstract/conclusion to mathematical reasoning, or at least one non-math benchmark (e.g., code, general knowledge) should be added to support the generality.
minor comments (6)
  1. [Table 1] The 14B ThinkSwitcher row is garbled: '42.5 1006565.8 6018' appears to be a formatting corruption, likely splitting the AIME25 accuracy and token count. Please fix the table typesetting and verify all numbers.
  2. [Appendix C] The mechanism analysis (cosine similarity ranking) is presented as supporting a 'suppression-and-reversion' hypothesis, but the three similarity values 0.926, 0.919, 0.916 are extremely close and no statistical test or confidence interval is given. Please soften the causal interpretation or add a proper significance test.
  3. [§4.1] The labels y_m(q) are called self-supervised, but they require executing the backbone model and checking correctness of responses; this is better described as self-generated supervision rather than fully self-supervised in the usual sense. Please clarify the terminology.
  4. [§6.2] The margin-loss ablation in Table 2 reports only nAUC-AC. Reporting the final accuracy and token counts for λmargin ∈ {0, 1, 2} would make the effect more transparent.
  5. [§6.3] Figure 5 shows nAUC-AC as a function of k, but the text says 'approximately 174 at k = 1' while the figure appears to begin around 174–176. Please align the text and figure values, and state the random-seed handling for the sampled responses.
  6. [Appendix D] The hyperparameter search uses Ray Tune, but the final hyperparameter values (learning rate, batch size, dropout) are not reported; please list the selected configuration for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the router's self-generated labels and empirical trade-off curves do not reduce to the paper's inputs by construction.

full rationale

ThinkSwitcher's derivation is an empirical pipeline rather than a circular one. The switcher is trained to regress empirical pass rates of short-CoT and long-CoT prompted responses generated by the same backbone (Eqs. 2-3), but this self-supervised labeling scheme does not define the reported accuracy or token savings: a poor regressor would produce poor routing, so the 20-30% token reduction at 1-2% accuracy loss is a measured outcome, not a consequence of the training objective. The decision rule (Eq. 5) compares predicted pass-rate margins against a threshold, and the margin loss (Eq. 7) supervises that margin; neither equation assumes the final trade-off. The Appendix C 'mechanism' analysis is explicitly presented as a hypothesis and supported by cosine-similarity measurements, not used as a derivation. There are no load-bearing self-citations or imported uniqueness theorems; the cited prior work (e.g., Ma et al. 2025, Qwen3) is external and used for context. The main validity caveats—the unspecified query embedding xq in Section 4.2 and the test-set selection of tau in Section 5.2—are evaluation and reproducibility concerns, not circular reductions; the reported numbers could in principle have been unfavorable at the chosen operating point. The Limitations section acknowledges scope limits (math only, models up to 14B) but does not assert or conceal a definitional dependency. Overall, the central claim is self-contained and empirically falsifiable, so the circularity score is 0.

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

The paper's central claim rests on one known phenomenon (empty think block) plus the router's ability to predict label differences from query embeddings. The main free parameters are the threshold, margin loss weight, and sampling count.

free parameters (3)
  • Decision threshold τ = 0.04 (1.5B), 0.05 (7B), 0.03 (14B)
    Chosen per model in Section 5.2 to achieve a favorable accuracy-cost balance; directly controls the fraction of questions routed to long CoT and hence the headline token savings.
  • Margin loss weight λmargin = 1
    Selected from {0, 1, 2} via ablation in Section 6.2; affects switcher training objective.
  • Pass-rate sampling count k = 8
    Chosen based on saturation in Figure 5; affects label variance during training data construction.
assumptions (4)
  • domain assumption Prompting with an empty <think> block reliably induces short chain-of-thought with acceptable accuracy in LRMs.
    Core observation in Section 3.1 and 3.2; relies on the model's latent short-reasoning capability. If this is not robust, the switch has no efficient mode to route to.
  • ad hoc to paper The query embedding xq extracted from the LRM captures task difficulty relevant to mode selection.
    Section 4.2 uses this embedding as the only feature; the paper does not specify how it is computed or verify its sufficiency.
  • domain assumption Empirical pass rates over k=8 samples are a faithful estimate of mode performance.
    Used as regression targets in Section 4.1; noise in these labels limits switcher quality.
  • domain assumption The evaluation benchmarks do not overlap with training data.
    Claimed in Section 5.1; contamination would inflate the results. We have no independent verification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ThinkSwitcher: When to Think Hard, When to Think Fast." pith.science (2026). https://pith.science/paper/I6D57CIT

@misc{pith2026250514183,
  author       = {Pith},
  title        = {Pith review of: ThinkSwitcher: When to Think Hard, When to Think Fast},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I6D57CIT}},
  note         = {Machine review of arXiv:2505.14183}
}
read the original abstract

Large reasoning models (LRMs) excel at solving complex tasks by leveraging long chain-of-thought (CoT) reasoning. However, this often leads to overthinking on simple tasks, resulting in unnecessary computational overhead. We observe that LRMs inherently possess the capability for efficient short CoT reasoning, which can be reliably elicited through prompt design. To leverage this capability, we propose ThinkSwitcher, a framework that enables a single LRM to dynamically switch between short and long CoT modes based on task complexity. ThinkSwitcher introduces a lightweight switching module trained with supervision signals derived from the relative performance of each reasoning mode across tasks. Experiments on multiple reasoning benchmarks show that ThinkSwitcher reduces computational cost by 20-30% while maintaining high accuracy on complex tasks. This demonstrates the effectiveness of ThinkSwitcher as a scalable and efficient solution for unified LRM deployment.

Figures

Figures reproduced from arXiv: 2505.14183 by the authors.

Figure 1
Figure 1. Comparison of long and short CoTs generated [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of long CoT and induced short CoT on the MATH500. “R1” denotes DeepSeek-R1-Distill [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Dynamic mode selection during infer￾ence. Given a question embedding from the LRM, ThinkSwitcher dynamically chooses between short and long CoT reasoning based on estimated task difficulty. 32B), Sky-T1-32B, and Qwen3-8B. Additional fig￾ures illustrating these trends on other benchmarks are provided in Appendix A. This reduction in out￾put length implies improved computational effi￾ciency, potentially lowering FLOPs… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Trade-off between average accuracy and cost (measured by average output tokens) for the three DeepSeek [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Performance with different values of k used to estimate pass rates in training data construction. 6.3 Impact of Pass Rate Estimation Quality The switcher is trained using pass rates estimated over k sampled responses per pathway (SC and LC) for each training instance. …
Figure 8
Figure 8. Figure 8: Long and short CoT prompt template. As shown in [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 6
Figure 6. Figure 6: Comparison of long CoT and induced short CoT on the AIME24. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Comparison of long CoT and induced short CoT on the AIME25. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 9
Figure 9. Figure 9: Trade-off between average accuracy and cost for ThinkSwitcher across various datasets (Part 1 of 2). [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Trade-off between average accuracy and cost for ThinkSwitcher across various datasets (Part 2 of 2). [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

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

  1. Compress the Easy, Explore the Hard: Difficulty-Aware Entropy Regularization for Efficient LLM Reasoning

    cs.LG 2026-02 conditional novelty 6.0 of 10

    CEEH selectively applies entropy regularization to hard questions and a shortest-correct-length penalty to easy ones, reducing reasoning length while preserving accuracy across six math benchmarks.

  2. Scaling Up, Speeding Up: A Benchmark of Speculative Decoding for Efficient LLM Test-Time Scaling

    cs.CL 2025-08 conditional novelty 6.0 of 10

    N-gram based speculative decoding methods, especially SAM and hybrid SAM[EAGLE-3], achieve strong speedups in LLM test-time scaling by exploiting repetitive reasoning patterns.

  3. Hierarchical Budget Policy Optimization for Adaptive Reasoning

    cs.AI 2025-07 conditional novelty 6.0 of 10

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

  4. How Far Are We from Optimal Reasoning Efficiency?

    cs.CL 2025-06 conditional novelty 6.0 of 10

    The authors define a reasoning efficiency frontier and a gap metric (REG), then train models with REO-RL to shrink the gap by at least 50% with only small accuracy losses.

  5. 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

51 extracted references · 12 canonical work pages · cited by 5 Pith papers

  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]

    Abien Fred Agarap. 2019. Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375

    arXiv 2019

  4. [4]

    Anthropic. 2025. https://api.semanticscholar.org/CorpusID:276612236 Claude 3.7 sonnet system card

    work page 2025

  5. [5]

    Akhiad Bercovich, Itay Levy, Izik Golan, Mohammad Dabbah, Ran El-Yaniv, Omri Puny, Ido Galil, Zach Moshe, Tomer Ronen, Najeeb Nabwani, and 1 others. 2025. Llama-nemotron: Efficient reasoning models. arXiv preprint arXiv:2505.00949

    arXiv 2025

  6. [6]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, and 1 others. 2021. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374

    arXiv 2021

  7. [7]

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

    arXiv 2024

  8. [8]

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

    arXiv 2021

Show all 51 references
  1. [9]

    Codeforces. 2025. https://codeforces.com/ Codeforces - competitive programming platform . Accessed: 2025-03-18

    2025

  2. [10]

    Alejandro Cuadron, Dacheng Li, Wenjie Ma, Xingyao Wang, Yichuan Wang, Siyuan Zhuang, Shu Liu, Luis Gaspar Schroeder, Tian Xia, Huanzhi Mao, and 1 others. 2025. The danger of overthinking: Examining the reasoning-action dilemma in agentic tasks. arXiv preprint arXiv:2502.08235

    2025 arXiv

  3. [11]

    Di Zhang . 2025. https://doi.org/10.57967/hf/4687 Aime\_1983\_2024 (revision 6283828)

    2025 doi

  4. [12]

    Kanishk Gandhi, Ayush Chakravarthy, Anikait Singh, Nathan Lile, and Noah D Goodman. 2025. Cognitive behaviors that enable self-improving reasoners, or, four habits of highly effective stars. arXiv preprint arXiv:2503.01307

    2025 arXiv

  5. [13]

    Bofei Gao, Feifan Song, Zhe Yang, Zefan Cai, Yibo Miao, Qingxiu Dong, Lei Li, Chenghao Ma, Liang Chen, Runxin Xu, Zhengyang Tang, Benyou Wang, Daoguang Zan, Shanghaoran Quan, Ge Zhang, Lei Sha, Yichang Zhang, Xuancheng Ren, Tianyu Liu, and Baobao Chang. 2025. Omni- MATH : A un...

    2025

  6. [14]

    Google. 2025. https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/ Gemini 2.5 pro

    2025

  7. [15]

    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

  8. [16]

    Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Yuxiang Zhang, and 1 others. 2024. Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems. In Proceedings of ...

    2024

  9. [17]

    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. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track

    2021

  10. [18]

    Wenyue Hua and Yongfeng Zhang. 2022. System 1+ system 2= better world: Neural-symbolic chain of logic reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 601--612

    2022

  11. [19]

    Sergey Ioffe and Christian Szegedy. 2015. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning

    2015

  12. [20]

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

    2024 arXiv

  13. [21]

    Daniel Kahneman. 2003. Maps of bounded rationality: Psychology for behavioral economics. American economic review, 93(5):1449--1475

    2003

  14. [22]

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

    2011

  15. [23]

    Yu Kang, Xianghui Sun, Liangyu Chen, and Wei Zou. 2024. C3ot: Generating shorter chain-of-thought without compromising effectiveness. arXiv preprint arXiv:2412.11664

    2024 arXiv

  16. [24]

    Abhinav Kumar, Jaechul Roh, Ali Naseh, Marzena Karpinska, Mohit Iyyer, Amir Houmansadr, and Eugene Bagdasarian. 2025. Overthink: Slowdown attacks on reasoning llms. arXiv preprint arXiv:2502.02542

    2025

  17. [25]

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Princip...

    2023

  18. [26]

    Zhong-Zhi Li, Duzhen Zhang, Ming-Liang Zhang, Jiaxin Zhang, Zengyan Liu, Yuxuan Yao, Haotian Xu, Junhao Zheng, Pei-Jie Wang, Xiuyi Chen, and 1 others. 2025. From system 1 to system 2: A survey of reasoning large language models. arXiv preprint arXiv:2502.17419

    2025 arXiv

  19. [27]

    Gonzalez, and Ion Stoica

    Richard Liaw, Eric Liang, Robert Nishihara, Philipp Moritz, Joseph E. Gonzalez, and Ion Stoica. 2018. Tune: A research platform for distributed model selection and training. arXiv preprint arXiv:1807.05118

    2018 arXiv

  20. [28]

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let's verify step by step. In The Twelfth International Conference on Learning Representations

    2024

  21. [29]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations

    2019

  22. [30]

    Haotian Luo, Li Shen, Haiying He, Yibo Wang, Shiwei Liu, Wei Li, Naiqiang Tan, Xiaochun Cao, and Dacheng Tao. 2025. O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning. arXiv preprint arXiv:2501.12570

    2025 arXiv

  23. [31]

    Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. 2025. Reasoning models can be effective without thinking. arXiv preprint arXiv:2504.09858

    2025 arXiv

  24. [32]

    MAA. 2025. https://maa.org/maa-invitational-competitions/ American invitational mathematics examination - aime

    2025

  25. [33]

    Sadegh Mahdavi, Muchen Li, Kaiwen Liu, Christos Thrampoulidis, Leonid Sigal, and Renjie Liao. 2025. Leveraging online olympiad-level math problems for llms training and contamination-resistant evaluation. arXiv preprint arXiv:2501.14275

    2025 arXiv

  26. [34]

    Yu Meng, Mengzhou Xia, and Danqi Chen. 2024. Sim PO : Simple preference optimization with a reference-free reward. In Advances in Neural Information Processing Systems

    2024

  27. [35]

    Gonzalez, M Waleed Kadous, and Ion Stoica

    Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M Waleed Kadous, and Ion Stoica. 2025. Route LLM : Learning to route LLM s from preference data. In The Thirteenth International Conference on Learning Representations

    2025

  28. [36]

    Xiaoye Qu, Yafu Li, Zhaochen Su, Weigao Sun, Jianhao Yan, Dongrui Liu, Ganqu Cui, Daizong Liu, Shuxian Liang, Junxian He, and 1 others. 2025. A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond. arXiv preprint arXiv:2503.21614

    2025

  29. [37]

    QwenTeam. 2025. Qwq-32b: Embracing the power of reinforcement learning. ://qwenlm.github.io/blog/qwq-32b/. Accessed: 6 March 2025

    2025

  30. [38]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems

    2023

  31. [39]

    Matthew Renze and Erhan Guven. 2024. https://doi.org/10.1109/fllm63129.2024.10852493 The benefits of a concise chain of thought on problem-solving in large language models . In 2024 2nd International Conference on Foundation and Large Language Models (FLLM), page 476–483. IEEE

    2024

  32. [40]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. ArXiv, abs/1707.06347

    2017 arXiv

  33. [41]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, and 1 others. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300

    2024 arXiv

  34. [42]

    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

  35. [43]

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929--1958

    2014

  36. [44]

    Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Hanjie Chen, and 1 others. 2025. Stop overthinking: A survey on efficient reasoning for large language models. arXiv preprint arXiv:2503.16419

    2025 arXiv

  37. [45]

    Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, and 1 others. 2025. Kimi k1. 5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599

    2025 arXiv

  38. [46]

    NovaSky Team. 2025. Sky-t1: Train your own o1 preview model within \ 450. https://novasky-ai.github.io/posts/sky-t1

    2025

  39. [47]

    Benjamin Warner, Antoine Chaffin, Benjamin Clavié, Orion Weller, Oskar Hallström, Said Taghadouini, Alexis Gallagher, Raja Biswas, Faisal Ladhak, Tom Aarsen, Nathan Cooper, Griffin Adams, Jeremy Howard, and Iacopo Poli. 2024. Smarter, better, faster, longer: A modern bidirecti...

    2024 arXiv

  40. [48]

    Heming Xia, Yongqi Li, Chak Tou Leong, Wenjie Wang, and Wenjie Li. 2025. Tokenskip: Controllable chain-of-thought compression in llms. arXiv preprint arXiv:2502.12067

    2025

  41. [49]

    Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. 2025. Chain of draft: Thinking faster by writing less. arXiv preprint arXiv:2502.18600

    2025 arXiv

  42. [50]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388

    2025 arXiv

  43. [51]

    Weihao Zeng, Yuzhen Huang, Qian Liu, Wei Liu, Keqing He, Zejun Ma, and Junxian He. 2025. Simplerl-zoo: Investigating and taming zero reinforcement learning for open base models in the wild. arXiv preprint arXiv:2503.18892

    2025 arXiv

Pith tools

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