Pith. sign in

REVIEW 3 major objections 6 minor 3 cited by

LAPO: Internalizing Reasoning Efficiency via Length-Adaptive Policy Optimization

T0 review · 3 major / 6 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 Two-stage RL lets a model set its own reasoning budget, cutting tokens up to 40.9% while raising accuracy 2.3 points.

desk verdict A genuinely useful two-stage length-control recipe whose flagship 'internalization' claim is not actually tested: the self-declared budget is inserted into the prompt during training and never shown to emerge at inference. read the letter →

arxiv 2507.15758 v2 pith:A3ACZQBP submitted 2025-07-21 cs.AI cs.CL

Xingyu Wu , Yuchen Yan , Shangke Lyu , Linjuan Wu , Yiwen Qiu , Yongliang Shen , Weiming Lu , Jian Shao
show 2 more authors
Jun Xiao Yueting Zhuang
This is my paper
classification cs.AIcs.CL
keywords length-adaptivereasoningchain-of-thoughtefficiencyreinforcementlearningGRPOtokenbudgetcontrolmathematicaloverthinkingself-declaredplan
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 tries to establish that a reasoning model can learn to control its own output length, deciding how much to think per problem instead of having length limits imposed from outside. The proposed Length-Adaptive Policy Optimization (LAPO) runs two reinforcement-learning stages: a Discovery stage in which length-aware rewards and the model's own correct rollouts produce a per-problem target length $M(q)$, the median successful solution length, and an Internalization stage in which that target is written into the model's reasoning context as a self-declared budget inside the `` block and adherence is rewarded. If the claim holds, deployment needs no external token budget, router, or early-stopping rule. On mathematical reasoning benchmarks the paper reports up to 40.9% fewer tokens together with a 2.3-point accuracy gain, and it presents evidence that models scale reasoning length with problem difficulty and extrapolate beyond lengths seen in training.

What carries the argument

The load-bearing object is the problem-to-length mapping $M(q)$ together with the self-declared budget sentence. $M(q)$ is the median token count of correct GRPO (Group Relative Policy Optimization) rollouts for question $q$, with 4096 as the default for unsolved questions; during Internalization it is updated so newly solved questions use their median while previously solved questions move toward $\min(M(q), \mathrm{Median}(L^{(t)}_q))$. The budget sentence `<think> I will answer the question with $n$ tokens.` sits immediately after the thinking delimiter, reframing an external length instruction as the model's own opening plan, and the Gaussian adherence reward $R_2$ trains the model to fulfill that self-proposal. In the Discovery stage, the percentile band $[P_{30}, P_{70}]$ and the linear decay $f(d)=\max(0,1-d/100)$ make the discovered lengths robust to outliers while keeping all efficiency rewards conditional on correctness.

What would settle it

Run the LAPO-I model on MATH-500 with prompts that contain no length text at all, and check whether it still emits a budget sentence after `<think>` and whether per-problem token counts still track the trained $M(q)$ values; if the budget sentence disappears or lengths revert to the base model's distribution, the claimed internalization is absent.

Watch

Extended reading notes

Core claim

The central claim is that reasoning-length control can be converted from an external constraint into an intrinsic model capability, and that a two-stage RL procedure is enough to do it. During Discovery, GRPO rollouts supply the lengths of correct responses; LAPO keeps the 30th to 70th percentile band of those lengths as the reasonable range and records $M(q) = \mathrm{Median}(L_q)$ as the target for each question, using 4096 as a fallback for questions no rollout solved. During Internalization, each prompt is augmented with `<think> I will answer the question with $n$ tokens.` and the Gaussian reward $R_2 = \exp(-(|r_i|-n)^2/(2\sigma^2))$ is granted only for correct answers, so the model is rewarded for making good on its own stated plan. The paper maintains that placing the budget statement inside `<think>` is what makes the model internalize it, and that the resulting policy self-declares a budget at inference, resists conflicting external length instructions, and allocates computation roughly in proportion to problem difficulty.

Load-bearing premise

The approach depends on the median length of the model's own correct solutions being a trustworthy measure of how much thinking each problem really needs, and on the model still announcing its own budget and following it at use time, when no length target is in the prompt.

Editorial extensions

If this is right

  • On the stronger 1.5B base model used in the paper, the full pipeline reduces average tokens from 6,229 to 3,832 (38.5%) while raising average accuracy by 2.3 points, and on the weaker base model it cuts tokens 40.9% with a 1.2-point accuracy gain.
  • Among the methods compared, covering external-budget controllers, history-aware rewards, and think/no-think routers, LAPO-I reaches the highest average accuracy at a low token count under the same 4k training context.
  • Ablations show that exact length targets beat ranges and that moving the budget sentence outside `<think>` drops accuracy and makes the model follow external length instructions, which the paper reads as evidence that guidance position creates genuine internalization.
  • Using the median successful length as the target outperforms the mean and the minimum, balancing accuracy against compression.
  • On the knowledge-intensive GPQA benchmark, LAPO-D improves Pass@1 by 2.0 to 2.2 points while cutting tokens by 19.4 to 26.2 percent, evidence that the learned allocation transfers beyond mathematics.

Reading between the lines

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

  • A direct extension would be to apply the same discover-then-internalize recipe to tasks with more variable reasoning depth, such as code repair or multi-step tool use, where the model's own successful traces could define per-query budgets.
  • A stricter test than the paper's robustness check would be to remove all length language from the deployment prompt and verify that the model still opens `<think>` with its own budget statement; if it does not, the claimed internalization may be a learned template rather than a plan.
  • The 4096-token fallback for unsolved problems is also the training context cap, so hard-problem targets may be systematically inflated; a difficulty-aware fallback could improve efficiency on the hardest items.
  • Because the trained model already scales length with difficulty, its self-declared budget could be reused as a sampling hint, for example drawing more rollouts for high-budget problems, coupling test-time compute with the model's own estimate of complexity.
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

3 major / 6 minor

Summary. The paper proposes Length-Adaptive Policy Optimization (LAPO), a two-stage reinforcement-learning method for reducing reasoning token usage without sacrificing accuracy. In the Discovery stage, the model is trained with GRPO using a length-aware reward that encourages correct solutions to fall within a percentile-based length range; the per-problem median correct length M(q) is recorded as a target. In the Internalization stage, the prompt is augmented with the literal string "I will answer the question with n tokens." immediately after <think>, and the model is rewarded for correct solutions whose length is close to n via a Gaussian-shaped reward. The authors claim this makes length control an intrinsic model capability, enabling inference-time self-declared budgets. On MATH-500, AIME2024, AMC23, and OlympiadBench, they report average token reductions of 38.5%–41.0% with accuracy gains of 1.2–2.3 points over the base models. The paper also includes ablations on guidance form, target statistic, training data, and a robustness test with conflicting external length instructions.

Significance. If the central internalization claim were verified, LAPO would be a practically valuable method: it achieves large token reductions with small accuracy gains on competitive math benchmarks, and the two-stage design is conceptually clean. The paper also provides an Acc-Only ablation that partially isolates the effect of the length reward, which is a useful comparison. The code and project repository are linked, and the empirical tables are extensive. However, the paper's headline contribution—that length control is internalized so that the model autonomously proposes and follows a reasoning budget at inference—is not supported by the reported experiments, because the training procedure never requires the model to generate the budget statement itself. The measured token reductions and accuracy levels are real empirical results, but the claimed mechanism requires additional verification before the paper's main contribution can be accepted.

major comments (3)
  1. [§3.3, Eq. (6), Algorithm 1 line 21, §6.4] The internalization mechanism is not verified at inference. During training, every prompt contains the fixed prefix "I will answer the question with n tokens." after <think>, so the policy is optimized to continue a sequence in which n is already present. The reward RI rewards output length matching n, but it never rewards the model for generating the budget statement itself, nor does any training example require the model to produce that sentence autoregressively. At inference, when no n is supplied, either the model produces its own declaration through self-conditioning—a distribution never optimized—or it does not, in which case the stage-2 policy is simply a more concise policy learned from the stage-1 length reward. The robustness experiment in Table 4 only shows that LAPO-I ignores conflicting external length instructions; this is equally consistent with the model having learned a fixed output template and with the inserted n being a training-only hint. To support the paper's central claim, the authors should report, at inference, the fraction of rollouts that contain a self-declared budget, the distribution of declared budgets, and whether the declared budget predicts the eventual output length. Without such evidence, the "internalization" claim is not established.
  2. [Table 1 and all result tables] No error bars, confidence intervals, or multiple-seed results are reported anywhere in the paper. The claimed accuracy improvements are 1.2–2.3 points, which are small relative to typical run-to-run variance in RL training. For the central comparison LAPO-I versus its base model, and for the Acc-Only ablation that is used to attribute the gains, the paper needs at least 3–5 independent seeds or paired bootstrap intervals. Without them, the reader cannot assess whether the accuracy differences are statistically meaningful, even though the token reductions are large.
  3. [§3.2, Eq. (3), Eq. (8)] The target length M(q) is defined as the median of the model's own correct rollouts, so the "discovery" of natural reasoning patterns is self-referential: the model is trained to match a length statistic that it itself generated. The paper asserts that this statistic reflects an intrinsic complexity-to-length relationship, but it provides no stability analysis of M(q) across training episodes or across different initializations. Furthermore, the dual-strategy update in Eq. (8) takes min(M(q), median) for previously solved problems, making M(q) monotonically non-increasing over time; this can bias the target downward and induce over-shortening. The authors should show the evolution of M(q) over training and confirm that it stabilizes, or at least discuss this bias and its effect on the final policy.
minor comments (6)
  1. [§4 and §5] The paper contains two consecutive section headings both reading "EXPERIMENT SETUP"; the first appears to be an empty section and should be removed or merged.
  2. [§A.1 vs §5 and Table 5] The hyperparameter values for α and β are inconsistent: Section A.1 states "α and β ... were 0.7 and 0.8," while Section 5 and Table 5 report α = 0.7 and β = 0.7. Please correct this discrepancy.
  3. [Algorithm 1, line 14] Line 14 says "Update πθ using GRPO with rewards R1," but the reward used in the Discovery stage is defined as RD in Eq. (4). The algorithm should refer to RD to avoid confusion.
  4. [§6, intro paragraph] There is a typo: "including the the form of length guidance" should read "including the form of length guidance."
  5. [§A.1 prompts] The LAPO-I prompt in the appendix uses a placeholder "{length}" and the method section says n = M(q). Please clarify how the placeholder is set during training and at inference, and whether at inference the model is expected to generate the entire "<think> I will answer the question with ... tokens." string by itself.
  6. [Figure 4] The legend in Figure 4 uses the shorthand "DeepScaleR-1.5B-Pre" whereas the rest of the paper uses "DeepScaleR-1.5B-Preview." Please make the legend consistent.

Circularity Check

1 steps flagged · score 4.0 of 10

LAPO's target length M(q) is defined as the median of the model's own correct rollouts and then used as the reward target, making the 'discovery' of natural reasoning depth a self-referential loop; the final benchmark gains remain externally measured.

  1. self definitional [Section 3.2, Eqs. (1)-(3); Section 3.3, Eqs. (6)-(7); Algorithm 1, lines 10-11, 21-23]
    "Lq = {|ri| : I(yi = ygold) = 1, i∈ [1, N]} (1) ... M : q 7→ Lmedian(q) =Median(Lq) (3) ... R2(ri, n) = exp(− (|ri|−n)^2 / (2σ^2)) if Ci = 1; prompt′ q = promptq + “<think> I will answer the question with n tokens.” where n = M(q)."

    The target length that is supposed to represent 'appropriate reasoning depth' is not independently derived from problem structure: Eq. (1) defines Lq as the lengths of the current policy's own correct rollouts, and Eq. (3) sets M(q) to the median of those self-generated lengths. Stage 1 then rewards correct responses for falling inside [P30,P70] of that same self-sampled distribution (Eq. 5), and Stage 2 rewards adherence to n = M(q) (Eq. 7). Thus the model is trained to match a statistic of its own outputs; the 'discovery' of 'natural reasoning patterns' is a self-consistency loop, not an independent measurement of intrinsic problem complexity. The external benchmark token/accuracy numbers remain meaningful, so the circularity is partial rather than total.

full rationale

The concrete equation-level circularity is the self-referential target M(q): Eq. (1) collects lengths only from the policy's own correct rollouts, Eq. (3) defines the target as the median of those lengths, and Eqs. (5) and (7) use that same target in the reward. This means the 'natural reasoning pattern' the paper claims to discover is, by construction, a statistic of the very distribution being optimized. However, the headline efficiency and accuracy results are measured on external benchmarks against external baselines (Table 1), so they are not forced by this loop. No load-bearing self-citation is present; the reference list consists of external prior work. The Sec. 6.4 internalization test is an evidential gap rather than a circular derivation: Algorithm 1 line 21 injects the budget statement into the prompt after <think>, so the policy is never trained to emit that statement autoregressively, and robustness to conflicting external instructions is consistent with a training-conditional behavior; this affects the validity of the 'self-declared budget' claim but does not reduce any equation to its input. The M(q)=4096 fallback for unsolved problems is a design bias, not a circular step. Overall, the central method is self-referential in its target, but its measured outcomes are externally grounded, giving a partial circularity score of 4.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The method's central target (the per-problem median length) is a free, data-derived quantity, which gives the approach a self-referential character. There are no truly invented entities, only a prompt template. The main axioms are standard RL and domain assumptions, plus the untested assumption that the model self-declares a budget at inference.

free parameters (5)
  • alpha (length reward weight in Discovery) = 0.7
    Weights the length reward R1 against the correctness reward in Eq. 4. Chosen empirically, not fitted to a validation set; affects the strength of the efficiency pressure.
  • beta (adherence reward weight in Internalization) = 0.7 (text and Table 5), 0.8 (Appendix A.1)
    Weights the Gaussian adherence reward R2 in Eq. 6. Inconsistent reporting between sections is a reproducibility concern.
  • sigma (Gaussian width in R2) = not specified
    Controls how tightly the model must match the declared budget n in Eq. 7. No value is given, making exact reproduction impossible.
  • Percentiles P30/P70 for length range = 30th and 70th percentiles
    Defines the 'efficient range' in Eq. 2. Chosen by design, not tuned; different percentiles would change the reward shape.
  • Per-problem target length M(q) = median of correct rollouts (or 4096 for unsolved)
    The central per-problem target learned from the model's own rollouts. It is a fitted statistic that directly drives the Internalization reward; if the rollouts are unrepresentative, the target is poor.
assumptions (4)
  • standard math GRPO with the specified hyperparameters improves the policy as expected.
    The paper relies on GRPO as a working RL algorithm without proving any guarantees. This is a standard background assumption in RL fine-tuning.
  • domain assumption The correctness reward I(y_i = y_gold) is a reliable and complete signal for reasoning quality.
    The method assumes that achieving the final answer is the only objective and that the model's internal reasoning quality is fully captured by answer correctness. This is standard for math benchmarks.
  • domain assumption The model is capable of following the self-declared budget statement and of generating its own budget at inference.
    The Internalization stage relies on the model treating the inserted 'I will answer with n tokens' as a plan rather than an external command. The paper does not directly verify that the model generates this statement at inference, only that it ignores conflicting external instructions.
  • ad hoc to paper The median of correct rollouts is a stable statistic for a given problem.
    With only 8 rollouts per problem, the median can be noisy, especially for hard problems with few correct answers. The paper updates M(q) over time but does not measure the variance or stability of the estimate.
invented entities (1)
  • Self-declarative budget statement in the <think> block
    purpose: A prompt modification that reframes the token budget as the model's own plan, intended to produce internalized length control.
    This is a textual mechanism, not a physical entity. It has no independent handle outside the paper; its effectiveness is only measured through the downstream accuracy/efficiency metrics, which are not separated from the rest of the training pipeline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LAPO: Internalizing Reasoning Efficiency via Length-Adaptive Policy Optimization." pith.science (2026). https://pith.science/paper/A3ACZQBP

@misc{pith2026250715758,
  author       = {Pith},
  title        = {Pith review of: LAPO: Internalizing Reasoning Efficiency via Length-Adaptive Policy Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A3ACZQBP}},
  note         = {Machine review of arXiv:2507.15758}
}
read the original abstract

Large reasoning models have achieved remarkable performance through extended chain-of-thought sequences, yet this computational freedom leads to excessive token generation even for simple problems. We present Length-Adaptive Policy Optimization (LAPO), a novel framework that transforms reasoning length control from an external constraint into an intrinsic model capability. Unlike existing approaches that impose rigid limits or rely on post-hoc interventions, LAPO enables models to internalize an understanding of appropriate reasoning depth through a two-stage reinforcement learning process. In the first stage, models learn natural reasoning patterns by discovering the statistical distribution of successful solution lengths. The second stage leverages these patterns as meta-cognitive guidance, embedding them directly within the model's reasoning context to ensure inference-time flexibility. Experiments on mathematical reasoning benchmarks demonstrate that LAPO reduces token usage by up to 40.9% while improving accuracy by 2.3%. Our analysis reveals that models trained with LAPO develop emergent abilities to allocate computational resources based on problem complexity, achieving efficient reasoning without sacrificing quality.

Figures

Figures reproduced from arXiv: 2507.15758 by the authors.

Figure 1
Figure 1. Overview of Length-Adaptive Policy Optimization (LAPO) and its superior performance. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The LAPO framework consists of two stages: (1) Discovery stage learns natural reasoning [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Reasoning length allocation across mathematical problem difficulty levels. LAPO learns to scale computation with complexity. 0 2 4 6 8 10 12 14 Keyword Count per 1K Tokens But Wait, Alternatively, Perhaps First, Okay, Given Therefore, So, Self-correction & Verification Exploration & Alternatives Context Setting Conclusion Drawing DeepScaleR-1.5B-Pre LAPO-D LAPO-I [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Training dynamics evaluated on the MATH-500 validation set. Checkpoints were saved [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. Overthink-Triggered Slowdown Attacks on LVLM-Based Robotic Systems

    cs.CR 2026-07 unverdicted novelty 6.0 of 10

    Adversaries can use crafted scene text to trigger overthinking in LVLM-based robots, producing transferable slowdowns up to 6.96x latency amplification.

  2. BudgetThinker: Empowering Budget-aware LLM Reasoning with Control Tokens

    cs.LG 2025-08 conditional novelty 6.0 of 10

    A control-token insertion and two-stage training method that lets LLMs adhere to user-specified reasoning token budgets while preserving math accuracy.

  3. StaRPO: Stability-Augmented Reinforcement Policy Optimization

    cs.AI 2026-04 unverdicted novelty 5.0 of 10

    StaRPO improves LLM reasoning by adding autocorrelation function and path efficiency stability metrics to RL policy optimization, yielding higher accuracy and fewer logic errors on reasoning benchmarks.

Reference graph

Works this paper leans on

28 extracted references · 1 canonical work pages · cited by 3 Pith papers

  1. [1]

    L1: Controlling how long a reasoning model thinks with reinforcement learning

    Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning. arXiv preprint arXiv:2503.04697,

    arXiv

  2. [4]

    Thinkless: Llm learns when to think.arXiv preprint arXiv:2505.13379, 2025a

    Gongfan Fang, Xinyin Ma, and Xinchao Wang. Thinkless: Llm learns when to think.arXiv preprint arXiv:2505.13379, 2025a. Gongfan Fang, Xinyin Ma, and Xinchao Wang. Thinkless: LLM learns when to think. CoRR, abs/2505.13379, 2025b. doi: 10.48550/ARXIV .2505.13379. URL https://doi.org/10. 48550/arXiv.2505.13379. Kanishk Gandhi, Ayush Chakravarthy, Anikait Sing...

    arXiv doi:10.48550/arxiv.2505.13379

  3. [5]

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

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

    arXiv

  4. [6]

    Token- budget-aware llm reasoning

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

    arXiv

  5. [7]

    Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems

    Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Leng Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Yuxiang Zhang, et al. Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems. arXiv preprint arXiv:2402.14008,

    arXiv

  6. [8]

    Measuring massive multitask language understanding

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300,

    arXiv 2009

  7. [10]

    Hapo: Training language models to reason concisely via history-aware policy optimization

    Chengyu Huang, Zhengxin Zhang, and Claire Cardie. Hapo: Training language models to reason concisely via history-aware policy optimization. arXiv preprint arXiv:2505.11225, 2025a. Chengyu Huang, Zhengxin Zhang, and Claire Cardie. HAPO: training language models to reason concisely via history-aware policy optimization. CoRR, abs/2505.11225, 2025b. doi: 10....

    work page doi:10.48550/arxiv.2505.11225

  8. [11]

    Overthink: Slowdown attacks on reasoning llms

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

    arXiv

Show all 28 references
  1. [12]

    Adacot: Pareto-optimal adaptive chain-of-thought triggering via reinforcement learning

    Chenwei Lou, Zewei Sun, Xinnian Liang, Meng Qu, Wei Shen, Wenqi Wang, Yuntao Li, Qingping Yang, and Shuangzhi Wu. Adacot: Pareto-optimal adaptive chain-of-thought triggering via reinforcement learning. arXiv preprint arXiv:2505.11896,

    arXiv

  2. [13]

    O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning

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

    arXiv

  3. [14]

    Imitate, explore, and self-improve: A reproduction report on slow-thinking reasoning systems

    Yingqian Min, Zhipeng Chen, Jinhao Jiang, Jie Chen, Jia Deng, Yiwen Hu, Yiru Tang, Jiapeng Wang, Xiaoxue Cheng, Huatong Song, et al. Imitate, explore, and self-improve: A reproduction report on slow-thinking reasoning systems. arXiv preprint arXiv:2412.09413,

    arXiv

  4. [15]

    s1: Simple test-time scaling

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Cand `es, and Tatsunori Hashimoto. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393,

    arXiv

  5. [16]

    Routellm: Learning to route llms with preference data

    Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E Gonzalez, M Waleed Kadous, and Ion Stoica. Routellm: Learning to route llms with preference data. arXiv preprint arXiv:2406.18665,

    arXiv

  6. [17]

    Concise: Confidence-guided compression in step-by-step efficient reasoning

    Ziqing Qiao, Yongheng Deng, Jiali Zeng, Dong Wang, Lai Wei, Fandong Meng, Jie Zhou, Ju Ren, and Yaoxue Zhang. Concise: Confidence-guided compression in step-by-step efficient reasoning. arXiv preprint arXiv:2505.04881,

  7. [18]

    Scaling llm test-time compute optimally can be more effective than scaling model parameters

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314,

    arXiv

  8. [19]

    Stop overthinking: A survey on efficient reasoning for large language models

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

    arXiv

  9. [20]

    Learning when to think: Shaping adaptive reasoning in r1-style models via multi-stage rl

    Songjun Tu, Jiahao Lin, Qichao Zhang, Xiangyu Tian, Linjing Li, Xiangyuan Lan, and Dongbin Zhao. Learning when to think: Shaping adaptive reasoning in r1-style models via multi-stage rl. arXiv preprint arXiv:2505.10832,

  10. [21]

    Self-consistency improves chain of thought reasoning in language models

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

    arXiv

  11. [23]

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

    Sean Welleck, Amanda Bertsch, Matthew Finlayson, Hailey Schoelkopf, Alex Xie, Graham Neubig, Ilia Kulikov, and Zaid Harchaoui. From decoding to meta-generation: Inference-time algorithms for large language models. arXiv preprint arXiv:2406.16838,

    arXiv

  12. [24]

    Inference scaling laws: An empirical analysis of compute-optimal inference for problem-solving with language models

    Yangzhen Wu, Zhiqing Sun, Shanda Li, Sean Welleck, and Yiming Yang. Inference scaling laws: An empirical analysis of compute-optimal inference for problem-solving with language models. arXiv preprint arXiv:2408.00724,

    arXiv

  13. [25]

    Tokenskip: Controllable chain-of-thought compression in llms

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

  14. [26]

    Chain of draft: Thinking faster by writing less

    Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. Chain of draft: Thinking faster by writing less. arXiv preprint arXiv:2502.18600, 2025a. 13 Preprint Yuhui Xu, Hanze Dong, Lei Wang, Doyen Sahoo, Junnan Li, and Caiming Xiong. Scalable chain of thoughts via elastic reasoni...

    arXiv

  15. [28]

    Adaptthink: Reasoning models can learn when to think

    Jiajie Zhang, Nianyi Lin, Lei Hou, Ling Feng, and Juanzi Li. Adaptthink: Reasoning models can learn when to think. arXiv preprint arXiv:2505.13417, 2025a. Jiajie Zhang, Nianyi Lin, Lei Hou, Ling Feng, and Juanzi Li. Adaptthink: Reasoning models can learn when to think. CoRR, a...

    doi arXiv

  16. [2020]

    Thinkprune: Pruning long chain-of-thought of llms via reinforcement learning

    Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu, Kaizhi Qian, Jacob Andreas, and Shiyu Chang. Thinkprune: Pruning long chain-of-thought of llms via reinforcement learning. arXiv preprint arXiv:2504.01296,

    arXiv

  17. [2022]

    Thoughts are all over the place: On the underthinking of o1-like llms

    Yue Wang, Qiuzhi Liu, Jiahao Xu, Tian Liang, Xingyu Chen, Zhiwei He, Linfeng Song, Dian Yu, Juntao Li, Zhuosheng Zhang, et al. Thoughts are all over the place: On the underthinking of o1-like llms. arXiv preprint arXiv:2501.18585,

    arXiv

  18. [2023]

    Demystifying long chain- of-thought reasoning in llms

    Edward Yeo, Yuxuan Tong, Morry Niu, Graham Neubig, and Xiang Yue. Demystifying long chain- of-thought reasoning in llms. arXiv preprint arXiv:2502.03373,

    arXiv

  19. [2024]

    Learning to route llms with confidence tokens

    Yu-Neng Chuang, Helen Zhou, Prathusha Kameswara Sarma, Parikshit Gopalan, John Boccio, Sara Bolouki, and Xia Hu. Learning to route llms with confidence tokens. arXiv preprint arXiv:2410.13284,

    arXiv

  20. [2025]

    Do not think that much for 2+ 3=? on the overthinking of o1-like llms

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

    arXiv

Pith tools

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