Pith. sign in

REVIEW 5 major objections 6 minor 6 cited by

Optimizing Length Compression in Large Reasoning Models

T0 review · 5 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A new post-training method identifies and cuts "invalid thinking" tokens that appear after the correct answer is first derived, halving output length while losing roughly 2% accuracy.

desk verdict LC-R1 delivers a real, incremental win — extractor-guided </think> compression cuts ~50% of tokens at ~2% accuracy cost — but the VT headline is circular and the evidence lacks variance bars. read the letter →

arxiv 2506.14755 v2 pith:B7XZR6ZN submitted 2025-06-17 cs.AI cs.CL

Zhengxiang Cheng , Dongping Chen , Mingyang Fu , Tianyi Zhou This is my paper · ORCID
classification cs.AIcs.CL
keywords largereasoningmodelsinvalidthinkingvalidratecompressionGRPOlengthrewardcompressefficient
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 is trying to establish that a specific inefficiency in long-thinking models, which it calls "invalid thinking," can be pruned in post-training without meaningful accuracy loss. Invalid thinking is defined as the tokens produced after the model first states the correct answer, and the paper argues that this tail is largely redundant double-checking. It introduces LC-R1, a GRPO-based method with two rewards, one for overall brevity and one for stopping at the first correct derivation. On 7B and 1.5B reasoning models, the reported result is roughly half the output length with only about a 2% drop in accuracy, plus a large increase in the "valid thinking" rate. If true, this is a practical recipe for cutting the inference cost of reasoning models while preserving their problem-solving potential.

What carries the argument

The load-bearing object is the compression function $f$, realized by a finetuned parser called LC-Extractor, which locates the first occurrence of the ground-truth answer inside a thinking trace and truncates the trace at that point. Around this function the method builds a dual reward: a Length Reward, $1 - |o'_i|/\max_{j\in C}|o'_j|$, for overall conciseness, and a Compress Reward applied only to the final `</think>` token, equal to the fraction of tokens removed when the answer is correct and $-1$ if the model stops before finding the correct answer. GRPO is then run on the compressed trajectories $o'_i = f(o_i)$, so policy updates favor terminating thinking at the first correct derivation. LC-Extractor performs double duty, generating training targets and measuring the valid-thinking metric, so its reliability is central to both the training signal and the headline efficiency numbers.

What would settle it

Take a fresh sample of reasoning traces from the base and compressed models, have independent annotators mark the first point where the correct answer is actually established, and compare those marks with LC-Extractor's outputs; if agreement is low, or if retraining LC-R1 with a hand-corrected or differently seeded extractor changes the reported length/accuracy trade-off, the central claim is not stable.

Watch

Extended reading notes

Core claim

The paper's central claim is that the verbosity of large reasoning models is not diffuse: a substantial fraction of generated tokens occurs after the correct answer has already been derived, and this tail can be removed without harming the model's ability to solve problems. LC-R1 reports average length reductions of 46.3% on DeepSeek-R1-Distill-Qwen-7B and 51.9% on the 1.5B variant, with average accuracy drops of 1.84 and 2.14 percentage points respectively across seven benchmarks. The method also raises the valid thinking rate to above 97%, meaning nearly all thinking tokens precede the first correct answer. The paper further argues that pass@k performance on AIME25 is nearly unchanged after compression, concluding that the pruned "invalid thinking" segments are truly redundant rather than exploratory.

Load-bearing premise

The load-bearing premise is that the LC-Extractor parser reliably finds the first moment the model states the correct answer, because every training reward and the reported valid-thinking metric depend on that cut point; it was validated on only 100 human-annotated samples.

Editorial extensions

If this is right

  • On seven benchmarks spanning math, general knowledge, and code, LC-R1 reduces average token count by 46.3% on the 7B model and 51.9% on the 1.5B model, with average accuracy drops of 1.84% and 2.14% respectively.
  • Valid thinking rate rises to 97.1% (7B) and 98.6% (1.5B), meaning almost all generated thinking tokens now occur before the first correct answer.
  • Pass@k on AIME25 up to $k=128$ is nearly unchanged after compression, so repeated sampling still finds correct solutions at the same rate.
  • Compression is uniform across problem difficulty, and out-of-distribution benchmarks such as GPQA-Diamond and LiveCodeBench hold up better under LC-R1 than under SFT-based baselines.
  • Ablations show that both rewards are necessary: removing the Compress Reward lowers the valid thinking rate sharply, while removing the Length Reward reduces overall compression.

Reading between the lines

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

  • An extrapolation: the same redundancy pattern appears in the paper's measurements of much larger models, where 35-45% of tokens are invalid thinking, so similar training could plausibly yield comparable savings on models beyond the 7B scale, though LC-R1 itself is only trained and tested up to 7B.
  • Because LC-Extractor defines both the training signal and the headline metric, a natural testable extension is to use one extractor for training data and an independent hand-annotated extractor for evaluation, which would separate real efficiency gains from artifacts of the measurement.
  • The "stop at the first correct derivation" reward could generalize beyond the `</think>` token to other stopping decisions, such as terminating a tool-use loop or ending a search once a verified answer is found.
  • Compression-trained models are natural teachers for shorter-chain distillation, potentially letting smaller models inherit brevity without needing reinforcement learning themselves.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes LC-R1, a GRPO-based post-training method for compressing the reasoning chains of large reasoning models. The authors introduce a "Valid Thinking" (VT) rate (Eq. 1) that measures the fraction of a trace before the correct answer is first derived, and a compression function f (Eq. 2) that truncates the thinking process at that boundary using a learned parser, LC-Extractor. Training combines a Length Reward (Eq. 3) with a Compress Reward (Eq. 11) applied to the </think> token, and the GRPO objective is computed on the compressed trajectories (Eqs. 7-10). Experiments on DeepSeek-R1-Distill-Qwen-7B/1.5B across seven benchmarks report roughly 46-52% average length reduction with about 2% average accuracy drop, and ablations show that both rewards contribute. The paper also claims that compression does not impair exploration ability based on Pass@k curves and that the removed "invalid thinking" segments are truly redundant.

Significance. If the results are confirmed, LC-R1 is a practical and inexpensive method for roughly halving the inference cost of small reasoning models with only a small accuracy penalty. The paper deserves credit for releasing code, for evaluating on seven diverse benchmarks, and for including ablations and comparison with several baselines. The central quantitative result in Table 2 is measured on raw token counts and external answer checking, so the headline length-reduction number is not an artifact of the extractor. However, the paper's conceptual claims about "invalid thinking," the VT rates, and the training signal all rest on the LC-Extractor boundary, which is validated on only 100 samples; the training objective is also not derived as a standard policy-gradient surrogate. The lack of error bars and possible train/test overlap with MATH500 make the "marginal (~2%) drop" claim difficult to assess. These issues are fixable, so the contribution is promising but needs substantial revision.

major comments (5)
  1. [Sections 2.1, 3.1, 3.2, and Appendix B; Eq. (1), Eq. (11), Fig. 5] The LC-Extractor's boundary definition is ambiguous and its validation is too thin to support the load-bearing "valid thinking" construct. The text in Sections 2.1 and 3.1 says the valid prefix runs "until the correct answer is first derived," but the prompt in Figure 5 asks for "the first sentence that reveals the Ground Truth Answer." A trace can mention an answer before deriving it, so these are not equivalent. The only validation in Appendix B is 98% agreement on 100 samples, and the metric is agreement on extraction position, not whether the extracted prefix is sufficient for derivation. Because the Compress Reward in Eq. (11) and the VT metric in Eq. (1) both depend on this boundary, a systematic early-cut error would inflate VT rates and could penalize valid reasoning during training. Please provide a larger human study with a precise definition of "first derived," report inter-annotator agreement, and analyze cases where the first mention and first derivation differ.
  2. [Section 3.2, Eqs. (7)-(9)] The GRPO objective in Eq. (7) is computed on compressed trajectories o'_i = f(o_i), but the samples are drawn from the old policy as o_i. The likelihood ratio R_t(θ) = π_θ(o'_{i,t} | ...)/π_θ_old(o'_{i,t} | ...) is not the likelihood ratio of the actually sampled action sequence, because o'_i is a counterfactual sequence that the sampling policy did not generate. The expectation in Eq. (7) is therefore not the standard policy-gradient surrogate for the reward under π_θ, and no off-policy correction or derivation is provided. The paper should either derive the correct objective, or explicitly frame LC-R1 as a heuristic and provide analysis of the induced bias. This matters because the training objective is the core of the proposed method.
  3. [Section 4.3, Table 2] All experimental results are single-run, with no error bars or multiple seeds. The central claim of a "marginal (~2%) drop" in accuracy is an average of per-benchmark relative changes; on AIME25 (30 problems), the reported difference is under one problem, which is within plausible sampling noise. The paper should report variance across at least three seeds, or bootstrap confidence intervals, and ideally a significance test for the accuracy differences. The length-reduction numbers are large and likely robust, but the accuracy-drop claim needs statistical support.
  4. [Sections 4.1 and Appendix C.2] The training data is described as "past AIME competition problems with the MATH dataset" (Section 4.1), and the evaluation uses AIME25 and MATH500, which are subsets of the AIME and MATH problem distributions. The paper does not state that these evaluation benchmarks were excluded from the training set. If MATH500 (or AIME25) overlaps with training, the corresponding numbers in Table 2 are in-distribution and the generalization claims are weakened. Please clarify the exact train/test split and the contamination controls used.
  5. [Section 5, Figure 4] The Pass@k comparison between the original model and the retrained LC-R1 model does not establish that the removed suffix of the original trace is "truly redundant." The LC-R1 model is a different policy; its similar Pass@k curve reflects the new policy's exploration ability, not the redundancy of the original model's post-answer tokens. To support the redundancy claim, the authors should evaluate the original model with its own traces truncated at the extracted boundary, or otherwise directly measure the contribution of the removed suffix. In addition, the Pass@k curves in Figure 4 are based on 30 AIME25 problems and are presented without error bars, so "almost perfectly overlaps" is not quantitatively established.
minor comments (6)
  1. [Section 2.1 heading] The heading "Compression and Efficienct Reasoning Models" contains a typo: "Efficienct" should be "Efficient."
  2. [Appendix C.2, AIME25 description] The description of AIME25 as "International Conference on Artificial Intelligence in Medicine" is incorrect; AIME here stands for the American Invitational Mathematics Examination. This is a factual error that should be corrected.
  3. [Section 4.1 and Table 2] The text says "We use averaged Pass@1 as our primary metric," but Table 2 reports pass@10 for LiveCodeBench. Please clarify the exact evaluation metric used for each benchmark.
  4. [Acknowledgment] The acknowledgment "Many thanks to Yao Wan and for his invaluable support" appears to have a missing name after "and." Please complete the sentence.
  5. [Eq. (11)] The notation "ans(q)∈t(o'_i)" uses set-membership for a token sequence; please define it explicitly as "the ground-truth answer appears in the thinking token sequence t(o'_i)."
  6. [Section 4.2 and Table 2] The DPO baseline is described as being applied to the SFT-tuned model, but Table 2 lists "DPO" as a standalone method; the naming may confuse readers about the actual training pipeline.

Circularity Check

1 steps flagged · score 4.0 of 10

VT metric and Compress Reward are the same quantity by construction; headline length/accuracy results remain independent.

  1. self definitional [Eq. (1) and Sec. 2.1; Eq. (2) and Sec. 3.1; Eq. (11) and Sec. 3.2; reported VT in Sec. 4.3 and Tables 2-3]
    "VT= |Tokens in Valid Thinking| / |Total tokens in Thinking Process| (1) where “Valid Thinking” comprises the tokens from the start of the thinking process until the correct answer is first derived. ... f({R,A}) = {R ′,A} (2) ... ri,compress = 1− |t(o′ i)| / |t(oi)| (11)"

    By Eq. (2), o′_i = f(o_i) truncates the reasoning at the first occurrence of the correct answer, which is exactly the segment that Eq. (1) labels “Valid Thinking.” Therefore, for a correct sample, |t(o′_i)|/|t(o_i)| is the VT rate, and the Compress Reward in Eq. (11) is literally the invalid-token fraction, i.e., the complement of VT. LC-R1 is trained to maximize this quantity, so reporting VT > 97% in Tables 2-3 and calling it evidence of “superior efficiency” restates the training objective rather than independently confirming that post-first-answer tokens are invalid.

full rationale

The paper's central quantitative claims—roughly 50% raw sequence-length reduction with only about 2% accuracy drop—are evaluated with external token counts and answer checks, so they do not reduce to the paper's own definitions. The circularity is confined to the “invalid thinking” narrative and the Valid Thinking (VT) metric. In Eq. (1), valid thinking is defined as the prefix ending at the first occurrence of the correct answer; in Eq. (2), the compression function f truncates the trace at that same boundary; and in Eq. (11), the Compress Reward is 1 minus the ratio of compressed reasoning length to original reasoning length, i.e., the fraction of tokens removed by f. For correct samples this is exactly the complement of the VT rate. Thus the model is rewarded for maximizing the same quantity that the paper later reports as a favorable VT outcome. This makes the VT-based evidence for the “invalid thinking” phenomenon partially self-definitional: the extractor both constructs the training targets and measures the reported efficiency metric. However, the paper also validates the usefulness of compression through external accuracy and length measurements, and the ablations compare methods under the same evaluation protocol. No load-bearing self-citation or imported uniqueness theorem appears, and the pass@k exploration analysis, while indirect, is not circular by construction. The appropriate score is therefore a moderate 4: one significant construction-level circularity in the VT efficiency claim, with the main length-versus-accuracy contribution still independently grounded.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method's empirical claims rest on a handful of hand-chosen hyperparameters and on assumptions about the extractor, the definition of invalid thinking, and the answer verifier. The central quantitative result is not parameter-fitted in the sense of being derived from fitted constants, but the training objective and evaluation both depend on the LC-Extractor's notion of a valid prefix.

free parameters (5)
  • alpha (Length Reward weight) = 1.0
    Scales the Length Reward in Eq. 4; chosen by hand, directly controls how much compression is rewarded relative to accuracy and format.
  • gamma (Compress Reward weight) = 1.0
    Scales the bonus or penalty on the </think> token in Eq. 10; chosen by hand, controls how strongly the model is pushed to stop after the first correct answer.
  • beta (KL penalty coefficient) = 0.04
    Standard KL regularization weight in the GRPO objective Eq. 7; chosen by hand.
  • num_generations = 6 (7B), 8 (1.5B)
    GRPO group size per prompt; set by hand and affects the length-reward normalization baseline.
  • epsilon (PPO clip range) = 0.2
    Clip ratio in Eq. 7; standard value from GRPO, chosen by hand.
assumptions (4)
  • standard math GRPO/DAPO update rule is a valid RL optimizer for this setting.
    The objective in Eq. 7 is assumed to effectively optimize the reward; no convergence proof is given.
  • domain assumption LC-Extractor accurately finds the first occurrence of the ground-truth answer in a reasoning trace.
    Validated on 100 samples in Appendix B; all training rewards and VT measurements depend on this.
  • domain assumption Tokens after the first correct answer are redundant and can be removed without changing final-answer correctness or future exploration.
    This defines 'invalid thinking'; it is the premise behind the Compress Reward and the pass@k generalization claim.
  • domain assumption Math-Verify correctly determines whether the final answer matches the ground truth.
    Used for accuracy rewards and evaluation; errors in answer verification would misassign rewards and metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimizing Length Compression in Large Reasoning Models." pith.science (2026). https://pith.science/paper/B7XZR6ZN

@misc{pith2026250614755,
  author       = {Pith},
  title        = {Pith review of: Optimizing Length Compression in Large Reasoning Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B7XZR6ZN}},
  note         = {Machine review of arXiv:2506.14755}
}
read the original abstract

Large Reasoning Models (LRMs) have achieved remarkable success, yet they often suffer from producing unnecessary and verbose reasoning chains. We identify a core aspect of this issue as "invalid thinking" -- models tend to repeatedly double-check their work after having derived the correct answer. To address this specific inefficiency, we move beyond the general principles of Efficacy and Efficiency to propose two new, fine-grained principles: Brevity, which advocates for eliminating redundancy, and Sufficiency, which ensures critical reasoning steps are preserved. Guided by these principles, we introduce LC-R1, a post-training method based on Group Relative Policy Optimization (GRPO). LC-R1 employs a novel combination of a Length Reward for overall conciseness and a Compress Reward that is specifically designed to remove the invalid portion of the thinking process. Extensive experiments on multiple reasoning benchmarks demonstrate that LC-R1 achieves a significant reduction in sequence length (~50%) with only a marginal (~2%) drop in accuracy, achieving a favorable trade-off point on the Pareto frontier that prioritizes high compression. Our analysis further validates the robustness of LC-R1 and provides valuable insights for developing more powerful yet computationally efficient LRMs. Our code is released at https://github.com/zxiangx/LC-R1.

Figures

Figures reproduced from arXiv: 2506.14755 by the authors.

Figure 1
Figure 1. Comparison between inefficient reasoning model [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Pareto analysis of the Efficacy-Efficiency trade-off of different methods on two reasoning models. The x-axis [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. An overview of the LC-R1 training three-stage pipeline. (1) Valid Segment Extraction: First, an extractor model processes the original reasoning traces to identify the valid thinking portion and generate compressed sequences. (2) Reward Calculation: Next, these compressed sequences are used to compute our dual rewards—Length Reward and Compress Reward, with the latter applied exclusively as a bonus or penalty on the… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The impact of LC-R1 compression method on the AIME25 benchmark. Left: The Pass@k scores show that LC-R1 models maintain competitive performance compared to the originals, preserving the model’s potential. Right: Per-problem analysis on Deepseek-R1-Distill-Qwen-7B revea…
Figure 5
Figure 5. Figure 5: Our prompt for extraction of answer prefix. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: The annotation tool to evaluate the LC-Extratcor. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Case study of the comparison of LC-R1 and O1-Pruner. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Case study of the comparison of LC-R1 and the original model. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 6 Pith papers

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

  1. EvoThink: Evolving Thinking in Large Reasoning Models via Self-Pruning and Aha-Moment Preference Optimization

    cs.AI 2026-07 conditional novelty 6.0 of 10

    EvoThink reduces LRM overthinking by pruning redundant atomic reasoning steps (SPT) and training on diversity-selected wrong-to-right mutation data (AMPO), cutting tokens and improving math/code accuracy.

  2. 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.

  3. ConPress: Learning Efficient Reasoning from Multi-Question Contextual Pressure

    cs.CL 2026-02 conditional novelty 6.0 of 10

    Multi-question prompts elicit shorter chain-of-thought traces, and fine-tuning on those traces transfers the compression to single-question reasoning.

  4. Reconsidering Overthinking: Penalizing Internal and External Redundancy in CoT Reasoning

    cs.AI 2025-08 conditional novelty 6.0 of 10

    A dual-penalty RL method that compresses chain-of-thought traces by separately penalizing internal semantic stagnation and external post-answer continuation reduces reasoning length by about 40% while preserving accur...

  5. Failure Cases Are Better Learned But Boundary Says Sorry: Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training

    cs.CV 2025-08 reject novelty 5.0 of 10

    A paper whose abstract describes new adversarial training experiments, but whose full text is a different paper on CoT compression, leaving the claims unsupported.

  6. 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

42 extracted references · 25 linked inside Pith · cited by 6 Pith papers

  1. [1]

    Openai o1 system card

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

    arXiv 2024

  2. [2]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao ..., and Zhen Zhang. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025. URL https://arxiv.org/...

    arXiv 2025

  3. [3]

    Chain-of-thought prompting elicits reasoning in large language models, 2023

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. URL https://arxiv.org/abs/2201.11903

    arXiv 2023

  4. [4]

    Challenging the boundaries of reasoning: An olympiad-level math benchmark for large language models, 2025

    Haoxiang Sun, Yingqian Min, Zhipeng Chen, Wayne Xin Zhao, Zheng Liu, Zhongyuan Wang, Lei Fang, and Ji-Rong Wen. Challenging the boundaries of reasoning: An olympiad-level math benchmark for large language models, 2025. URL https://arxiv.org/abs/2503.21380

    arXiv 2025

  5. [5]

    Alex Gu, Baptiste Rozière, Hugh Leather, Armando Solar-Lezama, Gabriel Synnaeve, and Sida I. Wang. Cruxeval: A benchmark for code reasoning, understanding and execution. arXiv preprint arXiv:2401.03065, 2024

    arXiv 2024

  6. [6]

    Towards reasoning era: A survey of long chain-of-thought for reasoning large language models, 2025

    Qiguang Chen, Libo Qin, Jinhao Liu, Dengyun Peng, Jiannan Guan, Peng Wang, Mengkang Hu, Yuhang Zhou, Te Gao, and Wanxiang Che. Towards reasoning era: A survey of long chain-of-thought for reasoning large language models, 2025. URL https://arxiv.org/abs/2503.09567

    arXiv 2025

  7. [7]

    L1: Controlling how long a reasoning model thinks with reinforcement learning, 2025

    Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning, 2025. URL https://arxiv.org/abs/2503.04697

    arXiv 2025

  8. [8]

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

    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, 2024

    arXiv 2024

Show all 42 references
  1. [9]

    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, Hanjie Chen, Xia Hu, et al. Stop overthinking: A survey on efficient reasoning for large language models. arXiv preprint arXiv:2503.16419, 2025

    2025 arXiv

  2. [10]

    The danger of overthinking: Examining the reasoning-action dilemma in agentic tasks

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

    2025 arXiv

  3. [11]

    Adar1: From long-cot to hybrid-cot via bi-level adaptive reasoning optimization

    Haotian Luo, Haiying He, Yibo Wang, Jinluan Yang, Rui Liu, Naiqiang Tan, Xiaochun Cao, Dacheng Tao, and Li Shen. Adar1: From long-cot to hybrid-cot via bi-level adaptive reasoning optimization. arXiv preprint arXiv:2504.21659, 2025 a

    2025 arXiv

  4. [12]

    Dast: Difficulty-adaptive slow-thinking for large reasoning models, 2025

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

    2025

  5. [13]

    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, 2025

    2025 arXiv

  6. [14]

    O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning, 2025 b

    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, 2025 b . URL https://arxiv.org/abs/2501.12570

    2025 arXiv

  7. [15]

    Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, Chuning Tang, Congcong Wang, Dehao Zhang, Enming Yuan, Enzhe Lu, Fengxiang Tang, Flood Sung, Guangda Wei, Guokun Lai, Haiqing Guo, Han Zhu, Hao Ding, ...

    2025 arXiv

  8. [16]

    Qwen3, April 2025 a

    Qwen Team. Qwen3, April 2025 a . URL https://qwenlm.github.io/blog/qwen3/

    2025

  9. [17]

    Qwq-32b: Embracing the power of reinforcement learning, March 2025 b

    Qwen Team. Qwq-32b: Embracing the power of reinforcement learning, March 2025 b . URL https://qwenlm.github.io/blog/qwq-32b/

    2025

  10. [18]

    Yian Zhang, and Chris Alexiuk

    Akhiad Bercovich, Itay Levy, Izik Golan, Mohammad Dabbah, Ran El-Yaniv, Omri Puny, Ido Galil, Zach Moshe, Tomer Ronen, Najeeb Nabwani, Ido Shahaf, Oren Tropp, Ehud Karpas, Ran Zilberstein, Jiaqi Zeng, Soumye Singhal, Alexander Bukharin, ... Yian Zhang, and Chris Alexiuk. Llama...

    2025

  11. [19]

    Understanding r1-zero-like training: A critical perspective, 2025

    Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. Understanding r1-zero-like training: A critical perspective, 2025. URL https://arxiv.org/abs/2503.20783

    2025 arXiv

  12. [20]

    Dapo: An open-source llm reinforcement learning system at scale

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, et al. Dapo: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476, 2025

    2025 arXiv

  13. [21]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36: 0 53728--53741, 2023

    2023

  14. [22]

    OpenAI. Chatgpt. https://openai.com/o1/, 2024

    2024

  15. [23]

    Gemini 2.5 pro

    Google. Gemini 2.5 pro. https://cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/2-5-pro, 2025 a

    2025

  16. [24]

    Phi-4-reasoning technical report, 2025

    Marah Abdin, Sahaj Agarwal, Ahmed Awadallah, Vidhisha Balachandran, Harkirat Behl, Lingjiao Chen, Gustavo de Rosa, Suriya Gunasekar, Mojan Javaheripi, Neel Joshi, Piero Kauffmann, Yash Lara, Caio César Teodoro Mendes, Arindam Mitra, Besmira Nushi, Dimitris Papailiopoulos, Olli...

    2025 arXiv

  17. [25]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. URL https://arxiv.org/abs/2402.03300

    2024 arXiv

  18. [26]

    Code-r1: Reproducing r1 for code with reliable rewards

    Jiawei Liu and Lingming Zhang. Code-r1: Reproducing r1 for code with reliable rewards. 2025

    2025

  19. [27]

    Reinforce++: An efficient rlhf algorithm with robustness to both prompt and reward models, 2025

    Jian Hu, Jason Klein Liu, and Wei Shen. Reinforce++: An efficient rlhf algorithm with robustness to both prompt and reward models, 2025. URL https://arxiv.org/abs/2501.03262

    2025 arXiv

  20. [28]

    Cot-valve: Length-compressible chain-of-thought tuning, 2025 a

    Xinyin Ma, Guangnian Wan, Runpeng Yu, Gongfan Fang, and Xinchao Wang. Cot-valve: Length-compressible chain-of-thought tuning, 2025 a . URL https://arxiv.org/abs/2502.09601

    2025 arXiv

  21. [29]

    Training language models to reason efficiently, 2025

    Daman Arora and Andrea Zanette. Training language models to reason efficiently, 2025. URL https://arxiv.org/abs/2502.04463

    2025

  22. [30]

    Aytes, Jinheon Baek, and Sung Ju Hwang

    Simon A. Aytes, Jinheon Baek, and Sung Ju Hwang. Sketch-of-thought: Efficient llm reasoning with adaptive cognitive-inspired sketching, 2025. URL https://arxiv.org/abs/2503.05179

    2025

  23. [31]

    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, 2024

    2024 arXiv

  24. [32]

    Reasoning models can be effective without thinking, 2025 b

    Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. Reasoning models can be effective without thinking, 2025 b . URL https://arxiv.org/abs/2504.09858

    2025 arXiv

  25. [33]

    Qwen2.5: A party of foundation models, September 2024

    Qwen Team. Qwen2.5: A party of foundation models, September 2024. URL https://qwenlm.github.io/blog/qwen2.5/

    2024

  26. [34]

    Gemini 2.5 flash

    Google. Gemini 2.5 flash. https://developers.googleblog.com/en/start-building-with-gemini-25-flash/, 2025 b

    2025

  27. [35]

    Distilling system 2 into system 1, 2024

    Ping Yu, Jing Xu, Jason Weston, and Ilia Kulikov. Distilling system 2 into system 1, 2024. URL https://arxiv.org/abs/2407.06023

    2024 arXiv

  28. [36]

    The 23rd international conference on artificial intelligence in medicine (aime 2025)

    International Conference on Artificial Intelligence in Medicine . The 23rd international conference on artificial intelligence in medicine (aime 2025). https://aime25.aimedicine.info/. Accessed: 2025-06-10

    2025

  29. [37]

    Let's verify step by step

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In The Twelfth International Conference on Learning Representations, 2023

    2023

  30. [38]

    Training verifiers to solve math word problems

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

    arXiv

  31. [39]

    American mathematics competitions (amc)

    Mathematical Association of America . American mathematics competitions (amc). https://maa-amc.org/student-programs/amc/. Accessed: 2025-06-10

    2025

  32. [40]

    Gpqa: A graduate-level google-proof q&a benchmark

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling, 2024

    2024

  33. [41]

    Livecodebench: Holistic and contamination free evaluation of large language models for code

    Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974, 2024. doi:10.48550...

    2024 doi arXiv

  34. [42]

    Trl: Transformer reinforcement learning

    Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, Shengyi Huang, Kashif Rasul, and Quentin Gallouédec. Trl: Transformer reinforcement learning. https://github.com/huggingface/trl, 2020

    2020

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.