Pith. sign in

REVIEW 3 major objections 6 minor 22 references

EvoThink claims that large reasoning models become simultaneously more efficient and more capable when redundant atomic reasoning units are pruned and the model is trained on synthesized aha-moment transitions from wrong to right.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 11:10 UTC pith:V5JOBVWF

load-bearing objection A novel self-training recipe that cuts tokens and lifts accuracy across 12 model-benchmark cells, but the self-pruning signal is not independently validated and the reporting has several warts. the 3 major comments →

arxiv 2607.19962 v1 pith:V5JOBVWF submitted 2026-07-22 cs.AI

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

Xinbang Dai , Zheyu Xin , Huikang Hu , Lin Ren , Rihui Jin , Guohui Xiao , Guilin Qi , Kuicai Dong
show 2 more authors
Zhaocheng Du Yuyang Zhang
This is my paper
classification cs.AI
keywords EvoThinklarge reasoning modelsoverthinkingself-pruningaha-momentpreference optimizationreasoning efficiencymathematical reasoning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Large reasoning models often 'overthink,' re-deriving and re-verifying the same conclusions and burning tokens without adding insight. EvoThink attacks this at the level of individual reasoning steps: self-pruning training (SPT) decomposes a reasoning trace into atomic reasoning units, deletes any unit whose local conclusion duplicates the previous unit's, and fine-tunes the model on the pruned traces; aha-moment preference optimization (AMPO) then selects failed attempts with the most diverse intermediate claims, synthesizes from-wrong-to-right transitions, and applies DPO so the model learns to recover from mistakes. The paper claims this two-stage recipe reduces inference-time token usage while improving Pass@1 on math and code benchmarks across models from 1.5B to 32B parameters. If correct, it means efficiency and reasoning quality are not in tension: pruning redundancy makes models cheaper and better at once.

Core claim

Central discovery: a reasoning trajectory decomposes into atomic reasoning units—contiguous spans ending in a local conclusion—and consecutive units with duplicate conclusions are redundant and prunable without changing the final answer. Self-training on these pruned traces yields shorter, more direct reasoning with equal or better accuracy. The most informative failures are those with the most diverse local conclusions, and training on synthesized from-wrong-to-right transitions ('But is this correct? I think I missed something.') improves capability more than training on gold solutions. Together, SPT and AMPO reduce tokens and raise Pass@1.

What carries the argument

The key object is the atomic reasoning unit: a self-contained span of a reasoning trace that starts a new attempt and closes with a local conclusion (a candidate direction, intermediate claim, or tentative answer). The pruning rule compares each unit's local conclusion with the preceding unit's; if they match, the later unit is marked redundant and deleted. For AMPO, the load-bearing device is the diversity fitness F(r)=|{A(u_i)}|, the number of distinct local conclusions in a trajectory, which selects which failed attempt becomes the seed for a synthesized aha-moment transition. The mutation is anchored by the sentence 'But is this correct? I think I missed something,' followed by the corre

Load-bearing premise

The load-bearing premise is that the model itself reliably segments its reasoning into atomic units and correctly judges when one unit's local conclusion duplicates another's; if this self-pruning signal is unreliable, the training data is corrupted and the measured gains may reflect length regularization rather than genuine removal of redundancy.

What would settle it

Compare the model's atomic-unit segmentation and redundancy judgments against human/expert annotation on a sample of trajectories; near-chance agreement would show the pruning signal is untrustworthy. Alternatively, train with random pruning of units to the same length: if random pruning matches SPT's accuracy, the redundancy targeting is not the cause of the improvement.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Combined EvoThink (SPT+AMPO) reduces average output length while raising Pass@1 relative to the vanilla model on all four reported benchmarks (MATH-500, AIME24, AIME25, TACO).
  • SPT alone—unsupervised, using no gold answers—cuts token usage substantially (e.g., roughly a third on many settings) while keeping accuracy at or above several supervised baselines.
  • AMPO's from-wrong-to-right mutation data transfers better than SFT or DPO on gold solutions: in the paper's cross-dataset experiments, the AMPO-trained model retains or improves AIME performance when trained on MATH, whereas direct-supervision baselines collapse.
  • High-diversity failed trajectories are more effective mutation seeds than low-diversity ones, supporting the claim that exploring new reasoning paths drives the capability gain.
  • The efficiency gains appear across difficulty levels, with the largest reduction in the heavy tail on hard problems, indicating that pruning targets the over-verification tail rather than uniformly shortening reasoning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • I infer that the atomic-unit pruning rule could be applied at inference time as a post-hoc filter without retraining; since the paper only demonstrates training-time integration, an inference-time filter is a direct testable extension.
  • Diversity-of-conclusions as a fitness for failure mining is a transferable heuristic: prefer trajectories that visit many distinct intermediate claims when building preference pairs for other domains such as code repair or theorem proving.
  • The self-pruning signal may drift as the model's style changes over SPT iterations, since the same model annotates and is trained; using a frozen or periodically re-set annotator could change the results, a stability question the paper leaves open.
  • If pruning specificity (targeting duplicated conclusions) is the real driver, then random pruning matched for length should underperform on hard benchmarks; this comparison would isolate the mechanism from mere length regularization.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes EvoThink, a two-stage framework for improving the inference efficiency and reasoning capability of large reasoning models. The first stage, Self-Pruning Training (SPT), uses the model itself as an annotator to segment its own reasoning traces into 'atomic reasoning units,' prune units whose local conclusions duplicate the preceding unit, and iteratively fine-tune the model on the pruned traces. The second stage, Aha-Moment Preference Optimization (AMPO), samples incorrect generations, scores them with a diversity-based fitness function that counts distinct local conclusions, selects the highest-diversity failure, synthesizes a 'from-wrong-to-right' continuation ending in the known correct answer, and applies DPO. The method is evaluated on MATH-500, AIME24, AIME25, and TACO with three backbone LRMs (DeepScaleR-1.5B, Distill-Qwen-1.5B, QwQ-32B). Table 1 reports that EvoThink SPT+AMPO simultaneously reduces token usage and improves Pass@1 relative to the vanilla models in all 12 model-benchmark cells.

Significance. If the reported results are reliable, EvoThink offers a practical and largely self-supervised route to cheaper LRM inference without sacrificing (and sometimes improving) accuracy. The idea of pruning at the granularity of atomic reasoning units rather than whole trajectories is a useful conceptual advance, and the component ablation separating SPT and AMPO is informative. The evaluation spans multiple backbones and benchmarks, which strengthens the generality of the claims. However, the central claims rest on two pillars that need reinforcement: the reliability of the model's self-annotated pruning/fitness signals and the statistical significance of the Pass@1 differences. At present the evidence is suggestive rather than conclusive, so my assessment is conditional.

major comments (3)
  1. [§3.1; §3.2] The validity of the self-pruning signal is load-bearing for SPT and, through fitness F(r)=|{A(u_i)}|, for AMPO. In §3.1 the annotator A is the initial model M0 itself; it both segments trajectories into atomic units and labels a unit redundant when its local conclusion duplicates the previous unit's conclusion. No evidence is presented that these judgments are reliable: no human agreement, no comparison with a stronger annotator, no check of whether pruning preserves the final answer. The paper's own ablations show cases where SPT alone degrades accuracy (e.g., DeepScaleR-1.5B AIME24: 26.7 vs 29.2; Distill-Qwen-1.5B MATH500: 73.5 vs 75.4), which is consistent with a noisy pruning signal. I request a validation study of the annotator's segmentations and redundancy labels (e.g., human agreement on a sample of trajectories, or an automated metric of conclusion equivalence), and a report of
  2. [§4.1; Table 1] Pass@1 is estimated from 4 samples per problem (MATH-500, TACO) or 8 samples (AIME24, AIME25), and only point estimates are reported. The central claim of consistent Pass@1 improvement over all 12 cells rests on differences that are small relative to sampling noise; for example, DeepScaleR-1.5B AIME25 (35.8 vs 30.8) and QwQ-32B TACO (11.9 vs 10.7) are within one or two correct answers on datasets of that size. Please report bootstrap confidence intervals or raw counts, and identify which of the 12 cells survive a significance test. Without this, the accuracy-improvement component of the central claim is not statistically established.
  3. [§3.1] SPT is described as 'unsupervised fine-tuning on a set of concise solutions that it self-generates,' but the pruned output \tilde{o}_i is obtained by deleting units while keeping the original answer \hat{a}_i. If the original trajectory is wrong, the pruned trajectory is also wrong. The paper does not state whether the 100-solution set is filtered for correctness or how correctness interacts with the pruning loop. Training on wrong-but-concise traces could directly explain the accuracy degradations observed for SPT alone in Table 1. Please clarify the filtering (or its absence) and, if unfiltered, quantify the fraction of correct trajectories in the SPT training data and assess the effect of that fraction on the downstream results.
minor comments (6)
  1. [§4.2] The sentence 'when the vanilla model is DeepScaleR-1.5B, ThinkPrune-2k's Pass@1 drops to 72.1% compared to the vanilla model's 75.4%... EVOTHINK SPT+AMPO maintains a higher Pass@1 of 76.3%' uses numbers that in Table 1 correspond to Distill-Qwen-1.5B, not DeepScaleR-1.5B. Please correct the attribution.
  2. [§1] The quantitative claim 'over 65% of the tokens are spent on redundant computations' appears to be derived from the single example in Figure 1. Either support it with an aggregate measurement across a representative sample or soften the claim.
  3. [§4.3] The stopping threshold epsilon_L=0.1 is justified only by the MATH-500 trajectory in Figure 3; for AIME and TACO the text says 'remains within ten' but does not report the actual stopping iterations. Please report these values or the S(i) curves for all datasets.
  4. [§4.5] The sensitivity analysis for K is shown only for QwQ-32B; the statement that the pattern 'is consistent across other models and datasets' is not accompanied by data. Please either provide the additional plots or label the claim as anecdotal.
  5. [§4.1] The reference for veRL is listed as 'Hybridflow'; please clarify the relationship between veRL and HybridFlow, and ensure all framework citations are precise.
  6. [Figure 6] The legend 'Max Min' is unclear; the text refers to 'maximum diversity (Max)' and 'minimum diversity (Min)'. Please make the legend self-contained and define the y-axis (performance gain vs absolute Pass@1).

Circularity Check

0 steps flagged

No significant circularity: EvoThink's self-referential training loop is a validity concern, not a derivation that closes on its own inputs.

full rationale

EvoThink's claimed results are empirical and anchored to external measurements. SPT uses the LRM itself (A) to segment reasoning into atomic units and to label units redundant when their local conclusions duplicate the prior unit, and AMPO uses A to score failure diversity and to synthesize from-wrong-to-right mutations with the gold answer supplied (Sec. 3.1, 3.2). This is self-referential data generation, but it does not make the reported Pass@1 or #Tok results circular: Pass@1 is exact-matched against gold answers on MATH-500/AIME24/AIME25/TACO, and token counts are measured on newly generated responses, not on the pruned training trajectories. The only identity by construction is that a unit judged redundant is removed ('if they are the same ... the model marks u_t as redundant and removes it'); that is the method's operational definition, not a predicted quantity. No load-bearing self-citation or author-imported uniqueness theorem appears; all cited baselines and datasets are external. The paper does not validate the atomic-unit segmentation or redundancy labels against a gold standard, and its own SPT ablation shows occasional accuracy regression (e.g., DeepScaleR-1.5B AIME24 26.7 vs 29.2), but that is a robustness/correctness concern about the training signal, not circularity. The conclusion's admission that understanding why from-wrong-to-right is effective is out of scope is a mechanistic gap, not a circular step.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central claims rest on self-generated domain assumptions: the model can reliably segment its own thoughts into atomic units, redundancy can be judged by duplicate local conclusions, diversity-as-count-of-conclusions selects valuable failures, and synthetic wrong-to-right transitions transfer to better reasoning. None of these are machine-checked or independently annotated; they are instrumented via prompts. External benchmarks provide the only independent grounding.

free parameters (5)
  • epsilon_L (SPT stopping threshold) = 0.1
    Termination threshold for iterative self-training; chosen because observed length reduction S(i) stabilizes below 0.1 within ~5 iterations in Figure 3 (left). This controls the efficiency/accuracy trade-off.
  • K (wrong-answer population size) = 6
    Number of sampled responses used to form the Wrong Answer Group; set after Figure 6 shows performance gain stabilizes for K>6 on QwQ-32B with MATH-500 and TACO. The paper claims consistency across models but shows only one model.
  • beta (DPO coefficient) = 0.5
    Strength of preference optimization; set to enforce closer fitting to the preference data. No sensitivity analysis is reported.
  • concise solution set size per SPT iteration = 100
    Fixed number of pruned outputs used for self-training each iteration; presented as a design choice without ablation.
  • aha-moment transition prompt template = "But is this correct? I think I missed something."
    Hand-crafted sentence used to synthesize mutation data in §3.2; no ablation on alternative transition prompts is reported.
axioms (5)
  • domain assumption Atomic-unit decomposition and redundancy comparison by the same model are reliable enough for self-training.
    SPT's pruning quality rests entirely on the annotator's segmentation and conclusion-similarity judgments (§3.1); no human evaluation is reported.
  • domain assumption The number of distinct local conclusions F(r) is a valid proxy for the value of a failed reasoning trajectory.
    AMPO selects the trajectory maximizing F(r)=|C(r)| (§3.2); Figure 6 provides an empirical Max-vs-Min comparison but no independent validation of the fitness construct.
  • domain assumption From-wrong-to-right mutation data teaches a generalizable recovery pattern rather than answer memorization.
    AMPO's DPO pairs synthetic mutation data against original failures (§3.2); the transfer experiment (§4.4) is evidence but assumes the synthetic transition is the causal ingredient.
  • domain assumption Gold labels and gold solutions in MATH, AIME (AoPS), and TACO are correct and representative.
    Used both as supervision for mutation data and as ground truth for Pass@1 (§4.1).
  • domain assumption Pass@1 estimated from 4 or 8 samples with temperature 0.6/top-p 0.95 is a stable accuracy estimate.
    Point estimates are reported without error bars; differences of 1-2 points may fall within sampling noise.
invented entities (2)
  • Atomic reasoning unit no independent evidence
    purpose: Discrete segment of a <think> trajectory used as the unit of pruning and of fitness counting.
    Defined by prompting the annotator, not by an external annotation or formal criterion; its objectivity is assumed (§3.1).
  • Aha-moment mutation data no independent evidence
    purpose: Synthetic training examples that splice a model's wrong reasoning prefix to a correct answer via a transition sentence.
    The construct is operationalized only through the prompt and downstream performance; there is no independent measurement of an 'aha moment' (the case study in §4.6 is illustrative).

pith-pipeline@v1.3.0-alltime-deepseek · 12484 in / 13660 out tokens · 129714 ms · 2026-08-01T11:10:10.395443+00:00 · methodology

0 comments
read the original abstract

Large Reasoning Models (LRMs) often suffer from overthinking due to redundant verification steps. Existing approaches for mitigating overthinking, such as fast-slow thinking switching and reasoning trajectory compression, fail to make a fine-grained distinction between beneficial and redundant steps within the LRM's reasoning process, and may thus impair reasoning capability in their pursuit of efficiency. To simultaneously improve reasoning efficiency and capability, we propose EvoThink, a framework that reduces redundant verification and encourages the exploration of new reasoning paths. EvoThink comprises two key components: Self-Pruning Training (SPT), an unsupervised method that iteratively prunes redundant reasoning steps and self-trains on the concise trajectories; and Aha-Moment Preference Optimization (AMPO), which, inspired by genetic algorithms, identifies valuable failed reasoning attempts, synthesizes from-wrong-to-right aha-moment data, and optimizes the model to internalize this reasoning pattern. Extensive evaluations across mathematical reasoning and code generation benchmarks demonstrate that EvoThink not only substantially reduces inference-time token usage but also improves the reasoning capability of LRMs.

Figures

Figures reproduced from arXiv: 2607.19962 by Guilin Qi, Guohui Xiao, Huikang Hu, Kuicai Dong, Lin Ren, Rihui Jin, Xinbang Dai, Yuyang Zhang, Zhaocheng Du, Zheyu Xin.

Figure 1
Figure 1. Figure 1: An example of overthinking in DeepSeek-R1. The [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Self-Pruning Training learns concise reasoning by pruning redundant units, whereas Aha-Moment Preference Optimization uses a [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of reasoning lengths across difficulty levels [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Evolution of reasoning length under EVOTHINKSPT on MATH (left). Correlation between reasoning length and the number of local conclusions on MATH (right). 4.3 Improvement of Reasoning Efficiency [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance improvements of DeepScaleR-1.5B under [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: A case study on a math problem from AIME25. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance gain vs. initial population size and reasoning [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 11 linked inside Pith

  1. [1]

    Sketch-of-thought: Efficient LLM rea- soning with adaptive cognitive-inspired sketching

    [Ayteset al., 2025 ] Simon A Aytes, Jinheon Baek, and Sung Ju Hwang. Sketch-of-thought: Efficient LLM rea- soning with adaptive cognitive-inspired sketching. InPro- ceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 24296–24320,

    2025

  2. [4]

    Optimizing length compression in large reasoning models.arXiv preprint arXiv:2506.14755,

    [Chenget al., 2025b ] Zhengxiang Cheng, Dongping Chen, Mingyang Fu, and Tianyi Zhou. Optimizing length compression in large reasoning models.arXiv preprint arXiv:2506.14755,

    Pith/arXiv arXiv

  3. [5]

    The danger of overthinking: Examin- ing the reasoning-action dilemma in agentic tasks.arXiv preprint arXiv:2502.08235,

    [Cuadronet al., 2025 ] 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: Examin- ing the reasoning-action dilemma in agentic tasks.arXiv preprint arXiv:2502.08235,

    Pith/arXiv arXiv 2025

  4. [6]

    Deepseek-r1: In- centivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

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

    Pith/arXiv arXiv 2025

  5. [7]

    Token- budget-aware llm reasoning

    [Hanet al., 2025 ] Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. Token- budget-aware llm reasoning. InFindings of the Associ- ation for Computational Linguistics: ACL 2025, pages 24842–24855,

    2025

  6. [8]

    Measuring mathemat- ical problem solving with the math dataset

    [Hendryckset al., 2021 ] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathemat- ical problem solving with the math dataset. InProceed- ings of the Neural Information Processing Systems Track on Datasets and Benchmarks, volume 1,

    2021

  7. [10]

    Taco: Topics in algorithmic code generation dataset.arXiv preprint arXiv:2312.14852,

    [Liet al., 2023 ] Rongao Li, Jie Fu, Bo-Wen Zhang, Tao Huang, Zhihong Sun, Chen Lyu, Guang Liu, Zhi Jin, and Ge Li. Taco: Topics in algorithmic code generation dataset.arXiv preprint arXiv:2312.14852,

    Pith/arXiv arXiv 2023

  8. [12]

    O1-pruner: Length-harmonizing fine- tuning for o1-like reasoning pruning.arXiv preprint arXiv:2501.12570,

    [Luoet al., 2025a ] 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,

    Pith/arXiv arXiv

  9. [13]

    [Maet al., 2025 ] Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia

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

    Pith/arXiv arXiv 2025

  10. [14]

    s1: Simple test-time scaling

    [Muennighoffet al., 2025 ] Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Cand`es, and Tatsunori B Hashimoto. s1: Simple test-time scaling. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 20286–20332,

    2025

  11. [15]

    DynaThink: Fast or slow? a dynamic decision-making framework for large language models

    [Panet al., 2024 ] Jiabao Pan, Yan Zhang, Chen Zhang, Zuozhu Liu, Hongwei Wang, and Haizhou Li. DynaThink: Fast or slow? a dynamic decision-making framework for large language models. InProceedings of the 2024 Con- ference on Empirical Methods in Natural Language Pro- cessing, pages 14686–14695,

    2024

  12. [16]

    Direct preference optimization: Your lan- guage model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741,

    [Rafailovet al., 2023 ] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your lan- guage model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741,

    2023

  13. [17]

    DAST: Difficulty-adaptive slow-thinking for large reasoning mod- els

    [Shenet al., 2025 ] Yi Shen, Jian Zhang, Jieyun Huang, Shuming Shi, Wenjing Zhang, Jiangze Yan, Ning Wang, Kai Wang, Zhaoxiang Liu, and Shiguo Lian. DAST: Difficulty-adaptive slow-thinking for large reasoning mod- els. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 2322–2331,

    2025

  14. [18]

    Hybridflow: A flexible and efficient rlhf framework

    [Shenget al., 2025 ] Guangming Sheng, Chi Zhang, Zil- ingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. Hybridflow: A flexible and efficient rlhf framework. InProceedings of the Twen- tieth European Conference on Computer Systems, pages 1279–1297,

    2025

  15. [19]

    Dualformer: Controllable fast and slow thinking by learning with ran- domized reasoning traces

    [Suet al., 2025 ] DiJia Su, Sainbayar Sukhbaatar, Michael Rabbat, Yuandong Tian, and Qinqing Zheng. Dualformer: Controllable fast and slow thinking by learning with ran- domized reasoning traces. InThe Thirteenth International Conference on Learning Representations,

    2025

  16. [20]

    Stop overthinking: A survey on efficient reasoning for large lan- guage models.arXiv preprint arXiv:2503.16419,

    [Suiet al., 2025 ] Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Na Zou, et al. Stop overthinking: A survey on efficient reasoning for large lan- guage models.arXiv preprint arXiv:2503.16419,

    Pith/arXiv arXiv 2025

  17. [21]

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

    Pith/arXiv arXiv 2025

  18. [22]

    Qwen2.5 technical report.arXiv preprint arXiv:2412.15115, 2024

    [Yanget al., 2024 ] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Ji- axi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin...

    Pith/arXiv arXiv 2024

  19. [2021]

    Thinkprune: Pruning long chain-of-thought of llms via re- inforcement learning.arXiv preprint arXiv:2504.01296,

    [Houet al., 2025 ] Bairu Hou, Yang Zhang, Jiabao Ji, Yu- jian Liu, Kaizhi Qian, Jacob Andreas, and Shiyu Chang. Thinkprune: Pruning long chain-of-thought of llms via re- inforcement learning.arXiv preprint arXiv:2504.01296,

    Pith/arXiv arXiv 2025

  20. [2023]

    ThinkSwitcher: When to think hard, when to think fast

    [Lianget al., 2025 ] Guosheng Liang, Longguang Zhong, Ziyi Yang, and Xiaojun Quan. ThinkSwitcher: When to think hard, when to think fast. InFindings of the Associ- ation for Computational Linguistics: EMNLP 2025, pages 5185–5201,

    2025

  21. [2024]

    The overthinker’s DIET: Cutting token calories with DIfficulty-aware training

    [Chenet al., 2025 ] Weize Chen, Jiarui Yuan, Tailin Jin, Ning Ding, Huimin Chen, Zhiyuan Liu, and Maosong Sun. The overthinker’s DIET: Cutting token calories with DIfficulty-aware training. InThe Thirty-ninth Annual Con- ference on Neural Information Processing Systems,

    2025

  22. [2025]

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

    [Chenet al., 2024 ] 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,

    Pith/arXiv arXiv 2024