Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Overclocking LLM Reasoning: Monitoring and Controlling Thinking Path Lengths in LLMs

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Reasoning LLMs track their own thinking progress, and shifting that internal signal during inference can shorten the thinking phase, cut latency, and improve answer accuracy under tight token budgets.

desk verdict Useful monitoring/steering idea, but the causal story is under-supported and the accuracy claims are overstated; worth a serious look with added controls. read the letter →

arxiv 2506.07240 v1 pith:CMPQKTXV submitted 2025-06-08 cs.LG cs.AI

Roy Eisenstadt , Itamar Zimerman , Lior Wolf This is my paper
classification cs.LGcs.AI
keywords reasoningmodelsthinkingprogressvectoroverclockingoverthinkinghidden-stateinterventiontest-timescalingmechanisticinterpretabilitystructured
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

Reasoning models that mark off a `` phase need to decide when to stop deliberating. This paper claims that such models keep a running estimate of how far they are through the thinking phase, encoded linearly in the last-layer hidden states, and that this estimate is not a passive by-product but a signal the model actually uses. Reading that signal gives an interactive progress bar for the model's reasoning. More strongly, nudging the hidden state along the learned progress direction—the 'overclocking' intervention—shortens the thinking phase, cuts inference cost, and can raise the number of correct answers under tight token budgets. The paper positions this as the first direct evidence of an internal, causally active progress estimate in explicit reasoning.

What carries the argument

The load-bearing object is the thinking progress vector (TPV), a parameter vector $\theta \in \mathbb{R}^d$ fit by least-squares regression to predict the relative position $p = j/N_k$ of each thinking-phase token from its final-layer hidden state $h_j^{(k)}$. Its role is to turn an opaque hidden state into a scalar progress estimate, and then to act as an intervention direction: adding $\alpha\theta$ to $h$ raises $\theta^\top h_\alpha = \bar{p} + \alpha\|\theta\|^2$, pushing the model toward the 'later in the thinking phase' region of representation space. Because the edit is applied after the attention layers, each step's perturbation influences only the immediate next-token distribution while the autoregressive history still carries earlier edited tokens. A single-layer GRU over the same hidden states provides a smoother monitoring signal, but the linear TPV is the object used for control.

What would settle it

Run the same intervention with a random unit vector, or with a vector trained on shuffled position labels, in place of $\theta$, matching $\|\theta\|$ and $\alpha$. If random projections shorten the thinking phase as much as the TPV does, length control is not specific to progress encoding; a complementary check is whether overclocking preserves accuracy on non-mathematical open-ended reasoning tasks, where progress is harder to define.

Watch

Extended reading notes

Core claim

The paper's central claim is that a deep reasoning model tracks its own relative position inside the explicit thinking phase, and that this tracking can be both observed and manipulated. The authors train a linear projection (a 'thinking progress vector') on last-layer hidden states to predict the normalized token position $j/N$ within `<think>...</think>`, and show that the prediction transfers across prompts and datasets and improves with smoothing or a sequence model. They then intervene at every decoding step after the attention layers with $h_\alpha = h + \alpha\theta$, which changes the predicted progress but leaves cached states untouched. Positive $\alpha$ consistently shortens the generated thinking trajectory. Under low token budgets this shortening is not just cheaper: on Math-500 and GSM8K it produces more completed, answered, and correct responses than the base model and than temperature- or instruction-based acceleration baselines, and it combines with those baselines for further gains.

Load-bearing premise

The load-bearing premise is that shortened thinking is caused by the model's true sense of progress, not by the nudged vector simply biasing it toward stopping tokens; if a random direction of the same size also shortens reasoning, the progress-specific explanation would collapse.

Editorial extensions

If this is right

  • A reasoning model can be made more concise at inference time with no fine-tuning: shifting hidden states along the progress vector shortens thinking across two model sizes and two math benchmarks.
  • Accuracy under small token budgets rises with the intervention, so overclocking is a practical counter to overthinking when compute or context is limited.
  • The progress signal is complementary to prompt-based acceleration; combining TPV intervention with an instruction to be concise gives the largest gains in the paper's tables.
  • The same hidden-state readout can serve as a real-time progress bar, making the otherwise invisible thinking phase inspectable by users.
  • Token-level dips in predicted progress line up with self-verification behavior, so the readout exposes when the model re-solves a problem rather than advancing.

Reading between the lines

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

  • If the progress estimate is genuinely causal, the learned vector could become an early-stopping rule: generate until smoothed predicted progress crosses a threshold, then cut the remaining thinking tokens, a control loop the paper does not implement.
  • The method's open-domain reach is untested; applying the same regression and intervention to open-ended reasoning would clarify whether the progress signal is a general planning mechanism or a math-reasoning artifact.
  • A stronger causal check than the paper reports would compare the intervention with random projection directions of equal norm; if those shorten thinking too, the effect may be a generic output bias rather than a dedicated progress mechanism.
  • Because the intervention edits only the final hidden layer, it could be combined with speculative decoding or per-problem token-budget schedulers that decide how aggressively to overclock each question.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper studies whether reasoning LLMs encode their relative position inside the explicit <think>...</think> thinking phase. The authors train a linear probe (TPV) on last-layer hidden states to predict p = j/N (Equation 1), report low test MSE, and introduce an RNN/GRU-based smoother for visualizing progress. They then intervene by computing h_alpha = h + alpha*theta, reporting that positive alpha shortens thinking, increases the number of correct answers under token budgets, and reduces inference latency on Math500 and GSM8K for two DeepSeek-R1 distilled models. They interpret the results as evidence of an internal progress estimate and present 'overclocking' as a control mechanism that mitigates overthinking.

Significance. If the central causal claim were established, the paper would be a useful contribution: it connects monitoring and control of reasoning length, proposes a lightweight inference-time intervention that does not require fine-tuning, and publicly releases code. The progress-bar visualization is a practical diagnostic for reasoning models. However, the significance is currently limited by the absence of control interventions and by the lack of statistical rigor. The monitoring result is a correlational probe, and the intervention result is equally consistent with a generic 'conclude now' output bias. The paper's own tables also contradict the claim that accuracy improves without increasing errors. These issues must be addressed before the paper can support its strongest conclusions.

major comments (4)
  1. [Section 3.2 (h_alpha = h + alpha*theta)] The intervention shifts the hidden state along the direction trained to predict p=j/N, so the increase in predicted progress shown in Figure 6 is partly by construction. The observed shortening of thinking could be produced by a generic 'conclude now' bias rather than by an internal progress estimate. I ask for control interventions: random projection directions, directions trained on shuffled labels, and the direction that maximizes the logit of the </think> token. Without such controls, the causal interpretation in Sections 3.2 and 5 is underdetermined.
  2. [Section 4.1, Table 1] The statement that 'these increases in correct answers do not come at the cost of more errors, as the error rate remains unchanged' is contradicted by Table 1. On Math500 at 1024 tokens, #Answered - #Correct is 2 for Base and 12 for TPV alpha=100; on GSM8K at 512, it is 2 vs 11. Thus error counts and error rates increase in several regimes. The claim that overclocking improves accuracy without increasing errors is not supported by the reported data.
  3. [Section 4, Tables 1-2] All accuracy and token-length numbers are single-run raw counts with no variance, no repeated seeds, and no significance tests. Because generation uses temperature 0.6, the differences between methods may fall within sampling noise. The paper should report multiple runs or confidence intervals, or at least apply a bootstrap or other statistical test over problems, before claiming improvements of 80% or more.
  4. [Section 4.1, point (3)] The claim that 'our method outperforms both [baselines] by producing more correct answers' is internally inconsistent with Table 1: at Math500 with a 2048-token budget, Instruct yields 316 correct vs 300 for TPV alpha=100 and 300 for TPV alpha=100 Ins. This is later acknowledged in point (4), but the earlier blanket statement should be revised or qualified to reflect the regimes where prompting is stronger.
minor comments (4)
  1. [Section 4.1, point (2)] The text says 'under the same 512-token limit on GSM-8K' but then cites completion increases from 28 to 43 and 52, which are the Math500 512-token numbers; the GSM8K 512 completions are 227, 248, and 232. This conflation makes the baseline comparison difficult to follow.
  2. [Tables 1-2] Several table cells are merged or truncated (e.g., '316321 296' and '40247263 261'), making the data hard to verify. Please reformat the tables so every entry is clearly separated.
  3. [Example 2 (Appendix D)] The final answer in the overclocked output reads 'The number of ways is \boxed{36}' but should refer to miles rather than ways.
  4. [References and Section 3.1] References [10] and [11] appear to cite the same work with different venues and should be deduplicated. Also, Section 3.1 introduces a 'GRU' while Figures 3 and 4 say 'RNN'; please align the terminology.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: the intervention's predicted-progress readout is definitionally biased by the regression vector, but the overclocking token/accuracy results are empirical and give the central claim independent content.

  1. self definitional [Section 3.2 (Intervention Technique, h_alpha = h + alpha theta) and the discussion of Figure 6]
    "The intervention experiments shift the hidden representation h in the direction of projection vector θ by an amount α: hα =h+αθ . The altered representation has a new prediction value:θ T hα =θ T (h+αθ) = ¯p+α||θ|| 2."

    The vector θ was itself fitted to predict the relative-position labels p=j/N. Therefore the displayed 'predicted progress' of an intervened trajectory is raised by exactly α||θ||^2 by the intervention equation alone. The paper's statement that 'the intervened trajectory shows consistently higher predicted progress beyond a certain decoding step' is thus a restatement of the construction of θ and the intervention, not an independent measurement of a progress-tracking mechanism. The shortening in token count and the accuracy changes reported in Tables 1-2 are not similarly forced by the equations, so the circularity is partial rather than total.

full rationale

The monitoring half of the paper is not circular: a linear regressor is trained on held-out thinking trajectories with labels p=j/N, and its generalization to unseen trajectories is a legitimate empirical result. The circular step is confined to Section 3.2: because the intervention uses the same θ that defines the progress readout, the equation θ^T h_alpha = θ^T h + α||θ||^2 guarantees an increased predicted progress value. Consequently, Figure 6's higher predicted progress after intervention is constructed rather than discovered. The overclocking effects themselves—shorter thinking, more completed answers, and more correct answers under increased α—are empirical outcomes that do not reduce by construction, though their causal attribution to a specific 'progress estimate' is weakened by the absence of control directions such as random projections or end-of-thinking-token vectors. That absence is a validity/interpretability concern rather than an additional circular step. No load-bearing self-citations were found, and the paper's central claim retains independent empirical content, supporting a moderate circularity score rather than a high one.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

The method rests on a fitted probe, a hand-chosen intervention strength, and assumptions about where progress lives and how to intervene. The paper introduces no physical entities, but it posits two constructs: the TPV and an internal progress estimate. Neither has independent evidence outside the paper, so the ledger is dominated by fitted parameters and domain assumptions.

free parameters (3)
  • Thinking progress vector theta = Learned least-squares weights (dimension d, not reported numerically)
    Trained on hidden states from a small in-distribution sample (80 Math500, 30 GSM8K problems per model) to predict p=j/N; it is the object used both for monitoring and for the intervention.
  • Intervention strength alpha = Values 5, 25, 50, 100; headline numbers use 5 and 100
    Chosen by hand as a hyperparameter; no validation procedure or theory-based selection is given, and results vary strongly with it.
  • RNN progress predictor parameters = Learned GRU weights (dimension d), not reported
    Used for the monitoring comparison in Figures 3-4; not part of overclocking, but part of the evidence that progress is linearly decodable.
assumptions (4)
  • domain assumption Final-layer hidden states suffice for progress extraction and intervention
    Section 3.1 selects the final hidden layer for representation richness and proximity to the embedding; no layer ablation is reported.
  • domain assumption The relative-position label p=j/N is the correct ground truth for progress
    Eq. (1) assigns labels by linear interpolation; the paper assumes this matches an internal estimate rather than testing alternatives like absolute step or task difficulty.
  • domain assumption Post-attention intervention leaves cached states unmodified
    Section 3.2 relies on this to isolate the intervention to a single step; if attention caches are affected, the measured effect could be confounded with memory corruption.
  • domain assumption DeepSeek-R1-Distill models and math benchmarks represent the intended scope
    All experiments use two distill models on two math datasets; the Discussion's overclocking claims presuppose transfer beyond this setting without evidence.
invented entities (2)
  • Thinking progress vector (TPV)
    purpose: A linear projection of hidden states claimed to encode relative position inside <think>...</think>; used as a progress bar and as a steering direction.
    The vector is fit on the same models and datasets used for evaluation; no external benchmark or control direction independently validates it as a dedicated progress mechanism.
  • Internal progress estimate
    purpose: A hypothesized model-internal scalar tracking how close the model is to closing its thinking phase.
    Inferred from probe accuracy and intervention; the paper provides no separate neural or behavioral evidence that the model computes such an estimate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Overclocking LLM Reasoning: Monitoring and Controlling Thinking Path Lengths in LLMs." pith.science (2026). https://pith.science/paper/CMPQKTXV

@misc{pith2026250607240,
  author       = {Pith},
  title        = {Pith review of: Overclocking LLM Reasoning: Monitoring and Controlling Thinking Path Lengths in LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CMPQKTXV}},
  note         = {Machine review of arXiv:2506.07240}
}
read the original abstract

Recently, techniques such as explicit structured reasoning have demonstrated strong test-time scaling behavior by enforcing a separation between the model's internal "thinking" process and the final response. A key factor influencing answer quality in this setting is the length of the thinking stage. When the reasoning is too short, the model may fail to capture the complexity of the task. Conversely, when it is too long, the model may overthink, leading to unnecessary computation and degraded performance. This paper explores and exploits the underlying mechanisms by which LLMs understand and regulate the length of their reasoning during explicit thought processes. First, we show that LLMs encode their progress through the reasoning process and introduce an interactive progress bar visualization, which is then used to reveal insights on the model's planning dynamics. Second, we manipulate the internal progress encoding during inference to reduce unnecessary steps and generate a more concise and decisive chain of thoughts. Our empirical results demonstrate that this "overclocking" method mitigates overthinking, improves answer accuracy, and reduces inference latency. Our code is publicly available.

Figures

Figures reproduced from arXiv: 2506.07240 by the authors.

Figure 1
Figure 1. Applications of our method. (a) Monitoring the reasoning progress. (b) Overclocking it. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Linear vs Non-linear regression MSE loss for monitoring the thinking phase. For each model-dataset pair, 30 problems were sampled [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Qualitative analysis of RNN-based progress prediction: Predictions across time steps during the thinking phase. (a) Results on MATH-500 (in-domain); (b) results on GSM8K (zero-shot). in the Math-500 test set. As shown in the figures, both approaches, with and without smoothing, successfully predict the relative position, while the latter produces more precise results that can be used to create a clearer, more interp… view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Relative Progress Prediction Analysis: (a) We sample 1000 random tokens from the thinking trajectories across all examples in the test set of the Math-500 dataset. The x-axis denotes the ground-truth relative position within the thinking phase, while the y-axis shows t…
Figure 6
Figure 6. Figure 6: Predicted progress (p¯) values for two thinking sequences, one from the original generation and the other from the TPV-intervened process, on the same prompt in Example 1, using DeepSeek￾R1-Distill-Qwen-32B. For this prompt the original generation does not converge to …
Figure 7
Figure 7. Figure 7: Impact of α on Intervention: The x-axis represents different values of α in {0, 5, 25, 50, 100}, where α = 0 indicates no intervention and higher values of α amplify the amount of intervention. The y-axis shows the average number of tokens required to complete the answ…
Figure 8
Figure 8. Figure 8: Robustness of predicted progress under varied prompting styles and reasoning lengths. Left: Predicted and smoothed progress trajectories for the problem “How many 3-digit numbers can be formed using the digits 1 through 9 (no zeros), with no repeated digits?” under fou…
Figure 9
Figure 9. Figure 9: A math question and the associated thinking sequence generated by DeepSeek-R1-Distill [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Average change in p¯ values caused by significant tokens across the dataset, highlighting their impact on predicted progress. “hmm,” “wait,” and “right” often signal meaningful transitions in the model’s internal state. Some of these tokens reflect a recognition of in…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

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

    cs.AI 2025-07 conditional novelty 3.0 of 10

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

Reference graph

Works this paper leans on

41 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [1]

    L1: Controlling how long a reasoning model thinks with reinforcement learning.arXiv preprint arXiv:2503.04697, 2025

    Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning.arXiv preprint arXiv:2503.04697, 2025

    arXiv 2025

  2. [2]

    Do not think that much for 2+ 3=? on the overthinking of o1-like llms.arXiv preprint arXiv:2412.21187, 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, 2024

    arXiv 2024

  3. [3]

    A toy model of universality: Reverse engineering how networks learn group operations

    Bilal Chughtai, Lawrence Chan, and Neel Nanda. A toy model of universality: Reverse engineering how networks learn group operations. InInternational Conference on Machine Learning, pages 6243–6267. PMLR, 2023. 11

    work page 2023

  4. [4]

    Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    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

  5. [5]

    A mathematical framework for transformer circuits.Transformer Circuits Thread,

    Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Andy Jones, Jackson Kernion, Liane Lovitt, Kamal Ndousse, Dario Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Chris Olah. A...

  6. [6]

    Metacognition and cognitive monitoring: A new area of cognitive– developmental inquiry.American psychologist, 34(10):906, 1979

    John H Flavell. Metacognition and cognitive monitoring: A new area of cognitive– developmental inquiry.American psychologist, 34(10):906, 1979

    work page 1979

  7. [7]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    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, 2025

    arXiv 2025

  8. [8]

    In-context learning creates task vectors

    Roee Hendel, Mor Geva, and Amir Globerson. In-context learning creates task vectors. In Conference on Empirical Methods in Natural Language Processing, 2023

    work page 2023

Show all 41 references
  1. [9]

    Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    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

    2024 arXiv

  2. [10]

    The impact of reasoning step length on large language models

    Mingyu Jin, Qinkai Yu, Dong Shu, Haiyan Zhao, Wenyue Hua, Yanda Meng, Yongfeng Zhang, and Mengnan Du. The impact of reasoning step length on large language models. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors,Findings of the Association for Computational Linguisti...

    2024 doi

  3. [11]

    The impact of reasoning step length on large language models.arXiv preprint arXiv:2401.04925, 2024

    Mingyu Jin, Qinkai Yu, Dong Shu, Haiyan Zhao, Wenyue Hua, Yanda Meng, Yongfeng Zhang, and Mengnan Du. The impact of reasoning step length on large language models.arXiv preprint arXiv:2401.04925, 2024

    2024 arXiv

  4. [12]

    Language models use trigonometry to do addition

    Subhash Kantamneni and Max Tegmark. Language models use trigonometry to do addition. InICLR 2025 Workshop on Building Trust in Language Models and Applications, 2025. URL https://openreview.net/forum?id=CqViN4dQJk

    2025

  5. [13]

    Large language models are zero-shot reasoners.Advances in neural information processing systems, 35:22199–22213, 2022

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

    2022

  6. [14]

    Abstractive document summa- rization with summary-length prediction

    Jingun Kwon, Hidetaka Kamigaito, and Manabu Okumura. Abstractive document summa- rization with summary-length prediction. In Andreas Vlachos and Isabelle Augenstein, edi- tors,Findings of the Association for Computational Linguistics: EACL 2023, pages 618–624, Dubrovnik, Croat...

    2023 doi

  7. [15]

    Let’s verify step by step.arXiv preprint arXiv:2305.20050, 2023

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step.arXiv preprint arXiv:2305.20050, 2023

    2023 arXiv

  8. [16]

    s1: Simple test-time scaling.arXiv preprint arXiv:2501.19393, 2025

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling.arXiv preprint arXiv:2501.19393, 2025

    2025 arXiv

  9. [17]

    Progress mea- sures for grokking via mechanistic interpretability

    Neel Nanda, Lawrence Chan, Tom Lieberum, Jess Smith, and Jacob Steinhardt. Progress mea- sures for grokking via mechanistic interpretability. InThe Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=9XFSbDPmdW. 12

    2023

  10. [18]

    Metamemory: A theoretical framework and new findings

    Thomas O Nelson. Metamemory: A theoretical framework and new findings. InPsychology of learning and motivation, volume 26, pages 125–173. Elsevier, 1990

    1990

  11. [19]

    Zoom in: An introduction to circuits.Distill, 2020

    Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. Zoom in: An introduction to circuits.Distill, 2020. doi: 10.23915/distill.00024.001. https://distill.pub/2020/circuits/zoom-in

    2020 doi

  12. [20]

    In-context learning and induction heads.arXiv preprint arXiv:2209.11895, 2022

    Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, et al. In-context learning and induction heads.arXiv preprint arXiv:2209.11895, 2022

    2022 arXiv

  13. [21]

    Chatgpt: Optimizing language models for dialogue

    OpenAI. Chatgpt: Optimizing language models for dialogue. https://openai.com/ chatgpt, 2023. Accessed: 2025-05-11

    2023

  14. [22]

    Zero-shot strategies for length-controllable summa- rization.arXiv preprint arXiv:2501.00233, 2024

    Fabian Retkowski and Alexander Waibel. Zero-shot strategies for length-controllable summa- rization.arXiv preprint arXiv:2501.00233, 2024

    2024 arXiv

  15. [23]

    Scaling llm test-time compute opti- mally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314, 2024

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute opti- mally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314, 2024

    2024 arXiv

  16. [24]

    Between underthinking and overthinking: An empirical study of reasoning length and correctness in llms.arXiv preprint arXiv:2505.00127, 2025

    Jinyan Su, Jennifer Healey, Preslav Nakov, and Claire Cardie. Between underthinking and overthinking: An empirical study of reasoning length and correctness in llms.arXiv preprint arXiv:2505.00127, 2025

    2025 arXiv

  17. [25]

    Stop overthinking: A survey on efficient reasoning for large language models.URL https://arxiv

    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.URL https://arxiv. org/abs/2503.16419, 2025

    2025 arXiv

  18. [26]

    Thoughts are all over the place: On the underthinking of o1-like llms.arXiv preprint arXiv:2501.18585, 2025

    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, 2025

    2025 arXiv

  19. [27]

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

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

    2022

  20. [28]

    From decoding to meta-generation: Inference-time algorithms for large language models.arXiv preprint arXiv:2406.16838, 2024

    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, 2024

    2024 arXiv

  21. [29]

    Effectively controlling reasoning models through thinking intervention.arXiv preprint arXiv:2503.24370, 2025

    Tong Wu, Chong Xiang, Jiachen T Wang, and Prateek Mittal. Effectively controlling reasoning models through thinking intervention.arXiv preprint arXiv:2503.24370, 2025

    2025 arXiv

  22. [30]

    When more is less: Understanding chain-of-thought length in llms.arXiv preprint arXiv:2502.07266, 2025

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

    2025 arXiv

  23. [31]

    The clock and the pizza: Two stories in mechanistic explanation of neural networks

    Ziqian Zhong, Ziming Liu, Max Tegmark, and Jacob Andreas. The clock and the pizza: Two stories in mechanistic explanation of neural networks. InThirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id= S5wmbQc1We

    2023

  24. [32]

    hmm,” “wait,

    Barry J Zimmerman. Becoming a self-regulated learner: An overview.Theory into practice, 41 (2):64–70, 2002. 13 A Experimental Setup This section describes the experimental setup used for the empirical studies in Sec. 4. All experiments were conducted in PyTorch with official H...

    2002

  25. [34]

    **Calculate the number of 3-day periods in 28 days:** 28 3 ≈9.33 Since Suzanne can’t walk a fraction of a 3- day period, we take the integer part, which is 9

  26. [35]

    **Determine the total miles walked:** 9×4 = 36miles

  27. [36]

    Suzanne doesn’t walk on the 28th day, so 36 miles is the minimum

    **Check if 28 days is a multiple of 3:** 28÷3 = 9with a remainder of 1 Since 28 is not a multiple of 3, 9 periods of 3 days account for 27 days, leaving 1 day. Suzanne doesn’t walk on the 28th day, so 36 miles is the minimum. **Final Answer** The number of ways is \boxed{36}. ...

  28. [37]

    Each time, I’m adding the two previous numbers to get the next one

    F(9) = 21 + 13 = 34 Yes, that seems correct. Each time, I’m adding the two previous numbers to get the next one. So, the 9th Fibonacci number is 34. . . . let me recall another way. . . I remember there’s Binet’s formula, which uses the golden ratio. The formula is: F(n) = (Φn...

  29. [38]

    First, compute

    Let me try using this formula to compute F(9) . . . First, compute . . . Now, compute . . . Let me compute this step by step: . . . . . . Similarly, compute . . . . . . Now, plug these into Binet’s formula: F(9) = (Φ9 −Ψ 9)/ √ 5≈ (76.013−(−0.01316))/2.23607≈. . .≈34.0 Wow, tha...

  30. [39]

    So, I’m confident that the 9th Fibonacci number is 34

    F(9) = 34 Yep, that’s consistent. So, I’m confident that the 9th Fibonacci number is 34. </think> Example 3: An example of overthinking through excessive self-verification and redundant, duplicated computation in the thinking process of the base model (right), which is mitigat...

  31. [40]

    Substitute X back into the first equation to find Y: 13 4 +Y= 5 Y= 5− 13 4 = 20 4 − 13 4 = 7 4

  32. [41]

    I need to solve for X and Y

    Final Answer: X= 13 4 , Y= 7 4 </think> Original (892 Thinking Tokens) <think>Okay, so I have these two equations here: X + Y = 5 and 3X - Y = 8. I need to solve for X and Y . Hmm, let me think about how to approach this. I remember from class that there are a couple of method...

  33. [2021]

    https://transformer-circuits.pub/2021/framework/index.html

    2021

Pith tools

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