Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Scaling over Scaling: Exploring Test-Time Scaling Plateau in Large Reasoning Models

T0 review · 4 major / 5 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 Generating more candidate answers or more rethinking rounds improves reasoning models only up to a computable saturation point, beyond which each extra unit of test-time compute yields gains below any chosen threshold.

desk verdict A clean algebraic wrapper around the textbook geometric success formula, but the validation is circular and the predictive claim is not supported. read the letter →

arxiv 2505.20522 v2 pith:AIKGMX4H submitted 2025-05-26 cs.AI cs.CLcs.LG

Jian Wang , Boyan Zhu , Chak Tou Leong , Yongqi Li , Wenjie Li This is my paper
classification cs.AIcs.CLcs.LG
keywords test-timescalingplateaulargereasoningmodelssaturationpointself-consistencysequentialrefinementprobabilisticperformancemodelresourceallocation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether test-time scaling — generating more candidate solutions in parallel, or asking a reasoning model to rethink its answer round by round — has a measurable limit for large reasoning models that already reason internally. The authors build a probabilistic performance model in which success probability after $N$ units of computation is $F(N) = F_{\max}(1-(1-p_x)^N)$, compute the marginal gain of each additional unit, $\Delta F(N) = F_{\max} p_x (1-p_x)^N$, and solve for the budget at which that gain falls below an arbitrary threshold $\epsilon$. Their central finding is that parallel and sequential scaling, despite different mechanisms, reduce to the same saturation point $N^* = \left\lceil \frac{\ln(\epsilon/(F_{\max} p_x))}{\ln(1-p_x)} \right\rceil$. If this holds, a practitioner can stop allocating generations once $N^*$ is reached and save test-time compute without measurable loss. The paper reports that this predicted plateau correlates with empirically observed plateaus on MATH-500 (Pearson $r$ from 0.575 to 0.803) and that the same saturation pattern appears on AIME and GPQA.

What carries the argument

The carrier of the argument is the identity $F(N) = F_{\max}(1-(1-p_x)^N)$ together with its derived saturation point $N^* = \left\lceil \frac{\ln(\epsilon/(F_{\max} p_x))}{\ln(1-p_x)} \right\rceil$. The identity does the work by converting two different mechanisms — binomial sampling followed by voting in the parallel case, and geometric waiting time to an absorbing correct state in the sequential case — into one exponential-decay formula for marginal gain, $\Delta F(N) = F_{\max} p_x (1-p_x)^N$. The saturation point turns the intuitive idea of diminishing returns into a computable budget: the smallest $N$ such that one more generation buys less than $\epsilon$ of expected performance.

What would settle it

Estimate $p_x$ for each MATH-500 problem from only the first 8 or 16 of its 32 generations, compute $N^*$ from that estimate, and correlate it with the plateau observed on the remaining generations; if the rank correlation, currently 0.575–0.803 when $p_x$ uses all 32 generations, collapses toward zero, then the plateau is not predictable from independent evidence. The same test can be run with a completely fresh sample of generations per problem before comparing the predicted and observed plateau.

Watch

Extended reading notes

Core claim

The central claim is that test-time scaling over already-reasoning models is governed by one curve and one number. Viewing each of $N$ parallel candidate answers as an independent Bernoulli attempt gives success probability $1-(1-p_{\mathrm{sample}})^N$; viewing $N$ sequential rethinking rounds as a two-state Markov chain with an absorbing correct state gives $1-(1-p_{\mathrm{rethink}})^N$, the same functional form. The unified curve $F(N) = F_{\max}(1-(1-p_x)^N)$ turns the plateau into an operational quantity: the marginal gain $\Delta F(N) = F_{\max} p_x (1-p_x)^N$ decays exponentially, so the smallest $N$ at which it drops below the threshold $\epsilon$ is $N^* = \left\lceil \frac{\ln(\epsilon/(F_{\max} p_x))}{\ln(1-p_x)} \right\rceil$. The paper claims this bound predicts the empirically observed plateaus on MATH-500, with Pearson correlations between 0.575 and 0.803 across two DeepSeek-R1 distilled models and both scaling strategies, and that the same saturation shape appears on AIME 2024/2025 and GPQA.

Load-bearing premise

The load-bearing premise is that each problem's per-generation success probability, estimated in the paper's Appendix A from the very same 32 generations that later define where the plateau occurs, is a legitimate input for predicting that plateau; if that success rate came from independent data, the paper gives no evidence that its predicted stopping point would still match the observed one.

Editorial extensions

If this is right

  • Allocating more than $N^*$ generations per problem is predicted to waste compute, since each additional unit's expected gain is below $\epsilon$; the saved decoding tokens can be redirected to other problems or models.
  • Parallel scaling should be preferred when the budget allows, because the paper's experiments find it reaches higher accuracy than sequential rethinking at the same $N$ and shows the strongest plateau predictability ($r = 0.803$).
  • The plateau point makes the trade-off between inference compute and model scale explicit: a 1.5B model with test-time scaling matches or beats a 7B model without scaling, and the per-problem cost of that gain can be priced once $N^*$ is known.
  • Sequential scaling inherits the same exponential plateau shape despite its error-propagation risk, so its limit is set by the effective single-round correction probability $p_{\mathrm{rethink}}$ rather than by the strategy's mechanism.
  • Choosing the threshold $\epsilon$ converts a vague intuition about diminishing returns into an explicit policy: smaller $\epsilon$ buys extra performance per token at increasing marginal cost, and $N^*$ gives the exact budget for any chosen $\epsilon$.

Reading between the lines

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

  • The paper's own validation leaves a stricter protocol untested: estimate each problem's $p_x$ from only its first 8 or 16 generations and check whether $N^*$ still ranks the plateaus observed on later generations — the reported correlations would need to survive that split to confirm genuine predictive power.
  • An untested symmetry follows directly from the formula: $N^*$ depends on $p_x$ and $F_{\max}$ only through the ratio $\epsilon/(F_{\max} p_x)$, so a model-task pair with double the success probability and double the maximum performance should saturate at the same budget.
  • Because the curve $F(N) = F_{\max}(1-(1-p_x)^N)$ is posited rather than derived from the model's internals, a natural extension is to test whether the exponential-decay law still holds when $p_x$ varies across problems within a dataset, or when sampling and rethinking are interleaved in a hybrid strategy that the paper's two-state Markov model does not cover.
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 / 5 minor

Summary. This paper introduces TTSPM, a probabilistic model that describes test-time scaling performance as F(N)=Fmax(1-(1-p_x)^N), where p_x is a per-problem single-step success probability. From this model the authors derive a saturation point N* = ceil(ln(epsilon/(Fmax p_x))/ln(1-p_x)) and claim it applies uniformly to parallel (self-consistency) and sequential (multi-round rethinking) scaling. They report Pearson correlations between predicted and observed plateau points on MATH-500 (r = 0.575-0.803) and present scaling curves on AIME, MATH-500, and GPQA.

Significance. The algebraic derivation in Section 3 is correct, and the paper is clearly written about its assumptions. The potential practical value lies in the possibility of setting a compute budget N* that avoids diminishing returns. However, the empirical validation is undermined by in-sample estimation of p_x, an incorrect geometric MLE, and an undefined observed-plateau measure; the correlation in Figure 4 therefore does not establish predictive power for unseen generations. The 'unified' bound for parallel and sequential scaling is largely a consequence of the shared geometric functional form chosen in Eq. (8).

major comments (4)
  1. [Section 4.4 and Appendix A, Figure 4] The validation is in-sample: for each test problem, p_x is estimated from the same up-to-32 generations used to compute the observed plateau (Appendix A), so the predicted N* is a function of the observed outcome. The 80/20 split in Section 4.4 only estimates the scalar epsilon/Fmax; it does not make the per-problem p_x out-of-sample. As a result, the positive correlations in Figure 4 can arise even if the model has no predictive content for unseen generations. Please re-estimate p_x on an independent set of generations or otherwise break the circularity. Additionally, the y-axis 'Observed Actual Scaling Plateau N' is never defined in the main text or appendix, which makes the validation not reproducible.
  2. [Appendix A] The estimator p_hat(x) = k_x/32 for sequential scaling is not the MLE for a geometric distribution; the MLE is 1/k_x. With p_hat = k_x/32, the predicted N* decreases as the observed first-success round k_x increases, which is opposite to the reported positive correlations in Figure 4. If the intended estimator is 1/k_x, then N* becomes a deterministic monotone transform of the observed outcome, making the correlation in Figure 4 essentially a self-correlation. Please clarify and correct this estimator.
  3. [Section 3.2 and Table 1] Eq. (8) models the probability that at least one of N independent samples is correct, i.e., Pass@N or Hit@N. However, the headline metric in Table 1 is majority-voting accuracy (Acc.), which is a different quantity and does not follow Eq. (8) in general. The paper must state which metric the model is intended to describe and reconcile the plateau observed in Acc. with the model.
  4. [Section 3.2] The claimed unification of parallel and sequential scaling in Eq. (14) is by construction: both paradigms are assumed to follow the identical geometric performance function F(N)=Fmax(1-(1-p_x)^N). The paper does not derive this functional form from the mechanisms of sampling-with-majority-vote or iterative refinement, so the 'remarkable' convergence to a unified structure is a modeling assumption rather than an emergent finding.
minor comments (5)
  1. [Abstract] The phrase 'thescaling plateau' should be 'the scaling plateau'.
  2. [Section 4.2] The sentence 'and percentage points on GPQA' has a missing value; based on Table 1 it should read 'and 18.1 percentage points on GPQA'.
  3. [Appendix A] The value '1e−5' should be typeset as $10^{-5}$ for clarity.
  4. [Section 4.1] The footnote marker for AIME 2025 is placed awkwardly, and the AIME 2025 HuggingFace dataset is cited only via a footnote URL rather than being included in the reference list.
  5. [Figures 2 and 3] The legends in Figures 2 and 3 are difficult to read; the captions should explicitly state which color/line style corresponds to which scaling strategy and model.

Circularity Check

1 steps flagged · score 7.0 of 10

The Figure 4 'prediction' is in-sample: per-problem p_x is estimated from the same 32 generations that define the observed plateau, so the claimed correlation partly reduces to a fit.

  1. fitted input called prediction [Appendix A (Experimental Details); Section 4.4 (Verification of the Scaling Plateau), Fig. 4]
    "We estimate this probability for each problem in both validation and test sets using experimental data from up to N= 32 generations per problem. ... For Sequential Scaling: ... For a problem first solved at round k x ≤ 32, we use the maximum likelihood estimate for a geometric distribution: ˆprethink(x) = kx/32 if problem x is first solved at round kx ≤ 32; 1e−5 if problem x is not solved within 32 rounds."

    The predicted plateau in Fig. 4 is computed by Eq. (14) with p_x estimated on the very same test problems and from the very same up-to-32 generations that produce the 'observed actual scaling plateau' (for sequential scaling, the first-success round k_x). The 8:2 split in Sec. 4.4 calibrates only the scalar epsilon/Fmax; it does not hold out p_x. Since App. A sets p_hat_rethink = k_x/32, the predicted N* is a function of the observed first-success round, so the Pearson r in Fig. 4 measures agreement between a transform of an outcome and that outcome, not an out-of-sample prediction. The paper itself concedes p_x 'might not be constant across all problem instances' and that its estimation is 'nuanced,' but the reported validation never escapes the in-sample dependence.

full rationale

The algebraic derivation of N* (Eqs. 8-14) is self-contained once F(N)=Fmax(1-(1-p)^N) is assumed; assuming a saturation curve is a modeling choice, not circularity. The circularity enters at the empirical verification: Section 4.4 presents Fig. 4 as evidence that Eq. (14) predicts the observed plateau, but Appendix A estimates the problem-specific p_x from the same 32 generations that define the observed plateau. The 8:2 split only calibrates epsilon/Fmax, not p_x, so for each test problem the predictor and the predicted quantity are computed from the same data. In the sequential case the estimator p_hat=k_x/32 is a deterministic transform of the first-success round k_x that serves as the observed plateau, making the correlation partly tautological. Because the functional-form assumption and scalar calibration are not themselves in-sample, the paper is not wholly circular; however, its central predictive claim is not independently established. Score 7.

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

TTSPM rests on a posited exponential saturation curve with parameters p_x (fitted per problem from the test-set generations), Fmax (fitted as part of the epsilon/Fmax ratio on a validation split), and epsilon (a user threshold that is also estimated during validation). The axioms include independence of scaling units and a constant transition probability in sequential scaling, both flagged as limitations. The model is not derived from LRM mechanisms, so the bound N* is an algebraic consequence of an assumed functional form.

free parameters (3)
  • p_x (per-problem effective success probability) = parallel: #correct/32; sequential: k/32 or 1e-5
    Fitted from the 32 generations per problem in the test set; the central bound N* depends directly on p_x.
  • Fmax (maximum achievable performance) = not directly reported; estimated via epsilon/Fmax on validation set
    Fmax is described as hard to determine; in the validation procedure the ratio epsilon/Fmax is fitted on the validation set, making the bound dependent on a fitted ratio.
  • epsilon (threshold for negligible gain) = not reported; ratio epsilon/Fmax estimated on validation
    A user-defined hyperparameter in principle, but in Section 4.4 the unknown entry epsilon/Fmax is estimated from the validation set, so it becomes a fitted parameter.
assumptions (4)
  • domain assumption Success events for each scaling unit are conditionally independent given the problem and model
    Explicitly flagged as a limitation in Section 4.5; without this the binomial/geometric form of F(N) does not follow.
  • domain assumption Sequential scaling is a two-state Markov process with an absorbing correct state and constant transition probability
    Section 3.1 posits this to justify the geometric distribution; real rethink chains can have memory and non-constant correction probabilities.
  • ad hoc to paper The saturating performance model F(N)=Fmax(1-(1-p)^N) holds
    Section 3.1 introduces this as 'positing' the form; it is not derived from LRM mechanisms, so the plateau bound inherits this assumption.
  • standard math For the logarithmic derivation, 0 < eps/(Fmax*p) < 1
    Need to take logs and reverse the inequality; cases outside this range are handled by assigning N*=1, which limits the applicability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Scaling over Scaling: Exploring Test-Time Scaling Plateau in Large Reasoning Models." pith.science (2026). https://pith.science/paper/AIKGMX4H

@misc{pith2026250520522,
  author       = {Pith},
  title        = {Pith review of: Scaling over Scaling: Exploring Test-Time Scaling Plateau in Large Reasoning Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AIKGMX4H}},
  note         = {Machine review of arXiv:2505.20522}
}
read the original abstract

Large reasoning models (LRMs) have exhibited the capacity of enhancing reasoning performance via internal test-time scaling. Building upon this, a promising direction is to further scale test-time compute to unlock even greater reasoning capabilities. However, as we push these scaling boundaries, systematically understanding the practical limits and achieving optimal resource allocation becomes a critical challenge. In this paper, we investigate the scaling plateau of test-time scaling and introduce the Test-Time Scaling Performance Model (TTSPM). We theoretically analyze two fundamental paradigms for such extended scaling, parallel scaling and sequential scaling, from a probabilistic modeling perspective. Our primary contribution is the derivation of the saturation point on the scaling budget for both strategies, identifying thresholds beyond which additional computation yields diminishing returns. Remarkably, despite their distinct mechanisms, both paradigms converge to a unified mathematical structure in their upper bounds. We empirically validate our theoretical findings on challenging reasoning benchmarks, including AIME, MATH-500, and GPQA, demonstrating the practical utility of these bounds for test-time resource allocation. We hope that this work provides insights into the cost-benefit trade-offs of test-time scaling, guiding the development of more resource-efficient inference strategies for large reasoning models.

Figures

Figures reproduced from arXiv: 2505.20522 by the authors.

Figure 1
Figure 1. Conceptual illustration of parallel scaling [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Scaling curves across three benchmarks (including AIME 2025, MATH-500, and GPQA), [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Statistics of the number of generations per problem ( [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Correlation between the theoretically predicted and empirically observed scaling plateau [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Accuracy scaling curves across four benchmarks (a) AIME 2024, (b) AIME 2025, (c) [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

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. AdaTurn: Budget-Aware Test-Time Scaling for Active Visual Perception Agents

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Budget-conditioned forced-answer RL lifts 4-turn VisualProbe-Medium from 36.7% to 47.6% while keeping 32-turn performance competitive.

Reference graph

Works this paper leans on

32 extracted references · 8 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]

    Large language monkeys: Scaling inference compute with repeated sampling.arXiv preprint arXiv:2407.21787, 2024

    Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V Le, Christopher Ré, and Azalia Mirhoseini. Large language monkeys: Scaling inference compute with repeated sampling.arXiv preprint arXiv:2407.21787, 2024

    arXiv 2024

  3. [3]

    Sets: Leveraging self-verification and self-correction for improved test-time scaling.arXiv preprint arXiv:2501.19306, 2025

    Jiefeng Chen, Jie Ren, Xinyun Chen, Chengrun Yang, Ruoxi Sun, and Sercan Ö Arık. Sets: Leveraging self-verification and self-correction for improved test-time scaling.arXiv preprint arXiv:2501.19306, 2025

    arXiv 2025

  4. [4]

    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

  5. [5]

    Measuring mathematical problem solving with the math dataset

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. InThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021

    work page 2021

  6. [6]

    Two heads are better than one: Test-time scaling of multi-agent collaborative reasoning.arXiv preprint arXiv:2504.09772, 2025

    Can Jin, Hongwu Peng, Qixin Zhang, Yujin Tang, Dimitris N Metaxas, and Tong Che. Two heads are better than one: Test-time scaling of multi-agent collaborative reasoning.arXiv preprint arXiv:2504.09772, 2025

    arXiv 2025

  7. [7]

    Reasoning-as-logic-units: Scaling test-time reasoning in large language models through logic unit alignment.arXiv preprint arXiv:2502.07803, 2025

    Cheryl Li, Tianyuan Xu, and Yiwen Guo. Reasoning-as-logic-units: Scaling test-time reasoning in large language models through logic unit alignment.arXiv preprint arXiv:2502.07803, 2025

    arXiv 2025

  8. [8]

    Dancing with critiques: Enhancing llm reasoning with stepwise natural language self-critique.arXiv preprint arXiv:2503.17363, 2025

    Yansi Li, Jiahao Xu, Tian Liang, Xingyu Chen, Zhiwei He, Qiuzhi Liu, Rui Wang, Zhuosheng Zhang, Zhaopeng Tu, Haitao Mi, et al. Dancing with critiques: Enhancing llm reasoning with stepwise natural language self-critique.arXiv preprint arXiv:2503.17363, 2025

    arXiv 2025

Show all 32 references
  1. [9]

    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. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  2. [10]

    Metascale: Test-time scaling with evolving meta-thoughts.arXiv preprint arXiv:2503.13447, 2025

    Qin Liu, Wenxuan Zhou, Nan Xu, James Y Huang, Fei Wang, Sheng Zhang, Hoifung Poon, and Muhao Chen. Metascale: Test-time scaling with evolving meta-thoughts.arXiv preprint arXiv:2503.13447, 2025

    2025 arXiv

  3. [11]

    American invitational mathematics examination - aime

    MAA. American invitational mathematics examination - aime. https://huggingface.co/datasets/ AI-MO/aimo-validation-aime, February 2024. 10

    2024

  4. [12]

    Self-refine: Iterative refinement with self-feedback

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. InAdvances in Neural Information Processing Systems, volume 36, pages 4653...

    2023

  5. [13]

    Deepseek-r1 thoughtology: Let’s< think> about llm reasoning.arXiv preprint arXiv:2504.07128, 2025

    Sara Vera Marjanovi´c, Arkil Patel, Vaibhav Adlakha, Milad Aghajohari, Parishad BehnamGhader, Mehar Bhatia, Aditi Khandelwal, Austin Kraft, Benno Krojer, Xing Han Lù, et al. Deepseek-r1 thoughtology: Let’s< think> about llm reasoning.arXiv preprint arXiv:2504.07128, 2025

    2025

  6. [14]

    a1: Steep test-time scaling law via environment augmented generation.arXiv preprint arXiv:2504.14597, 2025

    Lingrui Mei, Shenghua Liu, Yiwei Wang, Baolong Bi, Yuyao Ge, Jun Wan, Yurong Wu, and Xueqi Cheng. a1: Steep test-time scaling law via environment augmented generation.arXiv preprint arXiv:2504.14597, 2025

    2025 arXiv

  7. [15]

    Introducing OpenAI o1.https://openai.com/o1/, September 2024

    OpenAI. Introducing OpenAI o1.https://openai.com/o1/, September 2024

    2024

  8. [16]

    OpenAI o3-mini system card

    OpenAI. OpenAI o3-mini system card. https://cdn.openai.com/o3-mini-system-card-feb10. pdf, January 2025

    2025

  9. [17]

    Thinking slow, fast: Scaling inference compute with distilled reasoners

    Daniele Paliotta, Junxiong Wang, Matteo Pagliardini, Kevin Y Li, Aviv Bick, J Zico Kolter, Albert Gu, François Fleuret, and Tri Dao. Thinking slow, fast: Scaling inference compute with distilled reasoners. arXiv preprint arXiv:2502.20339, 2025

    2025 arXiv

  10. [18]

    Learning adaptive parallel reasoning with language models.arXiv preprint arXiv:2504.15466, 2025

    Jiayi Pan, Xiuyu Li, Long Lian, Charlie Snell, Yifei Zhou, Adam Yala, Trevor Darrell, Kurt Keutzer, and Alane Suhr. Learning adaptive parallel reasoning with language models.arXiv preprint arXiv:2504.15466, 2025

    2025 arXiv

  11. [19]

    Optimizing test-time compute via meta reinforcement fine-tuning.arXiv preprint arXiv:2503.07572, 2025

    Yuxiao Qu, Matthew YR Yang, Amrith Setlur, Lewis Tunstall, Edward Emanuel Beeching, Ruslan Salakhutdinov, and Aviral Kumar. Optimizing test-time compute via meta reinforcement fine-tuning.arXiv preprint arXiv:2503.07572, 2025

    2025 arXiv

  12. [20]

    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. InFirst Conference on Language Modeling, 2024

    2024

  13. [21]

    Scaling test-time compute without verification or rl is suboptimal.arXiv preprint arXiv:2502.12118, 2025

    Amrith Setlur, Nived Rajaraman, Sergey Levine, and Aviral Kumar. Scaling test-time compute without verification or rl is suboptimal.arXiv preprint arXiv:2502.12118, 2025

    2025 arXiv

  14. [22]

    Heimdall: test-time scaling on the generative verification.arXiv preprint arXiv:2504.10337, 2025

    Wenlei Shi and Xing Jin. Heimdall: test-time scaling on the generative verification.arXiv preprint arXiv:2504.10337, 2025

    2025 arXiv

  15. [23]

    Adaptive rectification sampling for test-time compute scaling.arXiv preprint arXiv:2504.01317, 2025

    Zhendong Tan, Xingjun Zhang, Chaoyi Hu, Yancheng Pan, and Shaoxun Wang. Adaptive rectification sampling for test-time compute scaling.arXiv preprint arXiv:2504.01317, 2025

    2025

  16. [24]

    Think twice: Enhancing llm reasoning by scaling multi-round test-time thinking.arXiv preprint arXiv:2503.19855, 2025

    Xiaoyu Tian, Sitong Zhao, Haotian Wang, Shuaiting Chen, Yunjie Ji, Yiping Peng, Han Zhao, and Xiangang Li. Think twice: Enhancing llm reasoning by scaling multi-round test-time thinking.arXiv preprint arXiv:2503.19855, 2025

    2025 arXiv

  17. [25]

    Math-shepherd: Verify and reinforce llms step-by-step without human annotations

    Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. Math-shepherd: Verify and reinforce llms step-by-step without human annotations. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volu...

    2024

  18. [26]

    Self-consistency improves chain of thought reasoning in language models

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  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. InAdvances in Neural Information Processing Systems, volume 35, pages 24824–24837, 2022

    2022

  20. [28]

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

    Yangzhen Wu, Zhiqing Sun, Shanda Li, Sean Welleck, and Yiming Yang. Inference scaling laws: An empirical analysis of compute-optimal inference for problem-solving with language models. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  21. [29]

    Inftythink: Breaking the length limits of long-context reasoning in large language models.arXiv preprint arXiv:2503.06692, 2025

    Yuchen Yan, Yongliang Shen, Yang Liu, Jin Jiang, Mengdi Zhang, Jian Shao, and Yueting Zhuang. Inftythink: Breaking the length limits of long-context reasoning in large language models.arXiv preprint arXiv:2503.06692, 2025

    2025

  22. [30]

    Tree of thoughts: Deliberate problem solving with large language models

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. InAdvances in Neural Information Processing Systems, volume 36, pages 11809–11822, 2023. 11

    2023

  23. [31]

    Limo: Less is more for reasoning.arXiv preprint arXiv:2502.03387, 2025

    Yixin Ye, Zhen Huang, Yang Xiao, Ethan Chern, Shijie Xia, and Pengfei Liu. Limo: Less is more for reasoning.arXiv preprint arXiv:2502.03387, 2025

    2025 arXiv

  24. [32]

    What, how, where, and how well? a survey on test-time scaling in large language models.arXiv preprint arXiv:2503.24235, 2025

    Qiyuan Zhang, Fuyuan Lyu, Zexu Sun, Lei Wang, Weixu Zhang, Zhihan Guo, Yufei Wang, Irwin King, Xue Liu, and Chen Ma. What, how, where, and how well? a survey on test-time scaling in large language models.arXiv preprint arXiv:2503.24235, 2025. 12 A Experimental Details Our theo...

    2025 arXiv

Pith tools

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