Pith. sign in

REVIEW 3 major objections 6 minor 5 cited by

Steering LLM Thinking with Budget Guidance

T0 review · 3 major / 6 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 Budget guidance multiplies a frozen LLM's token probabilities by a lightweight predictor's remaining-length scores, making reasoning length a controllable inference-time budget that beats hard cutoffs at matched token counts.

desk verdict Genuinely novel inference-time budget control with credible empirical gains, but the calibration of the length predictor on guided prefixes is an unexamined load-bearing assumption. read the letter →

arxiv 2506.13752 v1 pith:WTG7LWHI submitted 2025-06-16 cs.CL cs.AI

Junyan Li , Wenshuo Zhao , Yang Zhang , Chuang Gan This is my paper
classification cs.CLcs.AI
keywords budgetguidancethinkingreasoninglengthcontrolclassifiertest-timeinferencetokenefficiencyGammadistributionchain-of-thought
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

Deep-thinking LLMs often reason far longer than a task requires, wasting tokens. This paper claims that a frozen LLM's reasoning length can be steered toward a target budget by multiplying its next-token probabilities pointwise with scores from a lightweight auxiliary predictor, with no fine-tuning of the LLM. The predictor models, for each candidate next token, a Gamma distribution over the remaining thinking length, and this budget-conditional reweighting makes the model naturally wind down its reasoning to fit the budget. The claim is that this soft steering beats hard early-termination baselines at matched token counts (e.g., 88.2% vs 86.0% on MATH-500 for a 7B model) and can save 37% of thinking tokens with minimal accuracy loss. The paper also reports that the trained predictor transfers across benchmark domains and learns to estimate question difficulty.

What carries the argument

The machinery is the multiplicative modulation $c_t = \mathrm{normalize}(u_t \circ a_t)$: each candidate token's probability in the frozen LLM is scaled by the predictor's Gamma-CDF score $a_t$, so tokens that are likely to fit the remaining budget get up-weighted and the renormalized product becomes the budget-conditional sampling distribution. The predictor parameterizes each token's remaining-length distribution as a Gamma over $\log L_t$, so only two vectors (shape $\lambda_t$ and rate $\alpha_t$) are needed per step instead of one full distribution per vocabulary item. To keep overhead low, modulation is applied only at the start of each reasoning paragraph (newline delimiters), which the authors measure as a 0.6% total latency increase for a 7B model. The predictor is trained once on reasoning traces from the unguided model (OpenR1-Math-220k) and then used at inference time with the LLM frozen.

What would settle it

On a fixed benchmark with a fixed budget, collect the guided traces and, at each modulated step, compare the predictor's claimed cumulative probability that generation ends within budget with the empirical fraction of traces that actually end within budget at that point. If the predictor is calibrated on the guided data, the Eq. (4) identity is doing what it claims; if the calibration is systematically off, the method is not sampling from the true budget-conditional distribution, and a reader should look for another explanation of the accuracy gains.

Watch

Extended reading notes

Core claim

The paper's central claim is expressed in Eq. (4): the budget-conditional distribution over the next token is obtained by pointwise multiplication and renormalization, $c_t = \mathrm{normalize}(u_t \circ a_t)$, where $u_t$ is the frozen LLM's unconditional next-token probability vector and $a_t$ is a vector of cumulative probabilities, one per candidate token, that the remaining thinking length $L_t$ stays within the budget $\bar{l}-t$. This identity follows from Bayes' rule, and the paper's substantive contribution is to make $a_t$ computable: a BERT-scale predictor, fed the LLM's hidden states, outputs shape and rate parameters of a Gamma distribution over $\log L_t$ for every candidate token, trained by maximum likelihood on reasoning traces from the unguided model. Because generation is modulated softly at paragraph starts rather than cut off, the model's reasoning style adjusts to the budget, producing concise traces under tight budgets and reflective traces under loose ones, which the paper argues is why it outperforms the abrupt-stop baseline budget forcing at matched average token counts across all tested models and benchmarks.

Load-bearing premise

The predictor is trained on reasoning traces from the unguided model, but the guidance changes what the model generates next; the paper assumes the predicted remaining-length probabilities stay valid under the guided policy, and it never retrains the predictor on guided traces or measures that distribution shift.

Editorial extensions

If this is right

  • Reasoning length becomes a tunable inference-time knob for deep-thinking LLMs without retraining, so latency and cost can be set per request rather than baked in at training time.
  • At matched average token counts, budget guidance is claimed to beat budget forcing across three models and four math benchmarks, with the largest gaps under the tightest budgets, up to 26% absolute accuracy gain on MATH-500.
  • A predictor trained only on math traces transfers to scientific, logical, tabular, and code reasoning benchmarks, implying budget control does not need per-domain training data.
  • Because the predictor's length estimates rise with problem difficulty and with prompt encouragement to think long, budget guidance also yields a per-question difficulty signal at generation time.

Reading between the lines

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

  • Because the predictor is trained only on unguided traces but deployed under a guided distribution, retraining or recalibrating it on guided traces is a direct testable extension: if that retraining improves the accuracy-token tradeoff, the current method is leaving performance on the table; if it does not, the distribution-shift concern is moot.
  • The pointwise multiplication identity is not specific to reasoning length; a predictor that estimates any token-conditional scalar property such as verbosity, confidence, or time-to-answer could steer generation the same way, a generalization the paper does not state.
  • The paper's own reported difficulty-correlation results suggest a routing scheme it does not propose: run easy questions under tight budgets and hard questions under loose ones, using the predictor's first-token length estimate as a difficulty gate.
  • The paper applies modulation only at paragraph starts and reports a 0.6% latency increase, but it does not report what full-step modulation would cost in latency or gain in steering, leaving the paragraph-start choice an unquantified trade of control for speed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes budget guidance, a test-time method for controlling the reasoning length of deep-thinking LLMs without fine-tuning the LLM. A lightweight predictor, trained on unguided reasoning traces, models a Gamma distribution over the remaining thinking length for each candidate token; at inference its CDF is multiplied pointwise with the LLM's next-token probabilities (Eq. 4) to steer generation toward a target budget. Experiments on MATH-500, AIME-2024, AMC, OlympiadBench, and four out-of-domain benchmarks across DeepSeek-R1-Distill-Qwen-7B/32B and Qwen3-8B report consistent accuracy gains over budget forcing at roughly similar token counts, plus cross-domain transfer and an emergent difficulty-estimation behavior. The central theoretical claim is that Eq. (1) makes the guided distribution exactly the budget-conditional distribution of the base model.

Significance. If the mechanism holds, the method is a practical, fine-tuning-free alternative to budget forcing, with a lightweight predictor, negligible added latency, and encouraging generalization across models and domains. The paper ships its source code and presents a clear framework that others can build on. The empirical trend is consistent across three models and several benchmarks, and the difficulty-estimation analysis in Section 4.3 is a useful probe of what the predictor learns. However, the theoretical guarantee rests on a distribution-shift assumption that is not tested, and some headline numbers in the abstract are not traceable to the reported tables. The central idea is interesting, but the current evidence supports it as a promising heuristic rather than as a verified instantiation of Bayes-conditional generation.

major comments (3)
  1. [§3.1, Eq. (1); §3.3, Eq. (6); §4.2] The Bayes derivation requires Pr(L_t ≤ l̄ - t | X, Y_<t, Y_t) to be the true conditional under the base model's distribution. The predictor is trained only on unguided traces (Eq. 6; OpenR1-Math-220k traces generated by DeepSeek R1), but at inference it is applied to prefixes Y_<t produced by the guided policy. The manuscript does not measure predictor calibration on guided prefixes, does not retrain or adapt the predictor on guided traces, and does not quantify how much guidance changes the input distribution. Without such evidence, Eq. (4) may not implement the claimed budget-conditional distribution, and the observed gains could be explained by the predictor's implicit difficulty estimation (Section 4.3). Please add a calibration analysis on guided prefixes or an ablation in which the predictor is retrained on guided traces, and report the effective distribution shift.
  2. [Abstract; Table 1; Figure 1] The headline claims "up to a 26% accuracy gain" and "maintaining competitive accuracy with only 63% of the thinking tokens" are not traceable to any table or figure. For example, Table 1 MATH-500 R1-7B reports Budget Guidance 88.2 vs. Budget Forcing 86.0, which is a relative gain of about 2.6%, and a token ratio of 1329/2598 ≈ 51%. Figure 1's "26% higher" appears to refer to a single illustrative example, not to a systematic result. Please state exactly which experimental setting yields the 26% and 63% numbers and make the abstract consistent with the reported tables.
  3. [§4.2.1, Table 1, Figure 3] The text says "ensure the average thinking length comparable between our method and the baseline," but the token counts in Table 1 are not exactly matched (e.g., MATH-500 R1-7B: 1547 vs. 1329; AIME-2024 R1-32B: 2936 vs. 2873). Since the central claim is better accuracy at the same token budget, the comparison should be made at exactly matched token counts, or the Figure 3 trade-off curves should be used with a clearly defined matching procedure and with confidence intervals. Otherwise the reported accuracy differences may be confounded by unequal budgets.
minor comments (6)
  1. [§3.2 vs. §3.4] The description in Section 3.2 says the predictor must "traverse all possible values for Y_t" and predict the remaining length for each, but the architecture in Section 3.4 takes only hidden states of the last generated token and outputs an n×2 matrix; it does not appear to use the candidate token's identity as input. Please clarify how the predictor conditions on Y_t = v_i and why the n×2 output is the right parameterization.
  2. [Eq. (5)] The notation Gamma(log(L_t); λ, α) is unusual: if log(L_t) has a Gamma distribution, then L_t is log-Gamma distributed. Also, L_t = 0 for the end-of-thinking token is outside the support, but at inference the predictor must score that token. Please explain how the end-of-thinking token is handled in the CDF computation.
  3. [Table 1] Several entries in Table 1 are run together and hard to parse (e.g., the AIME-2024 row for R1-7B Budget Forcing appears as "16.7201555.4" instead of separate accuracy and token columns). Please reformat the table.
  4. [Section 4.3] The t-test for the prompt-type comparison should state the test type and sample size, and the main accuracy results would benefit from confidence intervals or multiple-seed runs, since several differences (e.g., FOLIO 61.6 vs. 60.1) are small.
  5. [Appendix D] The data augmentation that creates samples of the form <think>ANSWER_MESSAGE</think>ANSWER_MESSAGE treats the final answer as if it were a thinking trace. This may produce unnatural training targets; please justify why this does not bias the predictor's length estimates.
  6. [Section 3.5, Eq. (7)] Modulation skipping means Eq. (4) is applied only at paragraph starts, so the final generation distribution is not the budget-conditional distribution at every token. The paper should state explicitly that this breaks the exact Bayes interpretation and explain why the approximation is expected to preserve the desired behavior.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the budget-conditional derivation rests on an exact Bayes identity, the predictor is trained on separate unguided traces with real length labels, and evaluation accuracy is an external target.

full rationale

Budget guidance's derivation chain is self-contained, with no step that reduces by construction to its own inputs. Equation (1) is an exact Bayes identity for the base LLM's own token distribution: the budget-conditional distribution is proportional to the unconditional token probabilities times the true conditional CDF of remaining length. The paper does not define the budget-conditional distribution via Eq. (4) as a premise; Eq. (4) is a consequence of Eq. (1) together with a learned approximation to that CDF. The predictor is trained with a standard maximum-likelihood objective (Eq. 6) on OpenR1-Math-220k traces using actual remaining-length labels, not on the evaluation accuracies or on the evaluation sets; MATH-500, AIME, AMC, and OlympiadBench accuracies are external metrics never used as a training signal. The only likely author-overlap citation is [14] in Appendix A, and it is used merely for a keyword-frequency analysis methodology, not to justify the central claim. The distribution-shift concern noted by the skeptical reader is a real correctness risk: the predictor is trained on unconditional traces but applied to guided prefixes, so its CDF estimates may be miscalibrated and Eq. (4) may not exactly implement the budget-conditional distribution. That is an unverified approximation in the argument, not a circularity, because the conclusion does not follow from the training objective by definition and the evaluation target is external to the fitted model. No fitted parameter is renamed as a prediction, and no load-bearing self-citation chain was found.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The method rests on a parametric modeling choice and a distribution-shift assumption that are not validated. No additional physical or mathematical entities are introduced. The predictor itself is a large learned component.

free parameters (1)
  • Predictor output parameters (λ, α) for every vocabulary token = Learned on OpenR1-Math-220k; values not reported
    These define the predicted Gamma distribution over log remaining length used in Eq. (5). The method's behavior and all accuracy results depend on this learned mapping, which is a large fitted component rather than a derived constant.
assumptions (4)
  • domain assumption Remaining thinking length L_t is a random variable whose conditional distribution can be estimated from the LLM's hidden states and the next token alone.
    The BERT predictor receives only the last-token hidden states; it never conditions directly on the candidate token as an input. This assumes sufficient information is available or learnable in the output layer.
  • domain assumption A Gamma distribution over log(L_t) adequately approximates the conditional distribution of remaining length.
    Eq. (5) imposes this parametric form; no goodness-of-fit or calibration analysis is provided.
  • ad hoc to paper The predictor trained on unguided traces remains valid when the generation policy changes due to guidance.
    Eq. (1) uses the Bayes factor under the true generative process, but training data come from the original model, and guided traces are never used to retrain the predictor. This is the main unexamined distribution-shift assumption.
  • domain assumption OpenR1-Math-220k is representative and non-overlapping with the evaluation benchmarks.
    The paper states the training task is different from inference tasks, but does not quantify overlap with MATH-500 or other eval sets. Some overlap could inflate the length predictor's accuracy on in-domain benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Steering LLM Thinking with Budget Guidance." pith.science (2026). https://pith.science/paper/WTG7LWHI

@misc{pith2026250613752,
  author       = {Pith},
  title        = {Pith review of: Steering LLM Thinking with Budget Guidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WTG7LWHI}},
  note         = {Machine review of arXiv:2506.13752}
}
read the original abstract

Recent deep-thinking large language models often reason extensively to improve performance, but such lengthy reasoning is not always desirable, as it incurs excessive inference costs with disproportionate performance gains. Controlling reasoning length without sacrificing performance is therefore important, but remains challenging, especially under tight thinking budgets. We propose budget guidance, a simple yet effective method for steering the reasoning process of LLMs toward a target budget without requiring any LLM fine-tuning. Our approach introduces a lightweight predictor that models a Gamma distribution over the remaining thinking length during next-token generation. This signal is then used to guide generation in a soft, token-level manner, ensuring that the overall reasoning trace adheres to the specified thinking budget. Budget guidance enables natural control of the thinking length, along with significant token efficiency improvements over baseline methods on challenging math benchmarks. For instance, it achieves up to a 26% accuracy gain on the MATH-500 benchmark under tight budgets compared to baseline methods, while maintaining competitive accuracy with only 63% of the thinking tokens used by the full-thinking model. Budget guidance also generalizes to broader task domains and exhibits emergent capabilities, such as estimating question difficulty. The source code is available at: https://github.com/UMass-Embodied-AGI/BudgetGuidance.

Figures

Figures reproduced from arXiv: 2506.13752 by the authors.

Figure 1
Figure 1. Deep-thinking models often produce excessively long reasoning traces, leading to high [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An overview of budget guidance. A lightweight predictor uses the LLM’s hidden states to predict a Gamma distribution over the remaining reasoning length for each candidate token. We then use the CDF of Gamma distribution to compute a predictor score, which is combined with the LLM’s output score to guide generation. The result is soft, token-level steering toward budget-conditioned reasoning without any LLM fine-tun… view at source ↗
Figure 3
Figure 3. Accuracy vs. thinking length on math benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Thinking length controllability measured on MATH-500 benchmark. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Correlation between question difficulties and [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: Sample reasoning traces generated with budget guidance under different thinking budgets. the cross-domain generalizability of our approach and its potential applicability to a wide range of reasoning scenarios. While the gains on out-of-domain tasks are less pronounced…
Figure 8
Figure 8. Figure 8: Reasoning keywords frequency comparison under different budget settings. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Budget forcing and budget guidance generate a similar number of thinking tokens under a [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Another example demonstrating that budget guidance can solve a question correctly using [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Another example illustrating how budget guidance effectively steers LLM reasoning. [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

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

  1. OS-Pruner: Pruning Chains-of-Thought of Reasoning Models via Optimal Stopping

    cs.AI 2026-07 conditional novelty 6.0 of 10

    A lightweight optimal-stopping policy on frozen reasoning LLMs cuts CoT length 20–60% with minimal accuracy loss by trading answer correctness against token cost via a tunable λ.

  2. Are Large Reasoning Models Interruptible?

    cs.CL 2025-10 conditional novelty 6.0 of 10

    Interrupting large reasoning models or changing the problem mid-thought sharply degrades accuracy, with up to 60% drops and distinct failure modes: reasoning leakage, panic, and self-doubt.

  3. Efficient Reasoning on the Edge

    cs.LG 2026-03 accept novelty 5.5 of 10

    LoRA adapters, budget-forced GRPO, dynamic switching, parallel verification and FPTQuant enable practical chain-of-thought reasoning on quantized Qwen2.5-7B for edge devices.

  4. Base Models Know How to Reason, Thinking Models Learn When

    cs.AI 2025-10 reject novelty 5.0 of 10

    A hybrid model that steers base LLM activations with vectors fitted to thinking-model traces recovers part of the performance gap, but the conclusion that thinking models only learn 'when' is not established by the ex...

  5. 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

32 extracted references · 12 canonical work pages · cited by 5 Pith papers

  1. [1]

    Aime problems and solutions, n.d

    Art of Problem Solving. Aime problems and solutions, n.d

    work page

  2. [2]

    Amc 12 problems and solutions, n.d

    Art of Problem Solving. Amc 12 problems and solutions, n.d

    work page

  3. [3]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technolo- gies, volume 1 (long and short papers), pages 4171–4186, 2019

    2019

  4. [4]

    Dhariwal and A

    P. Dhariwal and A. Nichol. Diffusion models beat gans on image synthesis.Advances in neural information processing systems, 34:8780–8794, 2021

    2021

  5. [5]

    Dugas, Y

    C. Dugas, Y . Bengio, F. Bélisle, C. Nadeau, and R. Garcia. Incorporating second-order functional knowledge for better option pricing.Advances in neural information processing systems, 13, 2000

    work page 2000

  6. [6]

    H. Face. Open r1: A fully open reproduction of deepseek-r1, January 2025

    work page 2025

  7. [7]

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    arXiv 2025

  8. [8]

    S. Han, H. Schoelkopf, Y . Zhao, Z. Qi, M. Riddell, W. Zhou, J. Coady, D. Peng, Y . Qiao, L. Benson, et al. Folio: Natural language reasoning with first-order logic.arXiv preprint arXiv:2209.00840, 2022

    arXiv 2022

Show all 32 references
  1. [9]

    T. Han, Z. Wang, C. Fang, S. Zhao, S. Ma, and Z. Chen. Token-budget-aware llm reasoning. arXiv preprint arXiv:2412.18547, 2024

    2024 arXiv

  2. [10]

    S. Hao, S. Sukhbaatar, D. Su, X. Li, Z. Hu, J. Weston, and Y . Tian. Training large language models to reason in a continuous latent space.arXiv preprint arXiv:2412.06769, 2024

    2024 arXiv

  3. [11]

    C. He, R. Luo, Y . Bai, S. Hu, Z. L. Thai, J. Shen, J. Hu, X. Han, Y . Huang, Y . Zhang, et al. Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems.arXiv preprint arXiv:2402.14008, 2024

    2024 arXiv

  4. [12]

    Hendrycks, C

    D. Hendrycks, C. Burns, S. Kadavath, A. Arora, S. Basart, E. Tang, D. Song, and J. Stein- hardt. Measuring mathematical problem solving with the math dataset.arXiv preprint arXiv:2103.03874, 2021

    2021 arXiv

  5. [13]

    Ho and T

    J. Ho and T. Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022. 10

    2022 arXiv

  6. [14]

    B. Hou, Y . Zhang, J. Ji, Y . Liu, K. Qian, J. Andreas, and S. Chang. Thinkprune: Pruning long chain-of-thought of llms via reinforcement learning.arXiv preprint arXiv:2504.01296, 2025

    2025 arXiv

  7. [15]

    Jaech, A

    A. Jaech, A. Kalai, A. Lerer, A. Richardson, A. El-Kishky, A. Low, A. Helyar, A. Madry, A. Beutel, A. Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    2024 arXiv

  8. [16]

    N. Jain, K. Han, A. Gu, W.-D. Li, F. Yan, T. Zhang, S. Wang, A. Solar-Lezama, K. Sen, and I. Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code.arXiv preprint arXiv:2403.07974, 2024

    2024 arXiv

  9. [17]

    J. Li, E. Beeching, L. Tunstall, B. Lipkin, R. Soletskyi, S. Huang, K. Rasul, L. Yu, A. Q. Jiang, Z. Shen, et al. Numinamath: The largest public dataset in ai4maths with 860k pairs of competition math problems and solutions.Hugging Face repository, 13:9, 2024

    2024

  10. [18]

    Lightman, V

    H. Lightman, V . Kosaraju, Y . Burda, H. Edwards, B. Baker, T. Lee, J. Leike, J. Schulman, I. Sutskever, and K. Cobbe. Let’s verify step by step. InThe Twelfth International Conference on Learning Representations, 2023

    2023

  11. [19]

    W. Ma, J. He, C. Snell, T. Griggs, S. Min, and M. Zaharia. Reasoning models can be effective without thinking.arXiv preprint arXiv:2504.09858, 2025

    2025 arXiv

  12. [20]

    Muennighoff, Z

    N. Muennighoff, Z. Yang, W. Shi, X. L. Li, L. Fei-Fei, H. Hajishirzi, L. Zettlemoyer, P. Liang, E. Candès, and T. Hashimoto. s1: Simple test-time scaling.arXiv preprint arXiv:2501.19393, 2025

    2025 arXiv

  13. [21]

    X. Qi, Y . Zeng, T. Xie, P.-Y . Chen, R. Jia, P. Mittal, and P. Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to!arXiv preprint arXiv:2310.03693, 2023

    2023 arXiv

  14. [22]

    D. Rein, B. L. Hou, A. C. Stickland, J. Petty, R. Y . Pang, J. Dirani, J. Michael, and S. R. Bowman. Gpqa: A graduate-level google-proof q&a benchmark. InFirst Conference on Language Modeling, 2024

    2024

  15. [23]

    B. T. Willard and R. Louf. Efficient guided generation for large language models.arXiv preprint arXiv:2307.09702, 2023

    2023 arXiv

  16. [24]

    X. Wu, J. Yang, L. Chai, G. Zhang, J. Liu, X. Du, D. Liang, D. Shu, X. Cheng, T. Sun, et al. Tablebench: A comprehensive and complex benchmark for table question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 25497–25506, 2025

    2025

  17. [25]

    A. Yang, B. Yang, B. Hui, B. Zheng, B. Yu, C. Zhou, C. Li, C. Li, D. Liu, F. Huang, G. Dong, H. Wei, H. Lin, J. Tang, J. Wang, J. Yang, J. Tu, J. Zhang, J. Ma, J. Xu, J. Zhou, J. Bai, J. He, J. Lin, K. Dang, K. Lu, K. Chen, K. Yang, M. Li, M. Xue, N. Ni, P. Zhang, P. Wang, R. ...

    2024 arXiv

  18. [26]

    A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Xia, X. Ren, X. Ren, Y ...

    2024 arXiv

  19. [27]

    wait”and“alternatively

    Z. Yu, Y . Wu, Y . Zhao, A. Cohan, and X.-P. Zhang. Z1: Efficient test-time scaling with code. arXiv preprint arXiv:2504.00810, 2025. 11 Appendix A Quantitative Reasoning Behavior Analysis To quantitatively analyze how the predictor influences the reasoning behavior of LLMs un...

    2025 arXiv

  20. [28]

    But wait, in the Asymptote code, points are defined using the dir function, which gives points on the unit circle. So maybe the side length is 1, and the coordinates are on the unit circle.Hold on, in the Asymptote code, points are being drawn on a unit circle, but then points...

    2024

  21. [29]

    Group 2: 101

    Looking at the mapping, 010 is 2. Group 2: 101. That’s 5. Group 3: 001. That’s 1. Group 4: 110. That’s 6.So, putting them together, the octal number is 2 5 1 6. So, 2516 8.W ait, let me double- check to make sure I didn’t make a mistake in grouping or conversion. Starting from...

  22. [31]

    So, the third octal digit is 1. 4. 110: That’s 1*(2 2) + 1*(2 1) + 0*(2 0) = 4 + 2 + 0 = 6. So, the fourth octal digit is 6.Putting it all together, the octal number is 2 5 1 6, which is 2516. W ait, let me double-check to make sure I didn’t make a mistake in the grouping or t...

  23. [32]

    Converting each group:- 010 is 2 - 101 is 5 - 001 is 1 - 110 is 6So, 2516

    Then, grouping into threes from the right: 010 101 001 110. Converting each group:- 010 is 2 - 101 is 5 - 001 is 1 - 110 is 6So, 2516. Hmm, that seems correct. But just to be thorough, maybe I should convert the original binary number to decimal first and then see if that matc...

  24. [110]

    Hmm, seems correct.Just to be thorough, maybe I should convert the original binary number to decimal first and then to octal, to see if I get the same result

    Converting each group:010 is 2, 101 is 5, 001 is 1, 110 is 6.Combined, that’s 2516. Hmm, seems correct.Just to be thorough, maybe I should convert the original binary number to decimal first and then to octal, to see if I get the same result. Let’s try that. (thinking length: ...

Pith tools

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