Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Adaptive Termination for Multi-round Parallel Reasoning: An Universal Semantic Entropy-Guided Framework

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that semantic entropy, the spread of parallel answers' meanings, is a dependable training-free stop signal that lets multi-round reasoning finish earlier and score higher than a fixed budget.

desk verdict A plausible adaptive-stopping idea that isn't yet cleanly separated from a fixed round-2 policy; worth refereeing but needs the missing control. read the letter →

arxiv 2507.06829 v1 pith:3BDFVWRP submitted 2025-07-09 cs.CL

Zenan Xu , Zexuan Qiu , Guanhua Huang , Kun Li , Siheng Li , Chenchen Zhang , Kejiao Li , Qi Yi
show 4 more authors
Yuhao Jiang Bo Zhou Fengzong Lian Zhanhui Kang
This is my paper
classification cs.CL
keywords semanticentropymulti-roundparallelreasoningadaptiveterminationtest-timescalingoptimalstoppinglargelanguagemodelsanswerselectionuncertainty
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 is trying to establish that semantic entropy — how much the meanings of a model's parallel answers spread out — is a dependable, training-free stop signal for multi-round reasoning. Its proposed framework, SEAT, keeps refining a question while the answers disagree, and stops as soon as the entropy falls below either a precomputed threshold or a baseline set by the first reasoning round. The paper reports that accuracy climbs when this rule is applied: on math and science benchmarks, the adaptive version improves both a 32B and a 7B reasoning model, sometimes by double-digit percentage points, while using fewer rounds than the maximum budget. It also claims that early stopping prevents a 'semantic entropy collapse' in which small models repeat the same wrong answer with overconfidence.

What carries the argument

The load-bearing quantity is semantic entropy (SE), the entropy of the probability distribution over semantic clusters of sampled answers; in practice it is estimated by clustering the final-answer segments of $N$ parallel responses and weighting each cluster by the sum of its response probabilities. The termination policy is the second mechanism: a fixed rule stops when SE falls below the 20th-percentile threshold from a calibration sample, while the adaptive rule fixes the first round's SE as the baseline ($T=1$) and stops at the first later round with SE below that baseline. Together these convert a fixed $N\times M$ reasoning budget into an early-stopping procedure, with the terminal round's answers used as the final candidate pool.

What would settle it

Run SEAT on a question set built from known reasoning traps where the model's first-round answers are unanimous but incorrect; if the adaptive rule stops at round 2 and stays wrong while a forced round-3 refinement recovers the answer, the baseline assumption fails. A simpler global check is to compare SEAT with an always-stop-at-round-2 rule on the same five benchmarks: if the fixed rule matches or beats SEAT, then entropy timing contributes nothing beyond one extra refinement round.

Watch

Extended reading notes

Core claim

The central discovery is that the entropy of the semantic clusters formed by parallel answers is strongly negatively correlated with answer accuracy, and that this correlation can be exploited prospectively as a termination policy. Concretely, the paper reports accuracy falling from about 97% to 19% as semantic entropy rises on AIME-2025, and shows that a rule which continues to a later round only when entropy is high, and stops at the first round whose entropy is low, outperforms running a fixed multi-round budget on all five benchmarks tested. The rule is implemented in two variants: a calibrated threshold taken from the 20th percentile of the entropy distribution, and a threshold-free variant that uses the first round's entropy as a dynamic baseline, inspired by the secretary problem. The paper further reports that this stop rule prevents semantic entropy collapse in small models and composes with random selection, maximum-probability selection, and majority voting.

Load-bearing premise

The load-bearing assumption is that the first round's semantic entropy is a trustworthy baseline, so a low first-round entropy can never simply mean the model has converged early onto a confident wrong answer.

Editorial extensions

If this is right

  • If semantic entropy is a reliable quality signal, multi-round reasoning can be run with an adaptive compute budget instead of a fixed token or round budget, saving inference on questions the model already handles.
  • Because the threshold-free rule needs no pre-sampling, the same monitoring signal can transfer to new models or new parallel settings without recalibration.
  • Even a small parallel batch of two branches plus an entropy stop rule can yield accuracy gains that would otherwise require many more samples.
  • The reported semantic entropy collapse gives a diagnostic for small-model overconfidence during long reasoning traces, and early termination is a direct, training-free remedy.
  • The final-answer selection strategies of random pick, maximum probability, and majority voting all benefit from the entropy-guided stopping, so the method layers onto existing answer-aggregation schemes.

Reading between the lines

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

  • Because the reported figures show most runs stopping at round 2 and the paper does not compare against a fixed 'stop after round 2' rule, the gain attributable to entropy timing itself, rather than to simply having a second refinement round, is not yet isolated.
  • A direct stress test of the adaptive rule is to feed it questions where the first round's answers are nearly unanimous but wrong; the round-1 baseline would then signal 'converged' and stop refinement immediately, so such a benchmark could bound when the approach fails.
  • The entropy threshold was calibrated on math problems and then applied to a science benchmark, suggesting thresholds may transfer across quantitative domains, but whether SE generalizes to open-ended tasks such as summarization or dialogue remains untested.
  • A natural extension is to set the baseline from more than one round, for example a low quantile of the first two rounds, and measure the trade-off between early stopping and protection against an uninformative first round.
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 / 5 minor

Summary. The paper proposes SEAT, an inference-time framework that combines multi-round parallel reasoning with semantic-entropy-based adaptive termination. The framework runs N parallel chains for up to M rounds, feeds previous answers into subsequent prompts, computes semantic entropy over the round's answers, and stops either when SE falls below a pre-calibrated threshold or below the first-round SE baseline. The authors evaluate on five mathematical and scientific benchmarks with DeepSeek-R1-Distill-Qwen-7B/32B, reporting accuracy gains over a full-budget baseline and identifying a 'semantic entropy collapse' phenomenon in small models.

Significance. If the central claim holds, SEAT would be a practical, training-free method for allocating test-time compute, and the SE-collapse observation is interesting for small model reasoning. The paper's strengths include the use of publicly available models, five benchmarks, three answer-selection strategies, and an explicit threshold-free mechanism; the reported gains are large on AIME datasets. However, as detailed below, the missing fixed early-stop control and unresolved cross-round SE comparability mean that the causal role of the SE signal is not yet established.

major comments (3)
  1. [Section 3.3, Figure 4, Tables 1-2] The adaptive rule terminates at round 2 for over 70% of trajectories, but the paper never compares to a fixed stop-at-round-2 policy under identical prompts, N, and answer selection. Since the baseline is the full 8-round budget, the reported gains could be entirely an early-stopping or single-refinement effect rather than an effect of the SE-based stopping decision. Add explicit controls that always stop at round 2 (and optionally round 3) to isolate the contribution of the SE rule.
  2. [Section 2.3] The threshold-free rule sets the baseline to the round-1 SE and stops at the first later round with lower SE, but rounds 2 and beyond receive prompts that include the N previous answers (Section 2.1), so SE values are computed under different conditioning contexts and are not directly comparable. The paper provides no calibration or evidence that a lower SE in round 2 relative to round 1 indicates higher answer quality; it may simply reflect confirmation or anchoring. The choice T=1 is also not ablated. Please provide cross-round comparability evidence (e.g., evaluate SE on round-2-style prompts without previous answers, or report round-by-round SE distributions and their correlation with accuracy per round).
  3. [Section 2.2 and Section 3.1] The fixed threshold is derived by choosing the 20th percentile of the SE distribution after observing that roughly 80% of correct answers fall in that region on a proprietary 10k-problem set. This is a fitted parameter, and the paper does not report the sensitivity of results to the percentile choice, the calibration set size, or the 5% outlier exclusion. Without such analysis, the claim that the 80/20 pattern is a robust basis for a universal SE threshold is not supported. Please report a sensitivity sweep (e.g., percentiles 10, 20, 30) and the corresponding performance on the five benchmarks.
minor comments (5)
  1. [Throughout] There are several typos and formatting issues: 'MINERV A' should be 'MINERVA', 'propsoed' should be 'proposed', 'benchamrk' should be 'benchmark', and the title 'An Universal' should be 'A Universal'.
  2. [Figure 1] The caption for panel (b) says 'Performance on AIME-2025' while panel (a) says 'Math-500'; clarify the exact model and benchmark for the correlation plot, and add a legend to panel (a) for the two curves if both are present.
  3. [Equation 3] The notation in Equation 3 is ambiguous: the sum over clusters should be indexed by k, and the use of |C| in the prefactor is confusing because the number of clusters is also |C|. Please rewrite the expression with clear indices and define all symbols.
  4. [Section 3.1] The sentence 'each query across all datasets will be conducted 8 independent multi-round parallel inferences' is unclear about whether the reported accuracy is an average over 8 independent runs or over 8 samples per query; clarify the experimental protocol.
  5. [Figure 5] The figure has axes labeled 'Semantic Entropy' and 'Expect accuracy' but no legend for the E(Acc.) curve; add a legend and a clear caption describing that E(Acc.) is the expected accuracy over the N parallel responses.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the adaptive mechanism is threshold-free and not fitted to labels, and the fixed-threshold calibration is on separate data.

full rationale

The paper's central adaptive termination rule (Section 2.3) sets a baseline from the round-1 semantic entropy (T=1) and stops at the first later round whose SE falls below that baseline. This rule contains no fitted parameters and does not use any evaluation labels; its reported accuracy gains are measured outcomes, not constructions. The alternative fixed-threshold rule (Section 2.2) is calibrated on a separate proprietary dataset by taking the 20th percentile of the SE distribution, and the same SE distribution is later used to explain improvements on AIME/MATH-500/MINERVA/GPQA; however, the threshold only controls when to stop, and the final answer is selected from the terminal round's candidate pool by random/max-probability/majority-vote, so the benchmark accuracies are not forced by the calibration. SE itself is defined purely from the model's own answer probability estimates (Eqs. 2-4), independent of correctness. The paper does not rely on self-citation as a load-bearing premise; the secretary-problem analogy and classic references are used only as inspiration. A genuine experimental concern is the absence of an 'always stop after round 2' control, especially since Figure 4 shows over 70% of runs terminate at round 2; this confound could make the gains appear to come from the SE signal when they might instead come from simply running a second refinement round. That is a confounding-variable issue rather than a derivation that reduces to its inputs by construction, so under the stated circularity criteria no circular step is established.

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

The central claims rest on an empirical correlation between semantic entropy and answer accuracy, on transferring thresholds fitted to a proprietary math set to public benchmarks, and on the validity of using answer-only probabilities in the SE estimator. The fixed thresholds are fitted parameters; the adaptive rule avoids fitting but assumes round-1 SE is a reliable anchor.

free parameters (5)
  • Fixed SE thresholds for 32B model = 27 (N=2), 22 (N=4), 19 (N=8)
    Derived from the SE distribution of 10,000 proprietary math problems at the 20th percentile and rounded to integers (Section 3.1).
  • Fixed SE thresholds for 7B model = 23 (N=2), 20 (N=4), 17 (N=8)
    Derived from the same proprietary SE distribution, at the 20th percentile, for the 7B model (Section 3.1).
  • 20th percentile cutoff = 0.20
    Chosen from the observed 80/20 pattern, where 74-78% of correct answers fall in the lowest quintile; this is a hand-picked threshold, not derived from a theory (Section 2.2, Figure 3).
  • Outlier exclusion rate = ~5%
    Roughly 5% of long-tail SE values were excluded before percentile calibration (Section 3.1, footnote 5). The choice is not justified by a rule.
  • Exploration window T = 1
    The adaptive rule uses only the first round's SE as the baseline; T=1 is chosen for computational convenience, with no sensitivity analysis (Section 2.3).
assumptions (5)
  • domain assumption Semantic entropy over final-answer clusters approximates reasoning quality across rounds.
    The entire adaptive control rests on this empirical correlation, established in Figure 1 and Section 1.
  • domain assumption The 80/20 distribution observed on a proprietary math set transfers to AIME, MATH-500, MINERVA, and GPQA.
    Fixed thresholds are calibrated on proprietary math problems and applied to these benchmarks (Sections 2.2, 3.1).
  • domain assumption Multi-round feedback prompting, which feeds previous answers into the next prompt, improves answer quality.
    The framework assumes that conditioning on prior parallel answers helps refinement (Section 2.1, prompt template).
  • ad hoc to paper The secretary-problem analogy justifies stopping at the first later round with SE below the round-1 baseline.
    The paper draws inspiration from optimal stopping but provides no proof that this heuristic matches the secretary-problem optimality conditions (Section 2.3).
  • domain assumption Answer-only probability estimates, despite inflated probabilities and negative SE values, preserve the ranking of rounds.
    Section 2.1's footnote 2 acknowledges inflated estimates and negative SE values but asserts the correlation remains valid.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Termination for Multi-round Parallel Reasoning: An Universal Semantic Entropy-Guided Framework." pith.science (2026). https://pith.science/paper/3BDFVWRP

@misc{pith2026250706829,
  author       = {Pith},
  title        = {Pith review of: Adaptive Termination for Multi-round Parallel Reasoning: An Universal Semantic Entropy-Guided Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3BDFVWRP}},
  note         = {Machine review of arXiv:2507.06829}
}
read the original abstract

Recent advances in large language models (LLMs) have accelerated progress toward artificial general intelligence, with inference-time scaling emerging as a key technique. Contemporary approaches leverage either sequential reasoning (iteratively extending chains of thought) or parallel reasoning (generating multiple solutions simultaneously) to scale inference. However, both paradigms face fundamental limitations: sequential scaling typically relies on arbitrary token budgets for termination, leading to inefficiency or premature cutoff; while parallel scaling often lacks coordination among parallel branches and requires intrusive fine-tuning to perform effectively. In light of these challenges, we aim to design a flexible test-time collaborative inference framework that exploits the complementary strengths of both sequential and parallel reasoning paradigms. Towards this goal, the core challenge lies in developing an efficient and accurate intrinsic quality metric to assess model responses during collaborative inference, enabling dynamic control and early termination of the reasoning trace. To address this challenge, we introduce semantic entropy (SE), which quantifies the semantic diversity of parallel model responses and serves as a robust indicator of reasoning quality due to its strong negative correlation with accuracy...

Figures

Figures reproduced from arXiv: 2507.06829 by the authors.

Figure 1
Figure 1. (a) The key insight of this paper: Strong negative correlation between semantic entropy and R1-Distill [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overview of our proposed SEAT. and a model’s parallel reasoning performance, empirically showing that higher SE often accompanies lower accuracy (e.g., 97% → 19% accuracy drop on AIME2025 as SE increases). This suggests SE as a robust indicator for quality assessment in parallel reasoning. • The SEAT Framework: We propose the SEAT framework, a semantic entropy (SE)-guided adaptive termination framework that syne… view at source ↗
Figure 3
Figure 3. Semantic entropy distribution of correct and incorrect answers. The lowest 20% threshold is marked by [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Round distribution across models and datasets during different parallel settings with adaptive termination. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The evolution of SE and expect accuracy across different inference rounds of R7B model. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 10 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Large language monkeys: Scaling inference compute with repeated sampling

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

    arXiv 2024

  3. [3]

    Seed-grpo: Semantic entropy enhanced grpo for uncertainty-aware policy optimization

    Minghan Chen, Guikun Chen, Wenguan Wang, and Yi Yang. Seed-grpo: Semantic entropy enhanced grpo for uncertainty-aware policy optimization. ArXiv, abs/2505.12346, 2025. URL https://api.semanticscholar.org/CorpusID:278741064

    arXiv 2025

  4. [4]

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

    Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, et al. Do not think that much for 2+ 3=? on the overthinking of o1-like llms. arXiv preprint arXiv:2412.21187, 2024

    arXiv 2024

  5. [5]

    Universal self-consistency for large language model generation

    Xinyun Chen, Renat Aksitov, Uri Alon, Jie Ren, Kefan Xiao, Pengcheng Yin, Sushant Prakash, Charles Sutton, Xuezhi Wang, and Denny Zhou. Universal self-consistency for large language model generation. ArXiv, abs/2311.17311, 2023. URL https://api.semanticscholar.org/CorpusID:265498407

    arXiv 2023

  6. [6]

    Training verifiers to solve math word problems

    Karl Cobbe, Vineet Kosaraju, Mo 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, abs/2110.14168, 2021. URL https://api.semanticscholar.org/CorpusID:239998651

    arXiv

  7. [7]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Jun-Mei Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiaoling Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bing-Li Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Dama...

    arXiv 2025

  8. [8]

    Detecting hallucinations in large language models using semantic entropy

    Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy. Nature, 630: 0 625 -- 630, 2024. URL https://api.semanticscholar.org/CorpusID:270615909

    work page 2024

Show all 37 references
  1. [9]

    Ferguson

    Thomas S. Ferguson. Who solved the secretary problem. Statistical Science, 4: 0 282--289, 1989. URL https://api.semanticscholar.org/CorpusID:62172567

    1989

  2. [10]

    Cognitive behaviors that enable self-improving reasoners, or, four habits of highly effective stars

    Kanishk Gandhi, Ayush Chakravarthy, Anikait Singh, Nathan Lile, and Noah D Goodman. Cognitive behaviors that enable self-improving reasoners, or, four habits of highly effective stars. arXiv preprint arXiv:2503.01307, 2025

    2025 arXiv

  3. [11]

    Measuring mathematical problem solving with the math dataset

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Xiaodong Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. NeurIPS Datasets and Benchmarks, 2021

    2021

  4. [12]

    Openai o1 system card

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

    2024 arXiv

  5. [13]

    Scalable best-of-n selection for large language models via self-certainty

    Zhewei Kang, Xuandong Zhao, and Dawn Xiaodong Song. Scalable best-of-n selection for large language models via self-certainty. ArXiv, abs/2502.18581, 2025. URL https://api.semanticscholar.org/CorpusID:276618155

    2025

  6. [14]

    Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation

    Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. ArXiv, abs/2302.09664, 2023. URL https://api.semanticscholar.org/CorpusID:257039062

    2023 arXiv

  7. [15]

    Solving quantitative reasoning problems with language models

    Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, et al. Solving quantitative reasoning problems with language models. Advances in Neural Information Processing Systems, 35...

    2022

  8. [16]

    Internal consistency and self-feedback in large language models: A survey

    Xun Liang, Shichao Song, Zifan Zheng, Hanyu Wang, Qingchen Yu, Xunkai Li, Rong-Hua Li, Feiyu Xiong, and Zhiyu Li. Internal consistency and self-feedback in large language models: A survey. ArXiv, abs/2407.14507, 2024. URL https://api.semanticscholar.org/CorpusID:271310469

    2024 arXiv

  9. [17]

    Let's verify step by step

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In The Twelfth International Conference on Learning Representations, 2024

    2024

  10. [18]

    Learning from peers in reasoning models

    Tongxu Luo, Wenyu Du, Jiaxi Bi, Stephen Chung, Zhengyang Tang, Hao Yang, Min Zhang, and Benyou Wang. Learning from peers in reasoning models. ArXiv, abs/2505.07787, 2025. URL https://api.semanticscholar.org/CorpusID:278534473

    2025 arXiv

  11. [19]

    Uncertainty estimation in autoregressive structured prediction

    Andrey Malinin and Mark John Francis Gales. Uncertainty estimation in autoregressive structured prediction. In International Conference on Learning Representations, 2021. URL https://api.semanticscholar.org/CorpusID:231895728

    2021

  12. [20]

    Zettlemoyer, Percy Liang, Emmanuel J

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Fei-Fei Li, Hanna Hajishirzi, Luke S. Zettlemoyer, Percy Liang, Emmanuel J. Candes, and Tatsunori Hashimoto. s1: Simple test-time scaling. ArXiv, abs/2501.19393, 2025. URL https://api.semanticscholar.org/CorpusID:276079693

    2025 arXiv

  13. [21]

    Learning to reason with llms

    OpenAI. Learning to reason with llms. https://openai.com/index/learning-to-reason-with-llms/, 2024

    2024

  14. [22]

    Learning adaptive parallel reasoning with language models

    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, abs/2504.15466, 2025. URL https://api.semanticscholar.org/CorpusID:277994172

    2025 arXiv

  15. [23]

    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. ArXiv, abs/2311.12022, 2023. URL https://api.semanticscholar.org/CorpusID:265295009

    2023 arXiv

  16. [24]

    Shiryaev

    Albert N. Shiryaev. Optimal stopping rules. In International Encyclopedia of Statistical Science, 1980. URL https://api.semanticscholar.org/CorpusID:46379699

    1980

  17. [25]

    Scaling llm test-time compute optimally can be more effective than scaling parameters for reasoning

    Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute optimally can be more effective than scaling parameters for reasoning. In International Conference on Learning Representations, 2025. URL https://api.semanticscholar.org/CorpusID:278498044

    2025

  18. [26]

    Qwq-32b: Embracing the power of reinforcement learning

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

    2025

  19. [27]

    Think twice: Enhancing llm reasoning by scaling multi-round test-time thinking

    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, abs/2503.19855, 2025. URL https://api.semanticscholar.org/CorpusID:277314100

    2025 arXiv

  20. [28]

    Solving math word problems with process-and outcome-based feedback

    Jonathan Uesato, Nate Kushman, Ramana Kumar, Francis Song, Noah Siegel, Lisa Wang, Antonia Creswell, Geoffrey Irving, and Irina Higgins. Solving math word problems with process-and outcome-based feedback. arXiv preprint arXiv:2211.14275, 2022

    2022 arXiv

  21. [29]

    Beyond the 80/20 rule: High-entropy minority tokens drive effective reinforcement learning for llm reasoning

    Shenzhi Wang, Le Yu, Chang Gao, Chujie Zheng, Shixuan Liu, Rui Lu, Kai Dang, Xionghui Chen, Jianxin Yang, Zhenru Zhang, Yuqiong Liu, An Yang, Andrew Zhao, Yang Yue, Shiji Song, Bowen Yu, Gao Huang, and Junyang Lin. Beyond the 80/20 rule: High-entropy minority tokens drive effe...

    2025

  22. [30]

    Chi, and Denny Zhou

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed H. Chi, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. International Conference on Learning Representations, 2022

    2022

  23. [31]

    Thoughts are all over the place: On the underthinking of o1-like llms

    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 b

    2025 arXiv

  24. [32]

    From decoding to meta-generation: Inference-time algorithms for large language models

    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, abs/2406.16838, 2024. URL https://api.semanticscholar.or...

    2024 arXiv

  25. [33]

    Revisiting the test-time scaling of o1-like models: Do they truly possess test-time scaling capabilities? ArXiv, abs/2502.12215, 2025

    Zhiyuan Zeng, Qinyuan Cheng, Zhangyue Yin, Yunhua Zhou, and Xipeng Qiu. Revisiting the test-time scaling of o1-like models: Do they truly possess test-time scaling capabilities? ArXiv, abs/2502.12215, 2025. URL https://api.semanticscholar.org/CorpusID:276421903

    2025 arXiv

  26. [34]

    Alphaone: Reasoning models thinking slow and fast at test time

    Junyu Zhang, Runpei Dong, Han Wang, Xuying Ning, Haoran Geng, Peihao Li, Xialin He, Yutong Bai, Jitendra Malik, Saurabh Gupta, and Huan Zhang. Alphaone: Reasoning models thinking slow and fast at test time. arXiv preprint arXiv:2505.24863, 2025

    2025 arXiv

  27. [35]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  28. [36]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  29. [37]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.