Pith. sign in

REVIEW 4 major objections 5 minor 58 references

This paper proposes Dual-Dimensional Consistency (DDC), which couples a confidence-weighted Bayesian stopping rule with trend-aware pruning of reasoning paths, and claims this cuts token consumption by over 10× while maintaining or exceedin

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

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

T0 review · deepseek-v4-flash

2026-08-03 02:16 UTC pith:4E44RJ5Z

load-bearing objection Interesting coupling of Bayesian early stopping with trend-aware pruning, but the published stopping rule cannot produce the reported 10-27x savings — needs a major fix before the central claim is credible. the 4 major comments →

arxiv 2605.15100 v2 pith:4E44RJ5Z submitted 2026-05-14 cs.AI

Dual-Dimensional Consistency: Balancing Budget and Quality in Adaptive Inference-Time Scaling

Rongman Xu , Yifei Li , Tianzhe Zhao , Yanrui Wu , Bo Li , Hang Yan This is my paper
classification cs.AI
keywords inference-time scalingself-consistencyadaptive computationhallucination pruningBayesian early stoppingtoken efficiencyconfidence weightingLLM reasoning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

DDC is a framework for adaptive inference-time scaling that treats sampling width and path depth as one coupled decision rather than two orthogonal ones. It maintains that early consensus without quality weighting can lock in hallucinations, while fixed-depth pruning can truncate correct but temporarily uncertain chains. The framework reuses a single path-quality signal—the minimum sliding-window token confidence—both as a weight in a Bayesian consensus and as the input to a spectral trend analysis that prunes noisy or decaying paths. Across five benchmarks and five model sizes (1.7B to 32B), it reports over 10× token savings with accuracy equal to or better than strong baselines. The paper also concedes a boundary: if a model's confidence is systematically inverted, no sampling-based method can recover, and DDC would stop early on confident mistakes.

Core claim

The paper's central discovery is that the two established ways of saving inference compute—early stopping on consensus (inter-path) and pruning shaky reasoning chains (intra-path)—can be driven by the same confidence signal and reinforce each other. Path-level confidence, the minimum mean token probability over a sliding window, is used in two ways: as a weight in a Beta-Binomial posterior that terminates sampling when the leading answer exceeds a 0.5 majority with 0.95 posterior probability, and as the time series whose eigendecomposition yields a Structural Instability Score that flags hallucination-like trends. With these mechanisms, the paper reports, for example, a 15.6% accuracy gain o

What carries the argument

The central object is the path-level confidence, computed as the minimum group confidence over sliding windows of token probabilities. It serves two roles: it is the weight in the Beta-Binomial evidence accumulation that triggers early termination, and it is the signal whose windowed position-velocity phase space is eigendecomposed to give the Structural Instability Score R, with per-query pruning thresholds set by an interquartile-range outlier rule. The coupling of these two roles is the paper's mechanism for balancing budget and quality.

Load-bearing premise

The method assumes token-level probabilities are informative about correctness: correct paths are on average more confident than incorrect ones, and a model that is confidently wrong on a whole class of questions cannot be saved by any sampling-based strategy (as the paper itself concedes in Section B.3).

What would settle it

Measure the average confidence of correct and incorrect reasoning paths on any target benchmark. If confident-but-wrong paths have average confidence no lower than correct paths, the sufficient condition in the paper's acceleration theorem is violated; DDC should fail to beat a fixed-budget baseline. A concrete experiment: run DDC and Self-Consistency on a set of questions a model answers confidently-but-incorrectly; if DDC does not terminate early or loses accuracy relative to Self-Consistency, the central claim collapses.

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

If this is right

  • DDC reports over 10× average token reduction across MATH-500, AMC23, AIME24, AIME25, and GPQA-diamond, with accuracy equal or higher than Self-Consistency and DeepConf-High on most configurations.
  • On Qwen3-4B/AIME25, DDC's 82.1% accuracy is 15.6 points above Self-Consistency while consuming about 27× fewer tokens—the paper's strongest single efficiency-accuracy result.
  • The scaling-law curves in the paper show DDC outranking the baselines at every token budget from 0.89×10^7 to 20×10^7 tokens, implying it shifts the efficiency-accuracy frontier rather than just one operating point.
  • Latency results for Qwen3-32B on AIME25 show up to 12.4× speedup over the fastest baseline, because the 2048-token sliding-window analysis costs only about 1.25×10^-7 of an LLM forward pass.
  • Ablations indicate both components are load-bearing: removing the confidence-weighted Bayesian stopping raises tokens about 2.5× on MATH-500, and removing the trend analysis drops AIME25 accuracy by 6.8%.

Where Pith is reading between the lines

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

  • The paper does not test DDC on systematically miscalibrated models, but its own assumptions imply a testable boundary: on a dataset where a model's common error is confidently wrong, DDC should terminate even earlier than the baselines and lose accuracy, because the Bayesian posterior grows the leading answer's weight for the wrong answer.
  • A natural extension the paper leaves implicit: the same trend score could be applied to a single reasoning path in real time as a hallucination early-warning signal, without any ensemble or voting.
  • If the confidence-weighting mechanism generalizes, then any existing self-consistency-style method could be retrofitted by re-weighting votes with per-path minimum confidence; the paper's ablations suggest this alone already improves accuracy and token efficiency.
  • The paper's boundary claim implies a research program: post-hoc calibration of token probabilities before weighting would be the most direct way to push the method into systematically miscalibrated regimes—this is a speculation beyond the text.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Dual-Dimensional Consistency (DDC), an adaptive inference-time scaling method. DDC combines a Confidence-Weighted Bayesian Termination rule that stops sampling when the Beta posterior for the leading answer is sufficiently confident, and a Trend-Aware Stratified Pruning method that prunes low-quality reasoning paths based on token-level confidence trajectories. The authors report evaluations on five reasoning benchmarks across several Qwen and DeepSeek models, claiming an average token reduction of over 10x with accuracy equal or superior to baselines such as Self-Consistency, Adaptive-Consistency, and DeepConf.

Significance. The conceptual contribution is attractive: jointly managing sampling width and path depth is a natural way to attack the efficiency-accuracy trade-off in inference-time scaling. The paper includes extensive ablations, sensitivity analyses, latency comparisons, and a complexity analysis, which are useful assets. However, the central quantitative claim is not reproducible from the published algorithm because the pseudocode for Bayesian termination is inconsistent with the main-text stopping rule. Since the efficiency claim is the paper's primary contribution, this inconsistency is load-bearing and must be resolved before the results can be assessed.

major comments (4)
  1. [Appendix D, Algorithm 2 vs §4.2, Eq. (5)] The stopping rule in Algorithm 2 is not equivalent to Eq. (5). Eq. (5) stops when 1 - I_gamma(alpha,beta) > tau_stop with gamma=0.5 and tau_stop=0.95, i.e., P(p_u* > 0.5 | D) > 0.95. Algorithm 2 computes p_lower = I^{-1}_{0.5}(alpha,beta), which is the median of the Beta posterior, and stops when 1 - p_lower > 0.95, equivalent to median(p_u*) < 0.05. For any posterior with alpha >= beta, the typical case for the leading answer after confidence-weighted updates, this condition essentially never triggers. Thus the reported 27x savings on AIME25 and the average >10x claim cannot be generated by the algorithm as published. If the experiments used Eq. (5), the pseudocode is not the evaluated method; if they used Algorithm 2, the reported token budgets are unexplained. Please correct the pseudocode, define the inverse-Beta quantile precisely, and provide code or an erratum.
  2. [Table 3 vs Table 1] The internal numbers for the DDC baseline are inconsistent across tables. For Qwen3-4B on MATH500, Table 1 reports Acc. 92.8 / Tokens 7.5, whereas Table 3 reports Acc. 92.4 / Tokens 10.2 for the same configuration (Binit=16, L=2048, B=512). This ~36% token discrepancy is larger than any plausible sampling noise and directly affects the ablation conclusions, e.g., the claim that removing the weighted Bayesian term increases tokens by ~2.5x. Please reconcile the exact experimental conditions or report per-seed statistics.
  3. [§5.2/AIME25 and Appendix A.2] AIME25 has only 15 problems. The headline gain of 15.6% over Self-Consistency corresponds to roughly 2-3 questions, and token consumption ratios on this subset are highly sensitive to a few long generations. The paper says runs are repeated 5 times but reports no error bars or per-run ranges anywhere. Please report variance or confidence intervals for the key accuracy and token values, especially for AIME25 and AMC23.
  4. [§4.3, Eq. (9)] The text describes a velocity penalty whose magnitude should depend on the rate of confidence decline. In Eq. (9), however, the penalty is eta * I(align<0) * align^2, where align = |v_{1,x}| * sign(vbar). Since sign(vbar) contributes only the sign, the penalty magnitude is determined solely by the eigenvector alignment |v_{1,x}|, not by |vbar|. Two paths with very different rates of confidence decrease receive identical penalties if their primary eigenvector alignments match. Please either modify the formula to include |vbar| (or its normalized value) or revise the text to state that only the sign of the mean velocity is used.
minor comments (5)
  1. [Appendix D, Algorithm 2] The comment 'lower bound of the beta distribution's confidence interval' is misleading; I^{-1}_{0.5} is a median, not a lower bound. Use standard quantile notation such as Beta^{-1}(0.05; alpha, beta) if a lower confidence bound is intended.
  2. [§4.3] Typos: 'caculate' should be 'calculate' and 'immdiately' should be 'immediately'.
  3. [Table 5] Table 5 includes Qwen3-0.6B, but the main results in Table 1 do not include this model. Consider aligning the model lists to avoid confusion.
  4. [Appendix B.2.2, Proposition B.2] The proof of Proposition B.2 is only a paragraph and does not formally derive that the instability score R is monotone in the specified failure modes. For a labeled proposition, a more explicit argument is needed.
  5. [References] Several references are formatted incorrectly, e.g., [10] 'Gelman, Carlin, JB, Stern, and Rubin' and [25] 'Sandy Lovie' should follow the journal's author/editor format.

Circularity Check

0 steps flagged

No circularity: DDC's stopping rule and pruning are self-contained; reported gains are empirical measurements.

full rationale

I walked the claimed derivation chain. The Bayesian termination (Eqs. 4-7, Algorithm 2) is a standard sequential Beta update driven by path-level confidence weights; the final weighted majority vote (Eq. 11) and the pruning rule (Eqs. 8-10) are specified directly from token probabilities and sliding-window statistics. The headline efficiency numbers (e.g., 27x token reduction on AIME25 in Section 5.2) are measured token counts, not outputs derived from the assumptions, so no fitted parameter is being relabeled as a prediction. Assumptions 2 and 3 in Appendix B are stated and empirically supported, and the paper explicitly acknowledges in B.3 that systematic miscalibration cannot be fixed by sampling-based methods, which is an honest boundary rather than a hidden circularity. The self-citation to MUR [47] appears in related work and as inspiration for trend-based pruning, but the method is fully specified and ablated in this paper, so the citation is contextual, not load-bearing. The only notable technical flaw I found is that Algorithm 2's median-based check (1 - I^{-1}_{0.5}(α,β) > 0.95) is not equivalent to Eq. (5)'s CDF check (1 - I_{0.5}(α,β) > 0.95); this is a reproducibility/internal-consistency concern, not a circularity concern, and does not affect the circularity score.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 1 invented entities

The central claim rests on a small set of domain assumptions about token confidence being informative and error diversity across paths. The five explicit hyperparameters plus the unspecified top-k are chosen by hand and evaluated on the same benchmarks, which limits the independence of the empirical results.

free parameters (7)
  • Binit (init paths) = 16
    Number of initial paths used to calibrate thresholds; sensitivity analysis shows higher/lower values reduce accuracy or increase tokens.
  • L (sliding window size) = 2048 tokens
    Sliding window length for group confidence and trend analysis; Table 7 shows performance varies with L.
  • τstop = 0.95
    Stopping confidence threshold for Bayesian termination; chosen by hand, not derived.
  • γ = 0.5
    Absolute majority threshold in the stopping rule.
  • η = 0.5
    Penalty coefficient in the Structural Instability Score R.
  • tier percentiles = top 10% / bottom 20%
    Percentiles in Algorithm 1 that classify paths into tiers; chosen by hand.
  • k (top-k for global confidence) = unspecified
    Number of top tokens used in Eq. (1); not given in the hyperparameter settings, so the implementation is under-specified.
axioms (6)
  • domain assumption Token-level softmax probabilities reflect reasoning correctness (local and global confidence are meaningful).
    Used throughout Section 4.1; the authors state in B.3 that the method 'inherently relies on the strong correlation between confidence and correctness.'
  • domain assumption Assumption 2: correct paths receive higher confidence weights on average (μ_c > μ_i).
    Required for Theorem B.1 and for the weighted voting to improve over unweighted; validated in Figure 6 on MATH-500, a benchmark used later for evaluation.
  • domain assumption Assumption 3: error modes are not perfectly correlated; expected weighted support for the correct answer exceeds any single incorrect answer.
    Used in Proposition B.2 to argue self-reinforcing hallucinations are suppressed by consensus; cited from self-consistency wisdom-of-crowds.
  • standard math Beta-binomial conjugacy and the Power Prior weighted-likelihood update.
    Formalizes the weighted Bayesian update in Section B.1.1.
  • domain assumption Tukey Fences (Q3 + 1.5·IQR) provides a valid query-specific outlier threshold for the instability score.
    Used in Section 4.3 and B.2.1 to prune paths; assumes the score distribution of good paths is not heavy-tailed in a way that makes 1.5·IQR inappropriate.
  • domain assumption Eigendecomposition of the position-velocity covariance separates stable trends from stochastic noise in confidence trajectories.
    Core of trend-aware pruning Section 4.3, inspired by SSA; assumes a large λ1−λ2 gap marks recoverable trends, isotropic λ marks hallucination.
invented entities (1)
  • Structural Instability Score R no independent evidence
    purpose: Quantifies how much a reasoning path's confidence trajectory looks like noise or sustained decline; paths with R above a Tukey-fence threshold are pruned.
    R is an internal heuristic derived from confidence statistics; it has no falsifiable handle outside the paper's own pipeline, and its relationship to hallucination is asserted rather than independently demonstrated.

pith-pipeline@v1.3.0-alltime-deepseek · 21309 in / 16003 out tokens · 147597 ms · 2026-08-03T02:16:21.127644+00:00 · methodology

0 comments
read the original abstract

Large Language Models (LLMs) have demonstrated remarkable abilities in reasoning. However, maximizing their potential through inference-time scaling faces challenges in trade-off between sampling budget and reasoning quality. Current strategies remain inefficient as they typically treat sampling width and depth as orthogonal objectives, where width consensus methods risk reinforcing hallucinations, while depth pruning mechanisms prematurely truncate complex yet valid reasoning chains. Therefore, we propose Dual-Dimensional Consistency (DDC), a unified framework that bridges path quality with adaptive termination. By coupling Confidence-Weighted Bayesian protocol with a Trend-Aware Stratified Pruning, our method ensures that computational resources are concentrated on high quality reasoning paths, filtering hallucinations while accelerating consensus. Evaluations across five benchmarks demonstrate that this approach reduces token consumption by over 10 times while maintaining or exceeding the accuracy of strong baselines across various LLMs.

Figures

Figures reproduced from arXiv: 2605.15100 by Bo Li, Hang Yan, Rongman Xu, Tianzhe Zhao, Yanrui Wu, Yifei Li.

Figure 1
Figure 1. Figure 1: Comparisons of different methods. (a) is traditional static scaling method with fixed [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the DDC inference process. Init Phase Confidence Calculation computes [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustrating the multi-granular confidence. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Inference-time scaling law on Qwen3-4B-Instruct. The horizontal axis denotes the token consumption, while the vertical axis represents the average accuracy on 5 benchmarks. Datasets and Models. We evaluate on five challeng￾ing reasoning benchmarks: MATH-500 [14], AMC23, AIME24, AIME25, and GPQA-diamond [31]. To make fair compraison, we evaluate on the same LLMs following previous works [9, 47], specificall… view at source ↗
Figure 5
Figure 5. Figure 5: Hyperparameter sensitivity analysis. We report average accuracy and token consumption [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Analyzing the confidence distribution for problems of varying complexity. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Empirical verification. To empirically validate Theorem B.1, we con￾ducted Monte Carlo simulations across vary￾ing question difficulties, where Easy Questions denotes level 1 selected from MATH500 [14], Medium Questions denotes level 3 and Hard Questions denotes level 5. As shown in [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Scaling behavior of different models. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗

Review history (2 revisions) →

discussion (0)

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

Reference graph

Works this paper leans on

58 extracted references · 31 linked inside Pith

  1. [1]

    Let’s sample step by step: Adaptive-consistency for efficient reasoning and coding with llms, 2023

    Pranjal Aggarwal, Aman Madaan, Yiming Yang, and Mausam. Let’s sample step by step: Adaptive-consistency for efficient reasoning and coding with llms, 2023. URL https:// arxiv.org/abs/2305.11860

    Pith/arXiv arXiv 2023

  2. [2]

    Bernardo and Adrian F

    José M. Bernardo and Adrian F. M. Smith. Bayesian theory.Wiley Series in Probability and Statistics, 1994

    1994

  3. [3]

    Le, Christopher Ré, and Azalia Mirhoseini

    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, 2024. URLhttps://arxiv.org/abs/2407.21787

    Pith/arXiv arXiv 2024

  4. [4]

    Time series analysis and its applications - with r examples.Journal of the American Statistical Association, 97(458):656–657, 2006

    George Casella, Stephen Fienberg, and Ingram Olkin. Time series analysis and its applications - with r examples.Journal of the American Statistical Association, 97(458):656–657, 2006

    2006

  5. [5]

    Power prior distributions for regression models.Statistical Science, 15(1):46–60, 2000

    Ibrahim Ming Hui Chen. Power prior distributions for regression models.Statistical Science, 15(1):46–60, 2000

    2000

  6. [6]

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

    Pith/arXiv arXiv 2026

  7. [7]

    Test-time model adaptation for quantized neural networks,

    Zeshuai Deng, Guohao Chen, Shuaicheng Niu, Hui Luo, Shuhai Zhang, Yifan Yang, Renjie Chen, Wei Luo, and Mingkui Tan. Test-time model adaptation for quantized neural networks,

  8. [8]

    Bert: Pre-training of deep bidirectional transformers for language understanding, 2019

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. URL https://arxiv.org/ abs/1810.04805

    Pith/arXiv arXiv 2019

  9. [9]

    Deep think with confidence, 2025

    Yichao Fu, Xuewei Wang, Yuandong Tian, and Jiawei Zhao. Deep think with confidence, 2025. URLhttps://arxiv.org/abs/2508.15260

    Pith/arXiv arXiv 2025

  10. [10]

    Bayesian data analysis.Wiley Interdisciplinary Reviews Cognitive Science, 1(5):658–676, 2014

    Gelman, Carlin, JB, Stern, and Rubin. Bayesian data analysis.Wiley Interdisciplinary Reviews Cognitive Science, 1(5):658–676, 2014

    2014

  11. [11]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. 2017

    2017

  12. [12]

    J. M. Hammersley and D. C. Handscomb. Monte carlo methods.Methuen & Co Ltd, 1964

    1964

  13. [13]

    Reasoning with language model is planning with world model, 2023

    Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu. Reasoning with language model is planning with world model, 2023. URL https: //arxiv.org/abs/2305.14992

    Pith/arXiv arXiv 2023

  14. [14]

    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,

  15. [15]

    Monte carlo strategies in scientific computing.Technometrics, 44(4): 403–404, 2002

    Hesterberg and Tim. Monte carlo strategies in scientific computing.Technometrics, 44(4): 403–404, 2002

    2002

  16. [16]

    Large language models cannot self-correct reasoning yet, 2024

    Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. Large language models cannot self-correct reasoning yet, 2024. URL https://arxiv.org/abs/2310.01798

    Pith/arXiv arXiv 2024

  17. [17]

    The power prior: theory and applications.Statistics in Medicine, 2015

    Joseph, G., Ibrahim, Ming-Hui, Chen, Yeongjin, Gwon, Fang, and Chen. The power prior: theory and applications.Statistics in Medicine, 2015

    2015

  18. [18]

    Scalable best-of-n selection for large language models via self-certainty, 2025

    Zhewei Kang, Xuandong Zhao, and Dawn Song. Scalable best-of-n selection for large language models via self-certainty, 2025. URLhttps://arxiv.org/abs/2502.18581

    arXiv 2025

  19. [19]

    Reliability-aware adaptive self-consistency for efficient sampling in llm reasoning, 2026

    Junseok Kim, Nakyeong Yang, Kyungmin Min, and Kyomin Jung. Reliability-aware adaptive self-consistency for efficient sampling in llm reasoning, 2026. URL https://arxiv.org/ abs/2601.02970

    Pith/arXiv arXiv 2026

  20. [20]

    J. F. C Kingman. Random processes.Nature, 223(5202):219–219, 1969

    1969

  21. [21]

    Criticeval: Evaluating large language model as critic, 2024

    Tian Lan, Wenwei Zhang, Chen Xu, Heyan Huang, Dahua Lin, Kai Chen, and Xian ling Mao. Criticeval: Evaluating large language model as critic, 2024. URL https://arxiv.org/abs/ 2402.13764

    Pith/arXiv arXiv 2024

  22. [22]

    Modeling uncertainty trends for timely retrieval in dynamic rag, 2025

    Bo Li, Tian Tian, Zhenghua Xu, Hao Cheng, Shikun Zhang, and Wei Ye. Modeling uncertainty trends for timely retrieval in dynamic rag, 2025. URL https://arxiv.org/abs/2511. 09980

    2025

  23. [23]

    On the performance analysis of momentum method: A frequency domain perspective, 2025

    Xianliang Li, Jun Luo, Zhiwei Zheng, Hanxiao Wang, Li Luo, Lingkun Wen, Linlong Wu, and Sheng Xu. On the performance analysis of momentum method: A frequency domain perspective, 2025. URLhttps://arxiv.org/abs/2411.19671

    Pith/arXiv arXiv 2025

  24. [24]

    Escape sky-high cost: Early-stopping self-consistency for multi-step reasoning,

    Yiwei Li, Peiwen Yuan, Shaoxiong Feng, Boyuan Pan, Xinglin Wang, Bin Sun, Heda Wang, and Kan Li. Escape sky-high cost: Early-stopping self-consistency for multi-step reasoning,

  25. [25]

    Exploratory data analysis.John Wiley & Sons, Ltd, 2008

    Sandy Lovie. Exploratory data analysis.John Wiley & Sons, Ltd, 2008. 11

    2008

  26. [26]

    Non-myopic generation of language models for reasoning and planning, 2024

    Chang Ma, Haiteng Zhao, Junlei Zhang, Junxian He, and Lingpeng Kong. Non-myopic generation of language models for reasoning and planning, 2024. URL https://arxiv.org/ abs/2410.17195

    Pith/arXiv arXiv 2024

  27. [27]

    Self-refine: Iterative refinement with self-feedback, 2023

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. Self-refine: Iterative refinement with self-feedback, 2023. URL https://arxiv.org/abs/2303.17651

    Pith/arXiv arXiv 2023

  28. [28]

    Detecting hallucinations in large language model generation: A token probability approach, 2024

    Ernesto Quevedo, Jorge Yero, Rachel Koerner, Pablo Rivas, and Tomas Cerny. Detecting hallucinations in large language model generation: A token probability approach, 2024. URL https://arxiv.org/abs/2405.19648

    Pith/arXiv arXiv 2024

  29. [29]

    Learning transferable visual models from natural language supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. URL https://arxiv.org/abs/2103.00020

    Pith/arXiv arXiv 2021

  30. [30]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. URLhttps://arxiv.org/abs/1910.10683

    Pith/arXiv arXiv 2023

  31. [31]

    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, 2023. URLhttps://arxiv.org/abs/2311.12022

    Pith/arXiv arXiv 2023

  32. [32]

    Think just enough: Sequence-level entropy as a confidence signal for llm reasoning, 2025

    Aman Sharma and Paras Chopra. Think just enough: Sequence-level entropy as a confidence signal for llm reasoning, 2025. URLhttps://arxiv.org/abs/2510.08146

    arXiv 2025

  33. [33]

    Scaling llm test-time compute optimally can be more effective than scaling model parameters, 2024

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute optimally can be more effective than scaling model parameters, 2024. URL https://arxiv. org/abs/2408.03314

    Pith/arXiv arXiv 2024

  34. [34]

    Llm-check: Investigating detection of hallucinations in large language models

    Gaurang Sriramanan, Siddhant Bharti, Vinu Sankar Sadasivan, Shoumik Saha, Priyatham Kattakinda, and Soheil Feizi. Llm-check: Investigating detection of hallucinations in large language models. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Pa- quet, J. Tomczak, and C. Zhang, editors,Advances in Neural Information Processing Sys- tems, volume 37, page...

    2024

  35. [35]

    Ziegler, Ryan Lowe, Chelsea V oss, Alec Radford, Dario Amodei, and Paul Christiano

    Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea V oss, Alec Radford, Dario Amodei, and Paul Christiano. Learning to summarize from human feedback,

  36. [36]

    Silverman

    James Surowiecki and Mark P. Silverman. The wisdom of crowds.American Journal of Physics, 75(2):190–192, 2005

    2005

  37. [37]

    Understanding robust and exploratory data analysis.Wiley,

    Technometrics. Understanding robust and exploratory data analysis.Wiley,

  38. [38]

    Exploratory data analysis.Reading/Addison-Wesley, 1977

    John W Tukey. Exploratory data analysis.Reading/Addison-Wesley, 1977

    1977

  39. [39]

    A. Wald. Sequential tests of statistical hypotheses.Annals of Mathematical Statistics, 16(2): 117–186, 1945

    1945

  40. [40]

    Make every penny count: Difficulty-adaptive self-consistency for cost-efficient reasoning, 2025

    Xinglin Wang, Shaoxiong Feng, Yiwei Li, Peiwen Yuan, Yueqi Zhang, Chuyi Tan, Boyuan Pan, Yao Hu, and Kan Li. Make every penny count: Difficulty-adaptive self-consistency for cost-efficient reasoning, 2025. URLhttps://arxiv.org/abs/2408.13457

    Pith/arXiv arXiv 2025

  41. [41]

    Self-consistency improves chain of thought reasoning in language models, 2023

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models, 2023. URLhttps://arxiv.org/abs/2203.11171. 12

    Pith/arXiv arXiv 2023

  42. [42]

    Chain-of-thought prompting elicits reasoning in large language models, 2023

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. URLhttps://arxiv.org/abs/2201.11903

    Pith/arXiv arXiv 2023

  43. [43]

    L. Weiss. Sequential estimation.Wiley StatsRef: Statistics Reference Online, 2014

    2014

  44. [44]

    Wickens and D. Thomas. Categorical data analysis.Annual Review of Psychology, 49(1): 537–557, 1998

    1998

  45. [45]

    Inference scaling laws: An empirical analysis of compute-optimal inference for problem-solving with language models, 2025

    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, 2025. URLhttps://arxiv.org/abs/2408.00724

    Pith/arXiv arXiv 2025

  46. [46]

    ϕ- decoding: Adaptive foresight sampling for balanced inference-time exploration and exploitation,

    Fangzhi Xu, Hang Yan, Chang Ma, Haiteng Zhao, Jun Liu, Qika Lin, and Zhiyong Wu. ϕ- decoding: Adaptive foresight sampling for balanced inference-time exploration and exploitation,

  47. [47]

    Mur: Momentum uncertainty guided reasoning for large language models, 2025

    Hang Yan, Fangzhi Xu, Rongman Xu, Yifei Li, Jian Zhang, Haoran Luo, Xiaobao Wu, Luu Anh Tuan, Haiteng Zhao, Qika Lin, and Jun Liu. Mur: Momentum uncertainty guided reasoning for large language models, 2025. URLhttps://arxiv.org/abs/2507.14958

    Pith/arXiv arXiv 2025

  48. [48]

    Griffiths, Yuan Cao, and Karthik Narasimhan

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models, 2023. URLhttps://arxiv.org/abs/2305.10601

    Pith/arXiv arXiv 2023

  49. [49]

    Limo: Less is more for reasoning, 2025

    Yixin Ye, Zhen Huang, Yang Xiao, Ethan Chern, Shijie Xia, and Pengfei Liu. Limo: Less is more for reasoning, 2025. URLhttps://arxiv.org/abs/2502.03387

    Pith/arXiv arXiv 2025

  50. [50]

    Tokur: Token-level uncertainty estimation for large language model reasoning, 2025

    Tunyu Zhang, Haizhou Shi, Yibin Wang, Hengyi Wang, Xiaoxiao He, Zhuowei Li, Haoxian Chen, Ligong Han, Kai Xu, Huan Zhang, Dimitris Metaxas, and Hao Wang. Tokur: Token-level uncertainty estimation for large language model reasoning, 2025. URL https://arxiv.org/ abs/2505.11737

    Pith/arXiv arXiv 2025

  51. [51]

    URLhttps://arxiv.org/abs/2503.13288

    Pith/arXiv arXiv

  52. [56]

    X n wnI(yn =u ∗) # >max u̸=u∗ E

    Yuxin Zuo, Kaiyan Zhang, Li Sheng, Shang Qu, Ganqu Cui, Xuekai Zhu, Haozhan Li, Yuchen Zhang, Xinwei Long, Ermo Hua, Biqing Qi, Youbang Sun, Zhiyuan Ma, Lifan Yuan, Ning Ding, and Bowen Zhou. Ttrl: Test-time reinforcement learning, 2025. URL https://arxiv. org/abs/2504.16084. 13 A Implementation Details For all experiments, we set the temperature to 0.6 a...

    Pith/arXiv arXiv 2025

  53. [57]

    2.Decaying Hallucinationsare pruned via the velocity penalty term inR

    Stochastic Hallucinationsare pruned with high probability due to their maximal instability sores. 2.Decaying Hallucinationsare pruned via the velocity penalty term inR

  54. [58]

    Self-Reinforcing Hallucinationspass the intra-path filter but are suppressed in the inter- path consensus under Assumption 3. Proof. The first two claims follow directly from the construction of R and the properties of Tukey Fences for outlier detection [37]. The third claim follows from Assumption 3 and the consistency of weighted Bayesian aggregation [2...

    arXiv 2011

  55. [2021]

    URLhttps://arxiv.org/abs/2103.03874

    Pith/arXiv arXiv

  56. [2022]

    URLhttps://arxiv.org/abs/2009.01325

    Pith/arXiv arXiv 2009

  57. [2024]

    URLhttps://arxiv.org/abs/2401.10480

    Pith/arXiv arXiv

  58. [2025]

    URLhttps://arxiv.org/abs/2508.02180

    Pith/arXiv arXiv