pith. sign in

arxiv: 2605.27712 · v1 · pith:MHD3T642new · submitted 2026-05-26 · 💻 cs.AI

Prefix-Safe Bayesian Belief Tracking for LLM Reasoning Reliability:Separating Calibration from Ranking

Zhenghan Song , Yunyi Li , Yulong Liu This is my paper

Pith reviewed 2026-06-29 16:55 UTC · model grok-4.3

classification 💻 cs.AI
keywords Bayesian belief trackingprefix-safe observationsLLM reasoning reliabilitycalibrationrankingeventual success estimationsequential updating
0
0 comments X

The pith

Sequential Bayesian Belief Tracking separates calibration quality from ranking performance using prefix-safe observations in LLM reasoning.

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

The paper introduces Sequential Bayesian Belief Tracking to estimate the eventual success probability of a reasoning trace from prefix information alone. It works by first calibrating how likely each observation is under success versus failure, then recursively updating a simple two-state belief. Experiments across math benchmarks show that scalar-score versions mainly improve how closely the probabilities match reality, whereas ranking which traces will succeed requires richer structural signals. This split matters because many applications need reliability estimates before the trace finishes. The same tracker handles scores, text markers, self-verification, clusters, and latent features under the prefix-safety constraint.

Core claim

Sequential Bayesian Belief Tracking (SBBT) calibrates observation likelihoods and recursively updates a two-state belief for prefix-conditioned eventual-success estimation. Score-only SBBT improves Brier score for probability quality, while structure-aware observations deliver AUROC gains up to +0.110 against strong prefix-safe baselines on hard math traces; text markers and self-verification signals remain positive under same-prefix audits.

What carries the argument

Sequential Bayesian Belief Tracking (SBBT): a recursive two-state belief updater whose observation likelihoods are calibrated in advance.

If this is right

  • Score-only SBBT improves probability quality measured by Brier score.
  • Structure-aware observations produce AUROC gains that scalar scores alone do not achieve.
  • SBBT unifies tracking for scalar scores, text markers, self-verification, hidden clusters, token-pooling probes, and latent-trajectory features.
  • MATH-500 text markers and RIMO-N self-verification signals remain positive under same-prefix classifier audit.
  • Scalar scores mainly aid calibration while structure-aware signals aid ranking only when baselines have not already captured the rank evidence.

Where Pith is reading between the lines

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

  • Real-time reliability monitoring during generation becomes feasible without waiting for the completed trace.
  • Effort on new prefix-safe probes should target structural features rather than additional scalar scores to improve ranking decisions.
  • The calibration-ranking split may appear in other sequential estimation settings where only partial observations are available.
  • Deployed systems could route traces differently based on which observation type drives their reliability estimate.

Load-bearing premise

The chosen observations must stay strictly prefix-safe and add no information about the final answer beyond what the prefix already contains.

What would settle it

Showing that any structure-aware observation used for the AUROC gains actually leaks information about the eventual answer would eliminate the claimed ranking improvement.

Figures

Figures reproduced from arXiv: 2605.27712 by Yulong Liu, Yunyi Li, Zhenghan Song.

Figure 1
Figure 1. Figure 1: Prefix-safe belief tracking workflow. Offline calibration fits observation functions, likelihoods, transition [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Main evidence map. Panel A reports only positive structure-aware AUROC gains over the stan￾dard prefix-safe baseline set, excluding the PFC audit. Rows without rank gain are marked as stress tests and omitted from the positive-gain panel. This compact positive-gain map is paired with the signed-gap view in Appendix E. Stress rows remain visible in the signed ev￾idence. Panel B reports score-only Brier impr… view at source ↗
Figure 4
Figure 4. Figure 4: RIMO-N cross-model observation evidence. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Reliability curves for representative MATH-500 Level 5 and GSM8K rows. The curves compare raw last-prefix score, calibrated last-prefix score, EMA, calibrated EMA, score-only SBBT, and hybrid SBBT readouts. They support the calibration/ranking decoupling analysis and motivate treating identity state readout as a model score unless an outcome readout is fitted. D Additional Results [PITH_FULL_IMAGE:figures… view at source ↗
Figure 6
Figure 6. Figure 6: Rollout-based calibration diagnostic. Completed RIMO-N DeepSeek, RIMO-N Qwen3, and MATH-500 Level 5 belief joins compare raw source scores and SBBT beliefs against empirical continuation success. Panel A shows Brier improvement after the belief join. Panel B, Change in correlation after belief join, reports ∆ correlation = SBBT belief minus source score for Pearson and Spearman correlations. Together, the … view at source ↗
Figure 7
Figure 7. Figure 7: Utility operating points from existing split-seed summaries. [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Layerwise RIMO outcomes. Train-split hidden clustering is positive at early, mid, and final layers, while direct hidden probes and activation trajectories remain negative or near-tie. This separates a useful RIMO hidden-state signal from two weaker hidden-state variants under the same split-seed protocol. Figures 9, 10, 11, and 12 collect the appendix evidence for the paper’s multi-view argument: split-see… view at source ↗
Figure 9
Figure 9. Figure 9: Split-seed AUROC-gap distributions for representative main and stress-test rows. In the pre-PFC split-seed view, structure-aware rows are positive across most question-level split seeds for MATH-500, GSM8K, and the RIMO observation families shown here; AIME self-verification and Qwen3-MATH score rows sit near or below zero, which motivates keeping them as stress-test rows. 24 [PITH_FULL_IMAGE:figures/full… view at source ↗
Figure 10
Figure 10. Figure 10: Observation-family transfer matrix. Cells summarize main gains, auxiliary gains, stress-test rows, and not-applicable entries. 0.3 0.4 0.5 0.6 0.7 0.8 0.9 AUROC 0.786 MATH-500 Level 5 / 7B 0.646 GSM8K / 7B 5 25 50 75 100 Observed prefix fraction (%) 0.3 0.4 0.5 0.6 0.7 0.8 0.9 AUROC 0.390 RIMO-N / DeepSeek-Qwen-14B 5 25 50 75 100 Observed prefix fraction (%) 0.531 RIMO-N / Qwen3-14B Fixed-fraction online … view at source ↗
Figure 11
Figure 11. Figure 11: Fixed-fraction prefix diagnostics with question-cluster bootstrap intervals. Available exports provide 5%, 25%, 50%, 75%, and 100% prefix fractions for MATH-500 Level 5, GSM8K, and two RIMO-N rows. MATH-500 Level 5 and GSM8K show late-prefix online gains, while the RIMO online hybrid curves remain stress-test rows despite final hidden-cluster and self-verification gains. 25 [PITH_FULL_IMAGE:figures/full_… view at source ↗
Figure 12
Figure 12. Figure 12: Prefix-safe baseline headroom diagnostic. Rows with high best-baseline AUROC leave less rank headroom; RIMO hidden-cluster observations improve rank against standard prefix-safe baselines before adding the same-prefix PFC audit, while AIME and Qwen3-MATH remain stress-test cases. Marker size encodes positive split-seed fraction. E Signed AUROC Gaps and Evidence Coverage [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 13
Figure 13. Figure 13: reports the signed AUROC gaps that are intentionally not used as the main visual summary. The negative values are important stress-test evidence: they show that score-only, token-pooling, AIME-style, and some Qwen-family MATH settings can be dominated by strong prefix-safe baselines [PITH_FULL_IMAGE:figures/full_fig_p026_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Multi-view diagnostic coverage across evaluated settings. Rows summarize which dataset/model settings have rank, probability, early-prefix, utility, rollout, and answer-audit evidence. Cells encode diagnostic coverage only. F Reproducibility Details This appendix records reproducibility details for the reported rows. Exact commands, local paths, and environment setup are not part of the paper text; the pa… view at source ↗
read the original abstract

Long reasoning traces need reliability estimates before final answers are known. We study prefix-conditioned eventual-success estimation, $P(y=1 \mid o_{1:t})$, using prefix-safe observations. Sequential Bayesian Belief Tracking (SBBT) calibrates observation likelihoods and recursively updates a two-state belief, providing a common tracker for scalar scores, text and self-verification markers, hidden clusters, token-pooling probes, and latent-trajectory features. Across generated open-weight traces on MATH-500, GSM8K, AIME 2025, and RIMO-N, probability quality and ranking separate: score-only SBBT often improves Brier, while AUROC gains require structure-aware evidence beyond strong prefix-safe baselines. In the strongest hard math setting, structure-aware observations reach +0.110 AUROC against standard prefix-safe baselines. Under a same-prefix classifier audit, MATH-500 text markers and RIMO-N self-verification signals remain positive. Together, these findings support SBBT as a calibration-aware online inference framework and expose an evidence regime: scalar scores mainly support probability quality, while structure-aware prefix signals support ranking only when strong prefix-safe baselines have not already absorbed the rank evidence.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 2 minor

Summary. The paper claims to introduce Sequential Bayesian Belief Tracking (SBBT) for prefix-conditioned eventual-success estimation in LLM reasoning using prefix-safe observations. It separates probability quality (Brier score improvements from score-only SBBT) from ranking (AUROC gains up to +0.110 from structure-aware observations) across MATH-500, GSM8K, AIME 2025, and RIMO-N, with validation via same-prefix classifier audit confirming positive contributions from text markers and self-verification signals.

Significance. If the results hold, the contribution lies in providing a recursive Bayesian framework that distinguishes calibration effects from ranking performance in online settings. The direct audit of the prefix-safety assumption is a notable strength, supporting the claim that structure-aware evidence adds value beyond strong baselines only when not already absorbed.

minor comments (2)
  1. [Abstract] Reporting the +0.110 AUROC gain without error bars or details on experimental variance limits the ability to gauge result stability.
  2. [Abstract] Additional information on how likelihood calibration is performed and the train/test splits would help confirm the out-of-sample nature of the reported improvements.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive evaluation of the manuscript and the recommendation for minor revision. We appreciate the recognition that the direct audit of the prefix-safety assumption strengthens the claim regarding structure-aware evidence, and that the framework usefully separates calibration from ranking in online settings.

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper defines SBBT as a recursive two-state Bayesian update that takes calibrated observation likelihoods as input and produces prefix-conditioned success probabilities. The abstract and skeptic analysis indicate that likelihood calibration is performed as part of the method, with explicit same-prefix classifier audits and separation of Brier vs. AUROC metrics reported on held-out generated traces. No equations or steps reduce by construction to the evaluation metrics themselves; the +0.110 AUROC gain is presented as an empirical outcome under the prefix-safety assumption that the paper directly tests. The derivation chain remains self-contained against external benchmarks with no load-bearing self-citations or fitted-input renamings.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on a two-state success/failure model and on the ability to calibrate observation likelihoods from prefix data; both are introduced without independent external benchmarks in the abstract.

free parameters (1)
  • observation likelihoods
    Abstract states that SBBT 'calibrates observation likelihoods'; these values are fitted or chosen per observation type.
axioms (2)
  • domain assumption Two-state belief model suffices to capture eventual success
    Recursive update is defined on a binary success/failure state (abstract).
  • domain assumption Observations can be treated as conditionally independent given the latent state
    Standard Bayesian filtering assumption required for the recursive update.

pith-pipeline@v0.9.1-grok · 5744 in / 1473 out tokens · 48781 ms · 2026-06-29T16:55:35.772827+00:00 · methodology

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

46 extracted references · 37 canonical work pages · 17 internal anchors

  1. [1]

    Guillaume Alain and Yoshua Bengio. 2018. https://arxiv.org/abs/1610.01644 Understanding intermediate layers using linear classifier probes . Preprint, arXiv:1610.01644

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  2. [2]

    Angelopoulos, Stephen Bates, Adam Fisch, Lihua Lei, and Tal Schuster

    Anastasios N. Angelopoulos, Stephen Bates, Adam Fisch, Lihua Lei, and Tal Schuster. 2025. https://arxiv.org/abs/2208.02814 Conformal risk control . Preprint, arXiv:2208.02814

    work page arXiv 2025

  3. [3]

    Baum, Ted Petrie, George Soules, and Norman Weiss

    Leonard E. Baum, Ted Petrie, George Soules, and Norman Weiss. 1970. http://www.jstor.org/stable/2239727 A maximization technique occurring in the statistical analysis of probabilistic functions of markov chains . The Annals of Mathematical Statistics, 41(1):164--171

    work page arXiv 1970

  4. [4]

    Yonatan Belinkov. 2021. https://arxiv.org/abs/2102.12452 Probing classifiers: Promises, shortcomings, and advances . Preprint, arXiv:2102.12452

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  5. [5]

    Glenn W. Brier. 1950. https://doi.org/10.1175/1520-0493(1950)078<0001:VOFEIT>2.0.CO;2 Verification of forecasts expressed in terms of probability . Monthly Weather Review, 78(1):1--3

    work page doi:10.1175/1520-0493(1950)078 1950

  6. [6]

    Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt. 2024. https://arxiv.org/abs/2212.03827 Discovering latent knowledge in language models without supervision . Preprint, arXiv:2212.03827

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  7. [7]

    Lingjiao Chen, Matei Zaharia, and James Zou. 2023. https://arxiv.org/abs/2305.05176 Frugalgpt: How to use large language models while reducing cost and improving performance . Preprint, arXiv:2305.05176

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  8. [8]

    Ziye Chen, Chengwei Qin, and Yao Shu. 2025. https://arxiv.org/abs/2509.07711 Rimo: An easy-to-evaluate, hard-to-solve olympiad benchmark for advanced mathematical reasoning . Preprint, arXiv:2509.07711

    work page arXiv 2025

  9. [9]

    Maciej Chrabaszcz, Aleksander Szymczyk, Marcin Sendera, Tomasz Trzcinski, and Sebastian Cygert. 2026. https://arxiv.org/abs/2605.18549 Monitoring the internal monologue: Probe trajectories reveal reasoning dynamics . Preprint, arXiv:2605.18549

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  10. [10]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. https://arxiv.org/abs/2110.14168 Training verifiers to solve math word problems . Preprint, arXiv:2110.14168

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  11. [11]

    Jasper Dekoninck, Nikola Jovanovic, Tim Gehrunger, Kari Rognvaldsson, Ivo Petrov, Chenhao Sun, and Martin Vechev. 2026. https://arxiv.org/abs/2605.00674 Beyond benchmarks: Matharena as an evaluation platform for mathematics with llms . Preprint, arXiv:2605.00674

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  12. [12]

    A. P. Dempster, N. M. Laird, and D. B. Rubin. 1977. http://www.jstor.org/stable/2984875 Maximum likelihood from incomplete data via the em algorithm . Journal of the Royal Statistical Society. Series B (Methodological), 39(1):1--38

    work page arXiv 1977

  13. [13]

    Shrey Desai and Greg Durrett. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.21 Calibration of pre-trained transformers . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 295--302, Online. Association for Computational Linguistics

    work page doi:10.18653/v1/2020.emnlp-main.21 2020

  14. [14]

    Roy Eisenstadt, Itamar Zimerman, and Lior Wolf. 2025. https://arxiv.org/abs/2506.07240 Overclocking llm reasoning: Monitoring and controlling thinking path lengths in llms . Preprint, arXiv:2506.07240

    work page arXiv 2025

  15. [15]

    Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and 1 others. 2024. https://doi.org/10.1038/s41586-024-07421-0 Detecting hallucinations in large language models using semantic entropy . Nature, 630:625--630

    work page doi:10.1038/s41586-024-07421-0 2024

  16. [16]

    Tom Fawcett. 2006. https://doi.org/10.1016/j.patrec.2005.10.010 An introduction to roc analysis . Pattern Recognition Letters, 27(8):861--874. ROC Analysis in Pattern Recognition

    work page doi:10.1016/j.patrec.2005.10.010 2006

  17. [17]

    Yonatan Geifman and Ran El-Yaniv. 2019. https://proceedings.mlr.press/v97/geifman19a.html S elective N et: A deep neural network with an integrated reject option . In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 2151--2159. PMLR

    2019

  18. [18]

    Tilmann Gneiting and Adrian E Raftery. 2007. Strictly proper scoring rules, prediction, and estimation. Journal of the American statistical Association, 102(477):359--378

    2007

  19. [19]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 others. 2024. https://arxiv.org/abs/2407.21783 The llama 3...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. https://proceedings.mlr.press/v70/guo17a.html On calibration of modern neural networks . In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 1321--1330. PMLR

    2017

  21. [21]

    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, and 175 others. 2025. https://doi.org/10.1038/s41586-025-09422-z Deepseek-r1 incentivizes reasoning in llms through reinforcement lear...

    work page doi:10.1038/s41586-025-09422-z 2025

  22. [22]

    Hand and Robert J

    David J. Hand and Robert J. Till. 2001. https://doi.org/10.1023/A:1010920819831 A simple generalisation of the area under the roc curve for multiple class classification problems . Machine Learning, 45(2):171--186

    work page doi:10.1023/a:1010920819831 2001

  23. [23]

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. https://openreview.net/forum?id=7Bywt2mQsCe Measuring mathematical problem solving with the MATH dataset . In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)

    2021

  24. [24]

    Hugging Face H4 . 2023. https://huggingface.co/datasets/HuggingFaceH4/MATH-500 MATH-500 . Hugging Face dataset

    2023

  25. [25]

    Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, Scott Johnston, Sheer El-Showk, Andy Jones, Nelson Elhage, Tristan Hume, Anna Chen, Yuntao Bai, Sam Bowman, Stanislav Fort, and 17 others. 2022. https://arxiv.org/abs/2207.05221 Language models (mostly...

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  26. [26]

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. https://arxiv.org/abs/2305.20050 Let's verify step by step . Preprint, arXiv:2305.20050

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  27. [27]

    Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. https://arxiv.org/abs/2205.14334 Teaching models to express their uncertainty in words . Preprint, arXiv:2205.14334

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  28. [28]

    Liangchen Luo, Yinxiao Liu, Rosanne Liu, Samrat Phatale, Meiqi Guo, Harsh Lara, Yunxuan Li, Lei Shu, Yun Zhu, Lei Meng, Jiao Sun, and Abhinav Rastogi. 2024. https://arxiv.org/abs/2406.06592 Improve mathematical reasoning in language models by automated process supervision . Preprint, arXiv:2406.06592

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  29. [29]

    MathArena . 2025. https://huggingface.co/datasets/MathArena/aime_2025 AIME 2025 . Hugging Face dataset

    2025

  30. [30]

    L.R. Rabiner. 1989. https://doi.org/10.1109/5.18626 A tutorial on hidden markov models and selected applications in speech recognition . Proceedings of the IEEE, 77(2):257--286

    work page doi:10.1109/5.18626 1989

  31. [31]

    Simo Sarkka and Lennart Svensson. 2023. Bayesian Filtering and Smoothing, 2 edition. Institute of Mathematical Statistics Textbooks. Cambridge University Press

    2023

  32. [32]

    Tran, Yi Tay, and Donald Metzler

    Tal Schuster, Adam Fisch, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Q. Tran, Yi Tay, and Donald Metzler. 2022. https://arxiv.org/abs/2207.07061 Confident adaptive language modeling . Preprint, arXiv:2207.07061

    work page arXiv 2022

  33. [33]

    Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher Manning. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.330 Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback . In Proceedings of the 2023 Conference on Em...

    work page doi:10.18653/v1/2023.emnlp-main.330 2023

  34. [34]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  35. [35]

    Vilas, Safoora Yousefi, Besmira Nushi, Eric Horvitz, and Vidhisha Balachandran

    Martina G. Vilas, Safoora Yousefi, Besmira Nushi, Eric Horvitz, and Vidhisha Balachandran. 2025. https://arxiv.org/abs/2510.10494 Tracing the traces: Latent temporal signals for efficient and accurate reasoning . Preprint, arXiv:2510.10494

    work page arXiv 2025

  36. [36]

    Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. 2024. https://doi.org/10.18653/v1/2024.acl-long.510 Math-shepherd: Verify and reinforce LLM s step-by-step without human annotations . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages...

    work page doi:10.18653/v1/2024.acl-long.510 2024

  37. [37]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  38. [38]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  39. [39]

    Zhihui Xie, Jizhou Guo, Tong Yu, and Shuai Li. 2024. https://arxiv.org/abs/2405.18711 Calibrating reasoning in language models with internal consistency . Preprint, arXiv:2405.18711

    work page arXiv 2024

  40. [40]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025 a . https://arxiv.org/abs/2505.09388 Qwen3 technical report . Preprint, arXiv:2505.09388

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  41. [41]

    An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, Keming Lu, Mingfeng Xue, Runji Lin, Tianyu Liu, Xingzhang Ren, and Zhenru Zhang. 2024. https://arxiv.org/abs/2409.12122 Qwen2.5-math technical report: Toward mathematical expert model via self-improvement . Preprint, arXiv:2409.12122

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  42. [42]

    Chenxu Yang, Qingyi Si, Yongjie Duan, Zheliang Zhu, Chenyu Zhu, Zheng Lin, Li Cao, and Weiping Wang. 2025 b . Dynamic early exit in reasoning models. arXiv preprint arXiv:2504.15895

    work page arXiv 2025

  43. [43]

    Anqi Zhang, Yulin Chen, Jane Pan, Chen Zhao, Aurojit Panda, Jinyang Li, and He He. 2025. https://arxiv.org/abs/2504.05419 Reasoning models know when they're right: Probing hidden states for self-verification . Preprint, arXiv:2504.05419

    work page arXiv 2025

  44. [44]

    Chujie Zheng, Zhenru Zhang, Beichen Zhang, Runji Lin, Keming Lu, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. 2025. https://arxiv.org/abs/2412.06559 Processbench: Identifying process errors in mathematical reasoning . Preprint, arXiv:2412.06559

    work page arXiv 2025

  45. [45]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  46. [46]

    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 word.in bbl.in capitalize " " * FUNCT...