Pith. sign in

REVIEW 3 major objections 5 minor 77 references

Removing irrelevant and repetitive steps from a reasoning model's trace, guided by its own final-answer attention, substantially improves hallucination detection; REDE learns a lightweight projection to filter those steps before any detecto

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

2026-08-01 05:46 UTC pith:RZW6PY5I

load-bearing objection A solid empirical paper with a simple, effective idea—denoise reasoning traces before hallucination detection—that shows large AUROC gains; the main gaps are unvalidated attention supervision across architectures, missing code and baseline error bars, and a minor undisclosed self-citation. the 3 major comments →

arxiv 2607.22098 v1 pith:RZW6PY5I submitted 2026-07-24 cs.AI cs.LG

Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models

Junlin Fang , Do Nguyen-Thanh , Xiaogang Xu , Zhen Fang , Sean Du This is my paper
classification cs.AI cs.LG
keywords hallucination detectionlarge reasoning modelsreasoning tracesstep denoisingattention as supervisionrepresentation shapingkNN filteringREDE
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.

Large reasoning models emit long 'thinking' traces before answering, and this paper argues that those traces are littered with steps that actively hurt automatic hallucination detection. The central claim is that removing irrelevant and repetitive steps — using the model's own final-answer attention as an annotation-free guide — substantially improves the ability to tell whether the final answer is truthful. The paper introduces REDE, which learns a lightweight projection that makes informative steps cluster and noisy steps scatter, and then filters by k-NN distance before any detector runs. Across four benchmarks, two model families, and four detector types, detection AUROC improves consistently, with TruthfulQA rising from 68.63 to 87.32 under CCS. A sympathetic reader would care because the method is detector-agnostic and needs no human step labels, suggesting a cheap preprocessing step for more reliable hallucination detection.

Core claim

The paper's discovery, on its own terms, is that noisy reasoning steps — irrelevant asides and repetitions whose content is covered by later steps — are a hidden obstacle to hallucination detection in large reasoning models, and that they can be filtered without human labels by exploiting the model's internal attention. The final answer token's attention mass on each reasoning step provides an annotation-free score of step informativeness; training a lightweight projection to make high-attention steps cluster and low-attention steps scatter yields refined embeddings, and removing the steps with the largest k-NN distances in that space cleans the trace. The paper reports consistent AUROC impr

What carries the argument

The load-bearing object is the final-answer attention score (Eq. 1): the attention mass that the last token of the final answer assigns to each reasoning step, used as an automatic, annotation-free label of step informativeness. Steps are embedded via perplexity-weighted averaging of token hidden states (Eq. 2), then a lightweight two-layer MLP projection f_phi is trained with a three-term objective — compactness pulls informative steps together, dispersion keeps noisy steps scattered, and separation pushes the two groups apart (Eq. 3). At inference the projection is applied and steps are filtered by their k-NN cosine distance within the trace, with the most distant steps removed before any

Load-bearing premise

The whole method hinges on the premise that the attention the final answer token gives to each reasoning step is a dependable measure of that step's usefulness for truthfulness detection — a signal the paper validates only on AIME 2024 with Qwen3-8B and does not test on failure modes like hallucinated answers that attend heavily to a fabricated step.

What would settle it

On a benchmark where human annotation shows that early repetitive steps actually carry the decisive information (the opposite of the AIME pattern), or where final-answer attention concentrates on surface tokens rather than content-bearing steps, REDE's attention-supervised projection would be trained on wrong labels; if detector AUROC on the filtered trace then drops below the unfiltered baseline, the attention-as-supervision premise is falsified.

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

If this is right

  • Denoising is a detector-agnostic preprocessing step: CCS, supervised probing, perplexity, and verbalized certainty all improve when run on the filtered trace, so the benefit is not tied to any one detector.
  • No human annotation is needed for step-level supervision: final-answer attention provides the labels, and the projection is trained on frozen hidden states.
  • The learned projection transfers across datasets and to larger LRMs (32B), suggesting the denoising signal captures a general property of reasoning traces rather than dataset surface form.
  • Filtering is aggressive — up to 70% of steps are dropped — yet on MATH task accuracy only drops from 81.43% to 80.00% when the answer is regenerated from the filtered trace, so the removed steps are mostly uninformative for solving the problem as well.
  • Directly thresholding attention scores underperforms the learned projection, so the projection is doing real work beyond the raw signal.

Where Pith is reading between the lines

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

  • Since repetitive steps make up roughly half of annotated steps, the same attention-based signal could plausibly be used to make reasoning models generate shorter traces or to train them to suppress redundant thinking — a direction the paper notes only as future work.
  • The supervision signal is validated on one dataset/model pair; if attention correlates with answer-confidence rather than truthfulness, filtering might preferentially retain steps that lead to an answer regardless of correctness, potentially masking certain hallucination types.
  • The black-box proxy experiments suggest deployment on closed models using a separate white-box model to supply attention, but that introduces a distribution gap; the strong cross-dataset transfer hints that a single proxy projection might serve multiple target models, though this is not directly tested.
  • One could test whether the same filtered trace also improves consistency-based detectors such as self-consistency across multiple samples, since those detectors rely on trace-level features that may also be diluted by noise.

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

3 major / 5 minor

Summary. The paper proposes REDE, a preprocessing module for hallucination detection in large reasoning models (LRMs). REDE first scores each reasoning step by the attention mass from the final-answer token to that step (Eq. 1), uses the top/bottom attention quantiles as pseudo-labels to train a lightweight projection on step embeddings, and at inference removes the top ζ% of steps with largest k-NN cosine distance in the projected space. The filtered trace is then passed to downstream detectors (CCS, supervised probing, perplexity, verbalized certainty). Experiments on TruthfulQA, MATH, CodeElo, and MULTIHOPQA with Qwen3-8B/32B and DeepSeek-R1-Distill-Llama-8B/32B report consistent AUROC gains, e.g., 68.63→87.32 on TruthfulQA with CCS. Appendix E provides a conditional theoretical analysis based on assumptions about label-uninformativeness and local density separation.

Significance. If the empirical claims hold, the paper makes a useful, non-obvious contribution: it identifies noisy reasoning steps as an obstacle to hallucination detection and provides an annotation-free, detector-agnostic filtering method with substantial AUROC gains. The experimental scope is a clear strength: four datasets, two LRM families, four downstream detectors, extensive ablations, cross-dataset transfer, a black-box proxy experiment, human-annotation step-selection accuracy (86.37%), and task-accuracy preservation after 70% filtering. The central caveat is that the headline 'consistent improvement' rests on validation-set-selected hyperparameters, and the attention-supervision signal is directly validated on only one model/dataset combination; the theoretical analysis in Appendix E is conditional on assumptions that are not separately verified for the learned projection. No code is released, which limits reproducibility of the many configuration choices.

major comments (3)
  1. [Appendix C.4, Eq. (1)] The only human-annotated validation of the final-answer attention supervision is on AIME 2024 with Qwen3-8B (step-selection accuracy 86.37%). The main results for DeepSeek-R1-Distill-Llama-8B and for CodeElo/MULTIHOPQA traces do not include any analogous check. Since Eq. (1) is the sole supervision for the projection in Section 4.2, the claim that REDE 'generalizes across different LRM families' (Table 2) is supported only indirectly for DeepSeek. Please add step-selection accuracy against human annotations on at least one DeepSeek model and one non-math dataset, or otherwise validate that the attention signal is aligned with step informativeness outside Qwen3-8B/AIME.
  2. [Section 4.2, Section 5.2, Figure 6] REDE has several validation-selected hyperparameters: selection ratio ρ, drop ratio ζ, loss weights λ_disperse and λ_separate, projection dimension d′, number of neighbors k, and attention layer. The main tables report the method with configurations selected on a 100-example validation set, and Figure 6(b) shows performance peaking sharply at ζ=70%. This makes the headline 'consistent improvement' hard to interpret as a property of the method rather than of the selected configuration. I request a fixed-default hyperparameter run across all datasets, models, and detectors, or a nested evaluation, to show the gains are not an artifact of per-setting validation tuning.
  3. [Appendix E (Assumption E.4, Remark E.14, Eq. 3)] The theoretical chain does not establish that REDE's training objective actually creates the local density separation assumed in Assumption E.4 and used in Proposition E.5. Remark E.14 explicitly invokes an unverified 'population-level approximation' that the attention proxies are reasonably aligned with the true informative/noisy sets. Thus Proposition E.3, Proposition E.5, and Theorem E.11 are conditional on essentially the property the learned projection is supposed to produce. I suggest either empirically verifying Assumption E.4 (e.g., plotting empirical neighbor counts N_i(r) for human-labeled informative/noisy steps under the learned f_φ) or explicitly stating in the main text that the theory covers the downstream filtering step but not the projection-learning step.
minor comments (5)
  1. [Table 1] REDE results are reported with standard deviations over three seeds, but the baseline methods in Table 1 are not. Given the stated three-seed averaging, please report error bars for all baselines or state that baseline variability was not tracked.
  2. [Appendix A.4] No code or dataset link is provided. Given the number of hyperparameters and preprocessing choices (step segmentation, PPL weighting, validation selection), releasing code would materially improve reproducibility.
  3. [Section 4.3 vs. Appendix E] The nearest-neighbor parameter is denoted k in the main text and κ in Appendix E. Please unify the notation to avoid confusion.
  4. [Appendix C.4] The 86.37% step-selection accuracy is reported without confidence intervals or a class-balance baseline. The annotation has 73% noisy steps, so a trivial majority classifier would already be strong; please report precision/recall or a balanced accuracy metric.
  5. [Section 5.2 / Table 18] The position-based filtering baselines are evaluated only on Qwen3-8B. Since the DeepSeek traces have different length and repetition statistics (Table 5), a position-based comparison on DeepSeek would strengthen the claim that REDE's gain is not a length artifact for that model family.

Circularity Check

0 steps flagged

REDE's attention-derived supervision is internally derived, but the central hallucination-detection claim is tested on held-out labels and external annotations; no load-bearing circularity found.

full rationale

The derivation chain is not circular. REDE's projection is trained on contrastive pairs defined by final-answer attention (Eq. 1) and the objective in Eq. 3; however, the paper's central claim—that filtered traces improve hallucination detection—is evaluated on held-out truthfulness labels with four downstream detectors across two model families and four benchmarks, and cross-dataset transfer results are reported. Direct attention-based filtering (Table 4) is consistently weaker than the learned projection, so the gains are not simply attention scores reused at inference. The attention proxy is also checked against human annotations (Appendix C.4: 86.37% step-selection accuracy), which is an external sanity check rather than a fitted target. The only internal-support passages are in Appendix E, whose Assumption E.4 and Remark E.14 explicitly assume that the learned space separates informative from noisy steps and that the attention-based proxies are 'reasonably aligned' with true sets; these are formalized conditions, not the paper's empirical conclusion. Self-citations [66] and [70] appear as methodological precedent and baselines but are not load-bearing. There is at most a mild semantic overlap between 'informative' (defined as supporting the final answer in Appendix A.3) and the final-answer attention score, but the detection claim is not reduced to that overlap.

Axiom & Free-Parameter Ledger

6 free parameters · 7 axioms · 0 invented entities

The paper introduces no new physical entities. The central claim rests on several validation-selected hyperparameters, the validity of the attention signal, the human annotation ground truth, and the external judge. The theory adds assumptions that are plausible but not verified empirically.

free parameters (6)
  • Drop ratio zeta = 70% (default)
    Controls the fraction of steps removed at test time; performance peaks at 70% in Fig. 6b, selected on validation.
  • Selection ratio rho = 20-25% (default)
    Controls top/bottom attention proxies for training; validation-selected, Table 7.
  • Loss weights lambda_disperse, lambda_separate = default selected on validation
    Balance the contrastive loss terms; sensitivity shown in Fig. 13, but exact defaults are not stated.
  • Projection output dimension d' = 1024
    Ablation peaks at 1024 (Fig. 6c); validation-selected.
  • Number of nearest neighbors k = 15
    Validation-selected; Table 9 shows peak at 15.
  • Attention layer = final transformer layer
    Ablation in Fig. 14; last layer chosen as default.
axioms (7)
  • domain assumption Final-answer attention reflects step informativeness
    Section 3 and Eq. 1: the attention score from the final answer token to each reasoning step is assumed to indicate how much that step contributes to the final answer, serving as supervision for the projection.
  • domain assumption Human annotations on AIME 2024 are reliable ground truth for informative/noisy steps
    Appendix A.3: Fleiss' kappa 0.76, but only one dataset and one model (Qwen3-8B) are annotated; the validity of the noise taxonomy is assumed to generalize.
  • domain assumption External judge Qwen3-32B provides correct truthfulness labels
    Section 5: labels are obtained from a judge model; robustness to alternative labels is tested in Appendix C.5, but the judge itself is not independently verified.
  • ad hoc to paper Assumption E.1: noisy perturbation is approximately label-uninformative
    Appendix E.1: defines epsilon_mu and epsilon_c to bound mean and covariance of the noise term; the theory only holds in the strict-independence limit.
  • ad hoc to paper Assumption E.4: local density separation of informative and noisy steps in projected space
    Appendix E.3: assumes informative steps have at least 2*kappa neighbors within radius r_I and noisy steps at most kappa/2 within r_N; this is the core geometric assumption for kNN separation.
  • domain assumption Step segmentation via discourse markers and formatting yields semantically coherent steps
    Appendix A.3: the segmentation is heuristic; all subsequent attention scoring and embedding aggregation depend on these boundaries.
  • domain assumption Additive trace representation (Eq. 5)
    Appendix E.1: the trace representation is assumed to be the sum of prompt-answer component and averaged step contributions; used only in theory, but the practical detector representations may not be additive.

pith-pipeline@v1.3.0-alltime-deepseek · 38005 in / 13791 out tokens · 135324 ms · 2026-08-01T05:46:52.100211+00:00 · methodology

0 comments
read the original abstract

Large reasoning models (LRMs) generate long reasoning traces before producing final answers. While these traces may contain useful signals for hallucination detection, harnessing them is non-trivial because long trajectories often include noisy steps that obscure the cues relevant to truthfulness assessment. In this paper, we identify two prevalent forms of reasoning noises, i.e., irrelevant steps and repetitive steps, and show that both substantially degrade hallucination detection performance. Existing confidence-based scores and naive embedding-based filtering fail to reliably separate noisy from informative steps. To address this challenge, we propose REDE, a novel learning framework for denoising reasoning traces for hallucination detection. Specifically, REDE leverages final-answer attention as an automatic supervision signal to shape the step-level representation space, yielding refined embeddings in which noisy steps can be reliably identified and filtered. REDE can be readily plugged into diverse hallucination detectors by operating on the filtered reasoning trajectory after removing noisy steps. Extensive experiments on multiple reasoning benchmarks show that REDE consistently improves detection performance over competitive baselines.

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

77 extracted references · 16 linked inside Pith

  1. [1]

    Alain and Y

    G. Alain and Y . Bengio. Understanding intermediate layers using linear classifier probes.arXiv preprint arXiv:1610.01644, 2016

    Pith/arXiv arXiv 2016

  2. [2]

    S. An, R. Wang, T. Zhou, and C.-J. Hsieh. Don’t think longer, think wisely: Optimizing thinking dynamics for large reasoning models. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  3. [3]

    Azaria and T

    A. Azaria and T. Mitchell. The internal state of an LLM knows when it’s lying. InFindings of the Association for Computational Linguistics: EMNLP 2023, pages 967–976, 2023

    2023

  4. [4]

    F. Bao, C. Xu, and O. Mendelevitch. DeepSeek-R1 hallucinates more than DeepSeek-V3. https:// www.vectara.com/blog/deepseek-r1-hallucinates-more-than-deepseek-v3 , Jan. 2025. Vec- tara Blog, accessed 2025-04-02

    2025

  5. [5]

    Bhatnagar, Y

    R. Bhatnagar, Y . Sun, C. A. Zhang, Y . Wen, and H. Yang. HALT: Hallucination assessment via latent testing.arXiv preprint arXiv:2601.14210, 2026

    arXiv 2026

  6. [6]

    P. C. Bogdan, U. Macar, N. Nanda, and A. Conmy. Thought anchors: Which LLM reasoning steps matter?arXiv preprint arXiv:2506.19143, 2025

    arXiv 2025

  7. [7]

    Burns, H

    C. Burns, H. Ye, D. Klein, and J. Steinhardt. Discovering latent knowledge in language models without supervision. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  8. [8]

    C. Chen, K. Liu, Z. Chen, Y . Gu, Y . Wu, M. Tao, Z. Fu, and J. Ye. INSIDE: LLMs’ internal states retain the power of hallucination detection. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  9. [9]

    L. Chen, H. Zhao, T. Liu, S. Bai, J. Lin, C. Zhou, and B. Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. InEuropean Conference on Computer Vision, pages 19–35. Springer, 2024

    2024

  10. [10]

    Cheng, T

    J. Cheng, T. Su, J. Yuan, G. He, J. Liu, X. Tao, J. Xie, and H. Li. Chain-of-thought prompting obscures hallucination cues in large language models: An empirical evaluation. In C. Christodoulopoulos, T. Chakraborty, C. Rose, and V . Peng, editors,Findings of the Association for Computational Linguistics: EMNLP 2025. Association for Computational Linguistics, 2025

    2025

  11. [11]

    Cobbe, V

    K. Cobbe, V . Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakano, et al. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    Pith/arXiv arXiv 2021

  12. [12]

    Y . Cui, P. He, J. Zeng, H. Liu, X. Tang, Z. Dai, Y . Han, C. Luo, J. Huang, Z. Li, et al. Stepwise perplexity-guided refinement for efficient chain-of-thought reasoning in large language models. In Findings of the Association for Computational Linguistics: ACL 2025, pages 18581–18597, 2025

    2025

  13. [13]

    Devlin, M.-W

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

    2019

  14. [14]

    H. Do, J. Hwang, D. Han, S. J. Oh, and S. Yun. What defines good reasoning in LLMs? dissecting reasoning steps with multi-aspect evaluation.arXiv preprint arXiv:2510.20603, 2025. 12 Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models

    arXiv 2025

  15. [15]

    X. Du, C. Xiao, and Y . Li. HaloScope: Harnessing unlabeled LLM generations for hallucination detection.Advances in Neural Information Processing Systems, 37:102948–102972, 2024

    2024

  16. [16]

    Y . Fu, X. Wang, Y . Tian, and J. Zhao. Deep think with confidence.arXiv preprint arXiv:2508.15260, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Bi, et al. DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Hendrycks, C

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

    Pith/arXiv arXiv 2021

  19. [19]

    Ho, A.-K

    X. Ho, A.-K. D. Nguyen, S. Sugawara, and A. Aizawa. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. InProceedings of the 28th International Conference on Computational Linguistics, pages 6609–6625, 2020

    2020

  20. [20]

    B. Hong, J. Liu, K. Zhang, J. Sun, M. Zhang, and Z. Huang. Pruning long chain-of-thought of large reasoning models via small-scale preference optimization. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  21. [21]

    C. Hong, X. Guo, A. C. Singh, E. Choukse, and D. Ustiugov. Slim-SC: Thought pruning for efficient scaling with self-consistency. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 34488–34505, 2025

    2025

  22. [22]

    R. A. Horn and C. R. Johnson.Matrix analysis. Cambridge university press, 2012

    2012

  23. [23]

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

    Pith/arXiv arXiv 2025

  24. [24]

    Z. Hou, X. Lv, R. Lu, J. Zhang, Y . Li, Z. Yao, J. Li, J. Tang, and Y . Dong. T1: Advancing language model reasoning through reinforcement learning and inference scaling. InForty-second International Conference on Machine Learning, 2025

    2025

  25. [25]

    J. Hu, G. Tu, C. Shengyu, J. Li, J. Wang, R. Chen, Z. Zhou, and D. Shan. HARP: Hallucination detection via reasoning subspace projection. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  26. [26]

    Huggingfaceh4/aime_2024, 2025

    Hugging Face H4. Huggingfaceh4/aime_2024, 2025. URL https://huggingface.co/datasets/ HuggingFaceH4/aime_2024. Dataset

    2025

  27. [27]

    Jaech, A

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

    Pith/arXiv arXiv 2024

  28. [28]

    Jain and B

    S. Jain and B. C. Wallace. Attention is not explanation. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3543–3556, 2019

    2019

  29. [29]

    Kadavath, T

    S. Kadavath, T. Conerly, A. Askell, T. Henighan, D. Drain, E. Perez, N. Schiefer, Z. Hatfield-Dodds, N. DasSarma, E. Tran-Johnson, et al. Language models (mostly) know what they know.arXiv preprint arXiv:2207.05221, 2022

    Pith/arXiv arXiv 2022

  30. [30]

    L. Kuhn, Y . Gal, and S. Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. InThe Eleventh International Conference on Learning Representations, 2023. 13 Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models

    2023

  31. [31]

    J. R. Landis and G. G. Koch. The measurement of observer agreement for categorical data.biometrics, pages 159–174, 1977

    1977

  32. [32]

    Lee, K.-H

    H. Lee, K.-H. Park, H. Byun, J. Yeom, J. Kim, G.-M. Park, and K. Song. CED: Comparing embedding differences for detecting out-of-distribution and hallucinated text. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 14866–14882, 2024

    2024

  33. [33]

    Y . Li, Y . Huang, B. Yang, B. Venkitesh, A. Locatelli, H. Ye, T. Cai, P. Lewis, and D. Chen. SnapKV: LLM knows what you are looking for before generation.Advances in Neural Information Processing Systems, 37:22947–22970, 2024

    2024

  34. [34]

    Z. Li, J. Zhong, Z. Zheng, X. Wen, Z. Xu, Y . Cheng, F. Zhang, and Q. Xu. Making slow thinking faster: Compressing LLM chain-of-thought via step entropy. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  35. [35]

    Liang, B

    Z. Liang, B. Huang, Z. Wang, and M. Zhang. Hidden states as early signals: Step-level trace evaluation and pruning for efficient test-time scaling. InFindings of the Association for Computational Linguistics: ACL 2026, San Diego, California, USA, 2026. Association for Computational Linguistics

    2026

  36. [36]

    S. Lin, J. Hilton, and O. Evans. Teaching models to express their uncertainty in words.Transactions on Machine Learning Research, 2022. ISSN 2835-8856

    2022

  37. [37]

    S. Lin, J. Hilton, and O. Evans. TruthfulQA: Measuring how models mimic human falsehoods. In Proceedings of the 60th annual meeting of the association for computational linguistics (volume 1: long papers), pages 3214–3252, 2022

    2022

  38. [38]

    Z. Lin, S. Trivedi, and J. Sun. Generating with confidence: Uncertainty quantification for black-box large language models.Transactions on Machine Learning Research, 2024. ISSN 2835-8856

    2024

  39. [39]

    H. Lu, Y . Liu, J. Xu, G. Nan, Y . Yu, Z. Chen, and K. Wang. Auditing meta-cognitive hallucinations in reasoning large language models. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  40. [40]

    H. Lu, M. Pan, R. Li, G. Nan, J. Zhuang, Z. Zhao, Z. Sun, K. Wang, and Y . Liu. Streaming hallucination detection in long chain-of-thought reasoning.arXiv preprint arXiv:2601.02170, 2026

    arXiv 2026

  41. [41]

    W. Luo, G. Peng, W. Li, S. Wei, F. Song, L. Wang, N. Yang, X. Zhang, J. Jin, F. Wei, et al. Two pathways to truthfulness: On the intrinsic encoding of LLM hallucinations. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), San Diego, California, USA, 2026. Association for Computational Linguistics

    2026

  42. [42]

    Y . Luo, Y . Song, X. Zhang, J. Liu, W. Wang, G. Chen, W. Su, and B. Zheng. Deconstructing long chain-of-thought: A structured reasoning optimization framework for long cot distillation.arXiv preprint arXiv:2503.16385, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    Manakul, A

    P. Manakul, A. Liusie, and M. Gales. SelfCheckGPT: Zero-resource black-box hallucination detection for generative large language models. InProceedings of the 2023 conference on empirical methods in natural language processing, pages 9004–9017, 2023

    2023

  44. [44]

    Miralles-González, J

    P. Miralles-González, J. Huertas-Tato, A. Martín, and D. Camacho. Not all tokens are created equal: Perplexity attention weighted networks for ai generated text detection.arXiv preprint arXiv:2501.03940, 2025

    Pith/arXiv arXiv 2025

  45. [45]

    M. Niu, H. Haddadi, and G. Pang. Robust hallucination detection in LLMs via adaptive token selection. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. 14 Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models

    2025

  46. [46]

    AIME 2025 dataset, 2025

    OpenCompass Team. AIME 2025 dataset, 2025. URL https://huggingface.co/datasets/ opencompass/AIME2025. Problems from the American Invitational Mathematics Examination (AIME) 2025-I & II

    2025

  47. [47]

    S. Park, X. Du, M.-H. Yeh, H. Wang, and Y . Li. Steer LLM latents for hallucination detection. In International Conference on Machine Learning, pages 47971–47990. PMLR, 2025

    2025

  48. [48]

    Press, M

    O. Press, M. Zhang, S. Min, L. Schmidt, N. A. Smith, and M. Lewis. Measuring and narrowing the compositionality gap in language models. InFindings of the Association for Computational Linguistics: EMNLP 2023, pages 5687–5711, 2023

    2023

  49. [49]

    S. Quan, J. Yang, B. Yu, B. Zheng, D. Liu, A. Yang, X. Ren, B. Gao, Y . Miao, Y . Feng, et al. CodeElo: Benchmarking competition-level code generation of LLMs with human-comparable elo ratings.arXiv preprint arXiv:2501.01257, 2025

    Pith/arXiv arXiv 2025

  50. [50]

    J. Ren, J. Luo, Y . Zhao, K. Krishna, M. Saleh, B. Lakshminarayanan, and P. J. Liu. Out-of-distribution detection and selective generation for conditional language models. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  51. [51]

    Santosh, Y

    T. Santosh, Y . T. Elkhayat, O. Ichim, P. Shetty, D. Wang, Z. Ma, A. Nourbakhsh, and X. Liu. CoCoLex: Confidence-guided copy-based decoding for grounded legal text generation. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 19002–19018, 2025

    2025

  52. [52]

    Sellam, D

    T. Sellam, D. Das, and A. Parikh. BLEURT: Learning robust metrics for text generation. InProceedings of the 58th annual meeting of the association for computational linguistics, pages 7881–7892, 2020

    2020

  53. [53]

    P. Srey, X. Wu, and L. A. Tuan. Unsupervised hallucination detection by inspecting reasoning processes. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 22128–22140, 2025

    2025

  54. [54]

    Z. Sun, Q. Wang, H. Wang, X. Zhang, and J. Xu. Mechanistic detection and mitigation of hallucination in large reasoning models. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  55. [55]

    L. Tang, W. Gao, B. Zhao, L. Ma, B. Yang, Y . Zou, et al. Thinking by subtraction: Confidence-driven contrastive decoding for LLM reasoning.arXiv preprint arXiv:2602.18232, 2026

    arXiv 2026

  56. [56]

    Trivedi, N

    H. Trivedi, N. Balasubramanian, T. Khot, and A. Sabharwal. MuSiQue: Multihop questions via single- hop question composition.Transactions of the Association for Computational Linguistics, 10:539–554, 2022

    2022

  57. [57]

    S. Tu, Y . Li, Y . Bai, L. Hou, and J. Li. DeepPrune: Parallel scaling without inter-trace redundancy.arXiv preprint arXiv:2510.08483, 2025

    Pith/arXiv arXiv 2025

  58. [58]

    Varshney, W

    N. Varshney, W. Yao, H. Zhang, J. Chen, and D. Yu. A stitch in time saves nine: Detecting and mitigating hallucinations of LLMs by validating low-confidence generation.arXiv preprint arXiv:2307.03987, 2023

    Pith/arXiv arXiv 2023

  59. [59]

    C. Wang, W. Su, Q. Ai, and Y . Liu. Joint evaluation of answer and reasoning consistency for hallucination detection in large reasoning models. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 33377–33385, 2026. 15 Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models

    2026

  60. [60]

    S. Wang, E. Zhao, and X. Ren. Stepwise informativeness search for improving LLM reasoning. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 25291–25309, 2025

    2025

  61. [61]

    Wiegreffe and Y

    S. Wiegreffe and Y . Pinter. Attention is not not explanation. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pages 11–20, 2019

    2019

  62. [62]

    C. Wu, Q. Cao, C. Li, Z. Wang, C. Xue, Y . Fan, W. Xi, and X. He. Beyond token length: Step pruner for efficient and accurate reasoning in large language models.arXiv preprint arXiv:2510.03805, 2025

    arXiv 2025

  63. [63]

    H. Xia, C. T. Leong, W. Wang, Y . Li, and W. Li. TokenSkip: Controllable chain-of-thought compression in LLMs. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 3351–3363, 2025

    2025

  64. [64]

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025

  65. [65]

    Z. Yang, P. Qi, S. Zhang, Y . Bengio, W. Cohen, R. Salakhutdinov, and C. D. Manning. HotpotQA: A dataset for diverse, explainable multi-hop question answering. InProceedings of the 2018 conference on empirical methods in natural language processing, pages 2369–2380, 2018

    2018

  66. [66]

    Z. Yao, Y . Liu, Y . Chen, J. Chen, J. Fang, L. Hou, J. Li, and T.-S. Chua. Are reasoning models more prone to hallucination?arXiv preprint arXiv:2505.23646, 2025

    Pith/arXiv arXiv 2025

  67. [67]

    H. Yuan, B. Yu, H. Li, S. Yang, C. D. Wang, Z. Yu, X. Xu, W. Qi, and K. Chen. Not all tokens are what you need in thinking.arXiv preprint arXiv:2505.17827, 2025

    Pith/arXiv arXiv 2025

  68. [68]

    W. Zeng, Y . Wang, C. Hu, Y . Shi, C. Wan, H. Zhang, and X. Gu. Pruning the unsurprising: Efficient code reasoning via first-token surprisal.arXiv preprint arXiv:2508.05988, 2025

    arXiv 2025

  69. [69]

    X. Zeng, J. Lin, Y . Yan, F. Guo, L. Shi, J. Wu, and D. Zhou. HalluGuard: Demystifying data-driven and reasoning-driven hallucinations in LLMs. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  70. [70]

    Zhang, B

    J. Zhang, B. Guo, Y . Jiang, H. Wang, B. An, and X. Du. Harnessing reasoning trajectories for hallucination detection via answer-agreement representation shaping.International Conference on Machine Learning, 2026

    2026

  71. [71]

    Zhang, Y

    Z. Zhang, Y . Sheng, T. Zhou, T. Chen, L. Zheng, R. Cai, Z. Song, Y . Tian, C. Ré, C. Barrett, et al. H2O: Heavy-hitter oracle for efficient generative inference of large language models.Advances in Neural Information Processing Systems, 36:34661–34710, 2023

    2023

  72. [72]

    Zheng, W.-L

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. Xing, et al. Judging LLM-as-a-judge with MT-Bench and Chatbot Arena.Advances in neural information processing systems, 36:46595–46623, 2023

    2023

  73. [73]

    You are a factual question answering expert. Provide one concise and direct final answer

    H. Zubkova, J.-H. Park, and S.-W. Lee. SUGAR: leveraging contextual confidence for smarter retrieval. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2025. 16 Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models A. Additional experimental...

    arXiv 2025

  74. [74]

    This supports our default choice of computing the step score from the last answer token at the final transformer layer

    Layers 27–28 show a mild drop to around 84.5%, after which performance recovers and peaks at the final layer (86.44%). This supports our default choice of computing the step score from the last answer token at the final transformer layer. 27 Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models Table 10| Effe...

  75. [77]

    REDE(proxy)

    Area = 6 √ 21. S29 Alt verify: 16A2 = 12096. A2 = 756. ✓ S30 6 √ 21 ≈ 27.5. Reasonable. S31 S = 4× 6 √ 21 = 24 √ 21. S32 Now compute V via Cayley–Menger. S33 288V 2 = |CM|, 5×5 det. S34 CM with AB2 = 41, . . . , CD2 = 41. S35 W rite out full matrix. S36 Use row reduction. S37 Row ops: R′ 2 , R′ 3 , R′ 4 . S38 V erifyR′ 2 . ✓ S39 V erifyR′ 3 . ✓ S40 V erif...

    arXiv 2024

  76. [89]

    Find m + n + p

    Point I inside equidistant to all faces at distancem√n p . Find m + n + p. GT: 104 Pred: 102 ✗ S1 Equidistant point = incenter, distance = inradiusr. S2 In a triangle, incenter equidistant from sides. S3 2D: r = A/s for area A, semiperimeter s. S4 Incenter splits tet into 4 sub-tetrahedra. S5 Each has base = face, height =r. S6 V = 1 3 Bh. Basic formula. ...

  77. [445]

    S25 sin2 C = 189/445

    ✓ S24 cos2 C = 256/445 < 1, acute. S25 sin2 C = 189/445. S26 Area = 2 √ 189. S27 √ 445 cancels in product. S28 √ 189 = 3 √