Pith. sign in

REVIEW 3 major objections 4 minor 60 references

Showing a language model the gold answer while it writes its reasoning creates training data that correctness filters cannot save.

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

2026-08-02 01:45 UTC pith:MNBUO6H3

load-bearing objection Answer-conditioning hurts distilled reasoning, but the headline magnitude is partly entangled with an attempt-index confound; still worth a serious look. the 3 major comments →

arxiv 2607.14552 v1 pith:MNBUO6H3 submitted 2026-07-16 cs.CL cs.AI

Answer-Conditioned Chains of Thought Degrade Verifiable-Reasoning Distillation in Large Language Models

Jungseob Lee , Seungyoon Lee , Suhyune Son , Dongyub Jude Lee , Sungbin Han , Sugyeong Eo , Heuiseok Lim This is my paper
classification cs.CL cs.AI
keywords chain-of-thought distillationanswer conditioningverifiable reasoningrationalizationdata qualityself-improvementanswer leakagesupervised fine-tuning
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.

To distill reasoning ability, labs sample reasoning traces from a model, keep the ones that reach the right answer, and fine-tune on those survivors. Many pipelines try to increase yield by showing the model the gold answer and asking it to reason toward it. This paper argues that this answer-conditioning step quietly poisons the training data, and that a correctness filter cannot catch the damage because every kept chain is correct by construction. In a controlled experiment that changes only whether the generator sees the answer, training on answer-conditioned chains costs 16.2 points on the MATH-500 benchmark, with the loss growing with difficulty to about 27 points on olympiad problems. The mechanism is visible in the chains themselves: they state the answer early and rationalize backward, and the habit transfers to the student; the practical fix is to generate answer-blind, or to ask for the derivation first and the answer last.

Core claim

The central claim is that a correct final answer does not make a model-generated chain of thought a safe distillation target, because the correctness filter cannot see how the chain was produced. The paper's 'one-bit experiment' holds the generator, problem set, correctness filter, and fine-tuning recipe fixed and varies only whether generation is answer-conditioned—the gold answer shown with a request to reach it. That single bit costs 16.2 MATH-500 points on a strong reasoning model, with the penalty rising with difficulty (about 27 points on AIME) and extending to code where every kept chain passes its tests. The damage is a property of the data, not the generator: it transfers across tea

What carries the argument

The load-bearing instrument is the one-bit experiment: two otherwise identical generation arms, answer-blind and answer-leaked, sharing the same generator, problems, final-answer filter, and supervised fine-tuning recipe. The measurable mechanism is the answer-first rate (AFR), the fraction of chains that state the gold answer within the first 20% of the reasoning block; leaked chains have sharply higher AFR, and a within-corpus split into answer-first and derivation-first chains shows that the answer-first chains carry most of the accuracy loss (a 16.9-point difference-in-differences against blind controls). A length-matched control—truncating each blind chain to its leaked twin's length—ru

Load-bearing premise

The attribution of the penalty to answer-conditioning assumes the blind and leaked corpora are exchangeable except for the single intervention; in particular, the leaked chains are systematically shorter, and while length-matching controls for length, it does not control for other textual differences or for the possibility that the two arms differ in which problems get solved.

What would settle it

Run the one-bit experiment on a new family of thinking models with a randomized first-correct-chain sampling rule and exactly length-matched chains (e.g., by truncation or by generating leaked chains with a prompt that avoids shortening). If the accuracy gap between answer-blind and answer-leaked arms collapses to zero, or if the pre-training answer-first signature no longer orders the penalties across models, the central claim fails.

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

If this is right

  • Answer-blind self-distillation preserves base accuracy (within seed noise) and modestly beats training on gold data at a matched budget, so pipelines should generate chains without showing answers.
  • Correctness filtering cannot validate chains generated under answer-conditioning; data provenance—how a chain was produced—must be audited separately.
  • When showing the answer is unavoidable, a 'derive first, then state the answer' instruction recovers about two-thirds of the lost accuracy while keeping every example.
  • Candidate teachers can be screened before any fine-tuning by measuring the answer-first signature on a small unlabeled sample, predicting held-out penalties within a few points.
  • The penalty appears in domains that require derivation (math, code) and disappears on multiple-choice knowledge tasks, so benchmark choice determines whether the failure is visible.

Where Pith is reading between the lines

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

  • Editorial inference: the same early-answer rationalization habit may contaminate reinforcement-learning rollouts that are conditioned on a correct answer, so verifiable-reasoning RL pipelines should test whether answer-conditioned rollouts introduce the same off-policy bias.
  • Editorial inference: because the answer-first signature is a cheap, label-free proxy, it could be used as a real-time data-quality gate during generation, rejecting or discounting chains that state the answer in their first 20% before training begins.
  • Editorial inference: the paper's train/test length inversion (short rationalized chains at training time, long rambling outputs at test time) suggests that answer-conditioned fine-tuning teaches a compressed answer-first policy that fails by non-convergence; if so, inference-time length capping would mask rather than fix the underlying deficit.
  • Editorial inference: the blind twins of early-answer problems are themselves answer-first 48.9% of the time, so problem difficulty partly drives rationalization; an extension would test whether the penalty scales with the prevalence of such 'easy-to-rationalize' problems in a training corpus.

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 / 4 minor

Summary. The paper studies whether fine-tuning on an LLM's own answer-conditioned chains of thought (gold answer shown, with a request to reach it) is worse than fine-tuning on answer-blind chains, when both are passed through the same correctness filter. On 935 matched math problems with Qwen3-8B as generator and student, the leaked corpus costs 16.2 MATH-500 points relative to blind, growing to 27.2 points on AIME. The paper attributes the harm to chains that state the answer early ('rationalized' chains), supports this with a prompt ablation, a within-corpus difference-in-differences carrier experiment, a chain-excision control, and a pre-fine-tuning answer-first-rate signature that orders the penalty across eight thinking models from four families. It concludes that generating answer-blind is the safe choice and that no correctness filter can detect the damage in the data.

Significance. If the central estimate is valid, the finding is practically important: many self-improvement and distillation pipelines use answer-conditioned rationales, and this paper would show that such data can be negative-value despite passing the correctness filter. The paper has real strengths: matched problem sets, identical SFT recipe, three seeds for headline runs, length-matched controls, a difference-in-differences carrier design, held-out penalty predictions, and a released code/data package. The proposed answer-first signature is a cheap, falsifiable predictor. However, the headline causal isolation is weakened by an attempt-index confound in the sampling rule, and the text overclaims the universality of the sign of the effect. These issues are fixable within the manuscript's scope, but they need to be addressed before the 16.2-point attribution can be taken at face value.

major comments (3)
  1. [Section 3, Appendix G] The one-bit comparison as implemented confounds answer-conditioning with attempt index. The paper samples n=2 chains per problem and keeps the first correct chain. Under the leaked prompt, first-sample success is much more likely, so the leaked corpus is dominated by first-attempt chains, while the blind corpus contains a substantial share of second-attempt chains. The statement in Section 3 that the arms 'differ only in answer-conditioning and in the chain-length shift' omits this. Length matching does not equate attempt index, and the within-corpus controls (Sections 4.3, 6) operate inside the leaked corpus where attempt composition is nearly constant. The paper should report the per-arm attempt-index distribution and provide a control with matched attempt index (e.g., first-attempt correct chains only, or a randomized-order scheme), or argue explicitly why second-attempt chains are no
  2. [Section 6, Appendix H] The sentence in Section 6 that the penalty 'holds with the same sign across every generative setting' is contradicted by the cross-student transfer numbers in Appendix H: the three seed penalties for Qwen3-1.7B trained on Qwen3-8B corpora are 16.8, -0.2, and 4.0. One seed is negative, and the standard deviation is 8.9. The cross-student row in Table 5 is therefore a positive mean with a highly unstable sign, not a uniformly signed effect. The claim should be qualified to average sign, or the seed-level test should be reported. This is load-bearing because the 'damage travels with the data, not the student' conclusion currently rests on this row.
  3. [Section 4.3 / Section 6, excision experiment] The excision experiment shows that within the leaked corpus, the early-answer line itself carries most of the harm. This is an informative mechanism check, but it does not by itself rule out the attempt-index confound for the blind-vs-leaked contrast, because the excision operates entirely on chains that were all generated under the same prompt. The paper should not present the carrier and excision results as sufficient to identify the full causal quantity in Eq. (1). They support the qualitative conclusion and the rationalization mechanism, but the magnitude of the headline penalty still depends on the uncontrolled selection differences described above.
minor comments (4)
  1. [Table 1] The text refers to a 'tinted Blind column' and tinted rows, but the table as printed appears to have no visible tint. Please ensure the final PDF retains the intended shading or replace the color-dependent language with explicit row/column labels.
  2. [Appendix F, Table 9] The single-seed DiD value '+21.2' in Section 4.3 does not obviously match any cell in Table 9 (seed 42 gives +22.0, seed 123 +13.2, seed 7777 +18.0). Please clarify which seed or configuration produces +21.2.
  3. [Table 1(A)] The 'Out-of-band' arm is not defined in the main text or the appendix. Please give the exact prompt and state whether it was generated with a numerically equivalent memory optimization, as described for the two neutral arms in Appendix J.
  4. [Section 5 / Table 14] The cross-model correlation is reported as r=0.96 on n=8 with a wide Fisher-z interval. The paper appropriately caveats this, but the abstract and conclusion phrase it as a strong predictor. Consider stating the sample size and interval in those places so the small-sample nature is visible to a casual reader.

Circularity Check

0 steps flagged

No significant circularity: the leakage penalty is measured independently of the AFR signature, and the held-out predictions are genuine out-of-sample forecasts.

full rationale

The paper's central quantity, the leakage penalty Δ = acc(S_blind) − acc(S_leaked) (Eq. 1), is an independently measured fine-tuning outcome. It is not defined in terms of the answer-first-rate (AFR) signature, nor is AFR fitted to the outcome; instead, AFR is read off unlabeled generations before fine-tuning and then correlated with the later-measured penalty. The carrier experiment does reuse the same 0.2 threshold to define both the mechanism and the corpus split, but the endpoint there is the students' post-SFT MATH-500 accuracy, so the finding that Leaked-Early chains carry more harm is empirical rather than true by construction. The held-out penalty predictions in Table 4 were fixed from fits excluding those models and then checked against later observations, so they are not fitted inputs called predictions. No load-bearing self-citation or imported uniqueness theorem appears; the paper cites prior work for context, not to force its conclusion. The first-correct-chain sampling confound noted by skeptics is a potential threat to causal attribution, but it is a confound, not circularity: the arms are not equivalent by definition, and the penalty could in principle have gone either way. The stated limitations (small sample of eight models, some single-seed rows, non-length-controlled transfer in one appendix) further support reading the results as honest empirical findings rather than re-labeled assumptions. Overall, the derivation chain is self-contained and the main claims do not reduce to their inputs.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The central claim is empirical, so the ledger is dominated by design assumptions and the hand-chosen AFR threshold. There are no invented physical entities. The signature-penalty regression is a fitted predictor, but it is not used to define the penalty itself.

free parameters (3)
  • Answer-first rate threshold = 0.2
    Hand-chosen early-answer window defining the answer-first signature and carrier split; robustness swept over {0.1, 0.2, 0.3} in Appendix F, but the central mechanism claim depends on this threshold.
  • Evaluation token budget = 32768
    Fixed budget for MATH-500 evaluation; Appendix E budget sweep shows the penalty shape is robust, but the absolute penalty magnitude depends on this chosen budget.
  • Signature-penalty regression coefficients = not stated precisely (linear fit in Fig. 4)
    Linear fit of penalty on ΔAFR across eight models; used for held-out predictions in Table 4. The fit is not needed to define the penalty but is load-bearing for the prediction claim.
axioms (6)
  • domain assumption Symbolic final-answer matching is a sufficient correctness filter for keeping chains in the standard distillation recipe.
    Both arms of the one-bit experiment pass this filter; the paper's point is that this filter is insufficient, but the experiment presupposes it can identify correct final answers.
  • domain assumption SFT on the generator's own answer-blind chains approximates on-policy training from P(c|q), and targets from P(c|q,a) are off-policy and harmful.
    Section 4.2 invokes this off-policy framing to explain the mechanism; it is not derived from first principles.
  • domain assumption MATH-500, GSM8K, Minerva, AIME, MBPP+, and HumanEval+ accuracy under the vLLM thinking-mode harness is a valid measure of verifiable reasoning.
    Evaluation harness details are in Section 3 and Appendix G; all conclusions about verifiable-reasoning distillation rest on these benchmarks.
  • domain assumption The first-correct-chain-per-problem sampling with n=2 at temperature 0.6 gives representative chains for each condition.
    Generation details in Appendix G; if sampling is not representative, the matched corpora could be biased in unmeasured ways.
  • domain assumption The 935 training problems are disjoint from evaluation sets and drawn from the Nemotron-Cascade corpus.
    Stated in Section 3; necessary for the generalization claims to unseen problems.
  • ad hoc to paper The blind and leaked prompts differ only in answer visibility and the rationalization instruction, with no hidden confound in prompt wording.
    This is the core causal-identification assumption of the one-bit experiment; the prompt ablation addresses instruction wording but cannot fully rule out other wording effects.

pith-pipeline@v1.3.0-alltime-deepseek · 19392 in / 15229 out tokens · 166485 ms · 2026-08-02T01:45:44.577574+00:00 · methodology

0 comments
read the original abstract

A standard recipe for distilling the reasoning ability of large language models (LLMs) is to sample chains of thought from the model, keep those that reach the correct final answer, and fine-tune on the survivors. When sampling fails, a common fix shows the generator the gold answer and asks it to write a chain that reaches that answer. We show that this second step degrades the training data in a way that correctness filtering cannot catch. We run a controlled experiment that fixes the generator, the problem set, and the correctness filter, and varies only whether the chain is generated under answer-conditioning, the gold answer shown with a request to reach it. Training a strong instruction-tuned reasoning model on its own answer-conditioned chains sharply lowers its verifiable-reasoning accuracy. The loss grows with difficulty, reaching as much as about 27 points on the hardest competition problems. The mechanism is legible in the chains themselves, which rationalize backward from the shown answer instead of deriving it, with the early final-answer statement as the measurable symptom. The harm is a property of the data rather than the generator, read off unlabeled generations before any fine-tuning, ordering the penalty across eight thinking models from four families, and transferring across teacher families. A prompt ablation localizes it to the rationalize-toward instruction rather than the answer's bare visibility. The practical takeaway is to generate answer-blind, because no correctness filter can see this damage in the data.

Figures

Figures reproduced from arXiv: 2607.14552 by Dongyub Jude Lee, Heuiseok Lim, Jungseob Lee, Seungyoon Lee, Sugyeong Eo, Suhyune Son, Sungbin Han.

Figure 1
Figure 1. Figure 1: The one-bit experiment. A shared generator produces chains for the same problems under an answer-blind path and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Answer-position and length behavior of the answer-blind and answer-leaked arms in training chains and student [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Leakage penalty against the evaluation token bud [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The ∆AFR signature against the answer-leakage penalty across eight thinking models from four families. The dashed line is the linear fit and the shaded band its dispersion. Three points are out-of-sample, and labels are colored by in￾sample against held-out. Error bars show sampling variability on both axes. reported in Appendix I. The answer-first signature survives the capability and family confounds. Ap… view at source ↗

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

60 extracted references · 1 canonical work pages

  1. [1]

    Advances in Neural Information Processing Systems , volume =

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author =. Advances in Neural Information Processing Systems , volume =. 2022 , publisher =

    2022

  2. [2]

    International Conference on Learning Representations , year =

    Self-Consistency Improves Chain of Thought Reasoning in Language Models , author =. International Conference on Learning Representations , year =

  3. [3]

    2024 , publisher =

    Wang, Peiyi and Li, Lei and Shao, Zhihong and Xu, Runxin and Dai, Damai and Li, Yifei and Chen, Deli and Wu, Yu and Sui, Zhifang , booktitle =. 2024 , publisher =

    2024

  4. [4]

    2022 , publisher =

    Zelikman, Eric and Wu, Yuhuai and Mu, Jesse and Goodman, Noah D , booktitle =. 2022 , publisher =

    2022

  5. [5]

    International Conference on Machine Learning , year =

    The Wisdom of Hindsight Makes Language Models Better Instruction Followers , author =. International Conference on Machine Learning , year =

  6. [6]

    Measuring Mathematical Problem Solving With the

    Hendrycks, Dan and Burns, Collin and Kadavath, Saurav and Arora, Akul and Basart, Steven and Tang, Eric and Song, Dawn and Steinhardt, Jacob , booktitle =. Measuring Mathematical Problem Solving With the

  7. [7]

    arXiv preprint arXiv:2009.03300 , year =

    Measuring Massive Multitask Language Understanding , author =. arXiv preprint arXiv:2009.03300 , year =

    Pith/arXiv arXiv 2009

  8. [8]

    Rein, David and Hou, Betty Li and Stickland, Asa Cooper and Petty, Jackson and Pang, Richard Yuanzhe and Dirani, Julien and Michael, Julian and Bowman, Samuel R , journal =

  9. [9]

    Think you have Solved Question Answering? Try

    Clark, Peter and Cowhey, Isaac and Etzioni, Oren and Khot, Tushar and Sabharwal, Ashish and Schoenick, Carissa and Tafjord, Oyvind , journal =. Think you have Solved Question Answering? Try

  10. [10]

    Zellers, Rowan and Holtzman, Ari and Bisk, Yonatan and Farhadi, Ali and Choi, Yejin , booktitle =

  11. [11]

    Sakaguchi, Keisuke and Le Bras, Ronan and Bhagavatula, Chandra and Choi, Yejin , booktitle =

  12. [12]

    2020 , organization =

    Rajbhandari, Samyam and Rasley, Jeff and Ruwase, Olatunji and He, Yuxiong , booktitle =. 2020 , organization =

    2020

  13. [13]

    Communications of the ACM , volume =

    Algorithm = Logic + Control , author =. Communications of the ACM , volume =. 1979 , publisher =

    1979

  14. [14]

    arXiv preprint arXiv:2505.09388 , year =

    Pith/arXiv arXiv

  15. [15]

    arXiv preprint arXiv:2406.11704 , year =

    Nemotron-4 340B Technical Report , author =. arXiv preprint arXiv:2406.11704 , year =

    Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2407.21787 , year =

    Large Language Monkeys: Scaling Inference Compute with Repeated Sampling , author =. arXiv preprint arXiv:2407.21787 , year =

    Pith/arXiv arXiv

  17. [17]

    Snell, Charlie and Lee, Jaehoon and Xu, Kelvin and Kumar, Aviral , journal =. Scaling

  18. [18]

    2024 , note =

    Hosseini, Arian and Yuan, Xingdi and Malkin, Nikolay and Courville, Aaron and Sordoni, Alessandro and Agarwal, Rishabh , booktitle =. 2024 , note =

    2024

  19. [19]

    Zelikman, Eric and Harik, Georges and Shao, Yijia and Jayasiri, Varuna and Haber, Nick and Goodman, Noah D , journal =. Quiet-

  20. [20]

    arXiv preprint arXiv:2308.01825 , year=

    Scaling Relationship on Learning Mathematical Reasoning with Large Language Models , author=. arXiv preprint arXiv:2308.01825 , year=

    Pith/arXiv arXiv

  21. [21]

    arXiv preprint arXiv:2308.08998 , year=

    Reinforced Self-Training (ReST) for Language Modeling , author=. arXiv preprint arXiv:2308.08998 , year=

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2312.06585 , year=

    Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models , author=. arXiv preprint arXiv:2312.06585 , year=

    Pith/arXiv arXiv

  23. [23]

    Findings of EMNLP , year=

    Large Language Models are Better Reasoners with Self-Verification , author=. Findings of EMNLP , year=

  24. [24]

    arXiv preprint arXiv:2411.19865 , year=

    Reverse Thinking Makes LLMs Stronger Reasoners , author=. arXiv preprint arXiv:2411.19865 , year=

    Pith/arXiv arXiv

  25. [25]

    NeurIPS , year=

    LIMA: Less Is More for Alignment , author=. NeurIPS , year=

  26. [26]

    NeurIPS , year=

    Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting , author=. NeurIPS , year=

  27. [27]

    arXiv preprint arXiv:2307.13702 , year=

    Measuring Faithfulness in Chain-of-Thought Reasoning , author=. arXiv preprint arXiv:2307.13702 , year=

    Pith/arXiv arXiv

  28. [28]

    International Conference on Learning Representations (ICLR) , year =

    On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes , author =. International Conference on Learning Representations (ICLR) , year =

  29. [29]

    arXiv preprint arXiv:2305.20050 , year =

    Let's Verify Step by Step , author =. arXiv preprint arXiv:2305.20050 , year =

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2110.14168 , year =

    Training Verifiers to Solve Math Word Problems , author =. arXiv preprint arXiv:2110.14168 , year =

    Pith/arXiv arXiv

  31. [31]

    arXiv preprint arXiv:2501.12948 , year =

    Pith/arXiv arXiv

  32. [32]

    arXiv preprint arXiv:2406.06592 , year =

    Improve Mathematical Reasoning in Language Models by Automated Process Supervision , author =. arXiv preprint arXiv:2406.06592 , year =

    Pith/arXiv arXiv

  33. [33]

    Kazemi, Mehran and Kim, Najoung and Bhatia, Deepti and Xu, Xin and Ramachandran, Deepak , booktitle =

  34. [34]

    and Zhang, Hao and Stoica, Ion , booktitle =

    Kwon, Woosuk and Li, Zhuohan and Zhuang, Siyuan and Sheng, Ying and Zheng, Lianmin and Yu, Cody Hao and Gonzalez, Joseph E. and Zhang, Hao and Stoica, Ion , booktitle =. Efficient Memory Management for Large Language Model Serving with

  35. [35]

    2024 , note =

    Yu, Longhui and Jiang, Weisen and Shi, Han and Yu, Jincheng and Liu, Zhengying and Zhang, Yu and Kwok, James T and Li, Zhenguo and Weller, Adrian and Liu, Weiyang , booktitle =. 2024 , note =

    2024

  36. [36]

    arXiv preprint arXiv:2410.18982 , year =

    O1 Replication Journey: A Strategic Progress Report -- Part 1 , author =. arXiv preprint arXiv:2410.18982 , year =

    Pith/arXiv arXiv

  37. [37]

    Advances in Neural Information Processing Systems , volume=

    Training chain-of-thought via latent-variable inference , author=. Advances in Neural Information Processing Systems , volume=

  38. [38]

    arXiv preprint arXiv:2211.14275 , year =

    Solving math word problems with process- and outcome-based feedback , author =. arXiv preprint arXiv:2211.14275 , year =

    Pith/arXiv arXiv

  39. [39]

    arXiv preprint arXiv:2505.05410 , year =

    Reasoning Models Don't Always Say What They Think , author =. arXiv preprint arXiv:2505.05410 , year =

    Pith/arXiv arXiv

  40. [40]

    Zhang, Wenjing and Yan, Jiangze and Huang, Jieyun and Shen, Yi and Shi, Shuming and Chen, Ping and Wang, Ning and Liu, Zhaoxiang and Wang, Kai and Lian, Shiguo , journal =

  41. [41]

    doi:10.52202/079017-2066 , journal =

    Dan Zhang and Sining Zhoubian and Ziniu Hu and Yisong Yue and Yuxiao Dong and Jie Tang , year =. doi:10.52202/079017-2066 , journal =

    work page doi:10.52202/079017-2066 2066

  42. [42]

    doi:10.18653/v1/2025.acl-long.50 , journal =

    Chujie Zheng and Zhenru Zhang and Beichen Zhang and Runji Lin and Keming Lu and Bowen Yu and Dayiheng Liu and Jingren Zhou and Junyang Lin , year =. doi:10.18653/v1/2025.acl-long.50 , journal =

    doi:10.18653/v1/2025.acl-long.50 2025

  43. [43]

    doi:10.18653/v1/2025.acl-long.1230 , journal =

    Mingyang Song and Zhaochen Su and Xiaoye Qu and Jiawei Zhou and Yu Cheng , year =. doi:10.18653/v1/2025.acl-long.1230 , journal =

    doi:10.18653/v1/2025.acl-long.1230 2025

  44. [44]

    2024 , doi =

    Li, Chengpeng and Yuan, Zheng and Yuan, Hongyi and Dong, Guanting and Lu, Keming and Wu, Jiancan and Tan, Chuanqi and Wang, Xiang and Zhou, Chang , booktitle =. 2024 , doi =

    2024

  45. [45]

    2025 , doi =

    Li, Chengpeng and Xue, Mingfeng and Zhang, Zhenru and Yang, Jiaxi and Zhang, Beichen and Yu, Bowen and Hui, Binyuan and Lin, Junyang and Wang, Xiang and Liu, Dayiheng , booktitle =. 2025 , doi =

    2025

  46. [46]

    2025 , doi =

    Xiong, Feng and Xu, Hongling and Wang, Yifei and Cheng, Runxi and Wang, Yong and Chu, Xiangxiang , booktitle =. 2025 , doi =

    2025

  47. [47]

    2026 , doi =

    Li, Guozheng and Zhang, Xinyu , booktitle =. 2026 , doi =

    2026

  48. [48]

    2024 , doi =

    Luong, Trung Quoc and Zhang, Xinbo and Jie, Zhanming and Sun, Peng and Jin, Xiaoran and Li, Hang , booktitle =. 2024 , doi =

    2024

  49. [49]

    arXiv preprint arXiv:2509.05602 , year =

    Mitigating Spurious Correlations Between Question and Answer via Chain-of-Thought Correctness Perception Distillation , author =. arXiv preprint arXiv:2509.05602 , year =

    Pith/arXiv arXiv

  50. [50]

    Findings of the Association for Computational Linguistics: ACL 2025 , year =

    The Lessons of Developing Process Reward Models in Mathematical Reasoning , author =. Findings of the Association for Computational Linguistics: ACL 2025 , year =

    2025

  51. [51]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

    Outcome Accuracy is Not Enough: Aligning the Reasoning Process of Reward Models , author =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

  52. [52]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , year =

    Making Reasoning Matter: Measuring and Improving Faithfulness of Chain-of-Thought Reasoning , author =. Findings of the Association for Computational Linguistics: EMNLP 2024 , year =

    2024

  53. [53]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , year =

    Measuring Chain of Thought Faithfulness by Unlearning Reasoning Steps , author =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , year =

    2025

  54. [54]

    Findings of the Association for Computational Linguistics: ACL 2024 , year =

    Preemptive Answer ``Attacks'' on Chain-of-Thought Reasoning , author =. Findings of the Association for Computational Linguistics: ACL 2024 , year =

    2024

  55. [55]

    arXiv preprint arXiv:2606.00628 , year =

    Robust Reasoning via Dynamic Token Selection for Distribution-Aligned Self-Distillation , author =. arXiv preprint arXiv:2606.00628 , year =

    Pith/arXiv arXiv

  56. [56]

    2026 , doi =

    Li, Yuduo and Shi, Xiaofeng and Kou, Qian and Yu, Longbin and Zhou, Hua , journal =. 2026 , doi =

    2026

  57. [57]

    In-Token Rationality Optimization: Towards Accurate and Concise

    Zhu, Mingye and Liu, Yi and Fu, Zheren and Wang, Quan and Zhang, Yongdong , booktitle =. In-Token Rationality Optimization: Towards Accurate and Concise. 2026 , doi =

    2026

  58. [58]

    2025 , doi =

    Wang, Boxin and Lee, Chankyu and Lee, Nayeon and Lin, Sheng-Chieh and Dai, Wenliang and Chen, Yang and Yang, Zhuoling and Liu, Zihan and Shoeybi, Mohammad and Catanzaro, Bryan and Ping, Wei , journal =. 2025 , doi =

    2025

  59. [59]

    2025 , doi =

    Bercovich, Akhiad and Levy, Itay and Golan, Izik and Dabbah, Mohammad and El-Yaniv, Ran and Puny, Omri and others , journal =. 2025 , doi =

    2025

  60. [60]

    arXiv preprint arXiv:2603.05433 , year=

    Crisp: Compressed reasoning via iterative self-policy distillation , author=. arXiv preprint arXiv:2603.05433 , year=

    Pith/arXiv arXiv