Pith. sign in

REVIEW 2 major objections 4 minor 39 references

CASE: training-time causal alignment and inference-time attention masking push LLM answers to depend on the generated reasoning chain rather than on a direct instruction-to-answer shortcut.

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 14:12 UTC pith:XBPFRQ5Q

load-bearing objection Useful method, strong experiments, but the theory oversells and part of the evaluation is trained into the model. the 2 major comments →

arxiv 2607.18820 v1 pith:XBPFRQ5Q submitted 2026-07-21 cs.CL

CASE: Causal Alignment and Structural Enforcement for Improving Chain-of-Thought Faithfulness

Ziming Wang , Yinghua Yao , Changwu Huang , Ke Tang , Xin Yao This is my paper
classification cs.CL
keywords chain-of-thought faithfulnesscausal alignmentattention maskingcounterfactual reasoningselective-loss fine-tuninginformation-theoretic analysisLLM reasoningshortcut suppression
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.

The paper argues that LLM answers often diverge from their chain-of-thought reasoning because the instruction can shortcut directly to the answer, bypassing the CoT. It proposes CASE, which trains the model with counterfactual-CoT, biased-instruction, and empty-instruction datasets to strengthen the CoT-to-answer dependency, and at inference time masks all direct attention from answer tokens to instruction tokens. The central claim is that this two-stage intervention pushes the model toward the faithful causal chain instruction→CoT→answer. Experiments across three models and four benchmarks report a 37% average per-setting relative improvement in overall CoT faithfulness over the strongest baselines, stronger cross-dataset faithfulness transfer, and comparable average accuracy.

Core claim

CASE improves CoT faithfulness by strengthening the reasoning-to-answer path X→Y and suppressing the direct instruction-to-answer shortcut Z→Y. During training it constructs counterfactual-CoT, biased-instruction, and empty-instruction datasets and applies selective-loss fine-tuning: full supervision on the standard dataset, and answer-token-only losses on the auxiliary datasets. During inference it sets attention logits from instruction tokens to answer tokens to −∞, so answer tokens can attend only to the CoT and previously generated answer tokens. The paper shows that these training losses maximize lower bounds on I(X;Y) and I(X;Y|Z) and minimize an upper bound on I(Y;Z|X), and that the m

What carries the argument

The central object is the inference-time attention mask: for every answer token, all pre-softmax attention logits to instruction-token positions are set to −∞, making those attention weights zero so answer tokens cannot directly read instruction-token states. The accompanying training-side machinery is selective-loss SFT over four datasets: the standard instruction–CoT–answer set (full loss), a counterfactual-CoT set (answer loss to make answers respond to CoT changes), a biased-instruction set (answer loss to keep answers stable under misleading instruction cues), and an empty-instruction set (answer loss to make the CoT alone sufficient). Together they instantiate three information-theoret

Load-bearing premise

The inference-time mask only blocks direct token-level attention from answer to instruction; the central claim relies on the assumption that the instruction's influence cannot still reach the answer through the hidden states that generate the CoT (H_X), which the textual-level independence Y⊥Z|X does not guarantee.

What would settle it

Generate two instructions that are ambiguous but carry the same misleading retrieval cue (e.g., both mention an authority endorsing the same wrong answer), fix the same reasoning chain, and measure CDE under CASE. If the answer flips substantially despite the mask, the shortcut persists in hidden states. Alternatively, train a linear probe on H_X to decode whether the misleading cue is present; if decoding accuracy stays high after CASE, token-level masking has not enforced textual-level faithfulness.

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

If this is right

  • Faithfulness gains are not confined to the training distribution: in six cross-dataset settings CASE achieves the best overall faithfulness score, with a 61% average relative improvement over the strongest baseline per setting.
  • Training-time and inference-time components are complementary: training alone improves CIE and CDE, masking alone improves Filler Tokens and Early Answering, and the combination produces the highest overall score.
  • The inference-time mask preserves accuracy better than simply removing the instruction before answer generation (CoT-only decoding), because it retains the original instruction-conditioned CoT-side states.
  • The framework is a single-model, plug-in intervention that uses LoRA fine-tuning and an attention-mask change, with inference runtime comparable to other single-model methods and about 3× faster than a dual-model pipeline.
  • The selective-loss design prevents the model from being forced to generate counterfactual CoTs while still teaching how a given CoT determines the answer.

Where Pith is reading between the lines

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

  • The paper's Proposition 4 conditions on CoT-side hidden states H_X and concedes that instruction information retained there remains CoT-mediated; a natural next test is whether erasing instruction-specific features from H_X (e.g., representation steering) yields further faithfulness gains or reveals residual shortcut information that token-level masking cannot remove.
  • The data construction relies on an external strong LLM to write counterfactual CoTs and biased instructions; the paper audits consistency, but the method's ceiling is likely tied to how much causal variation that generator can produce — a testable extension is to measure CASE's faithfulness gain as a function of generator strength.
  • The faithfulness metrics used are perturbation-based; a more direct test of the causal claim would be to measure the actual intervention effect on hidden states, e.g., whether a probe trained to decode instruction cues from H_X fails after CASE training.
  • The selective-loss scheme supervises only answer tokens on the auxiliary sets; jointly training on counterfactual CoT tokens with a mixing schedule could trade some faithfulness for reasoning diversity, which the paper does not explore.

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

2 major / 4 minor

Summary. This paper proposes CASE, a framework for improving chain-of-thought (CoT) faithfulness by combining training-time causal alignment and inference-time structural enforcement. Training constructs counterfactual-CoT, biased-instruction, and empty-instruction datasets, and applies selective-loss SFT to strengthen the CoT-to-answer dependence while suppressing instruction shortcuts. Inference applies an attention mask that prevents answer tokens from attending directly to instruction tokens. The paper provides four information-theoretic propositions intended to show that these components promote the faithful chain Z→X→Y, and reports experiments on Llama-3.1-8B, Qwen3-8B, and DeepSeek-R1-Distill-Qwen-7B across ARC-Easy, ARC-Challenge, StrategyQA, and LogiQA. The empirical tables show consistent G-mean faithfulness improvements, stronger cross-dataset transfer, and competitive accuracy compared with Base LLM, DU_c, FRODO, and FoCus.

Significance. If the central claim is established, CASE would be a practical and useful method for improving CoT faithfulness in a single LLM, with clear engineering value: it is faster than dual-model baselines, is accompanied by released code, and includes quality audits of generated training data. The cross-dataset transfer results are a genuine strength, and the ablation study reasonably attributes different effects to training-time alignment and inference-time masking. The theoretical propositions are standard variational bounds and a computation-graph statement; their contribution is modest but coherent. However, the headline empirical claim is weakened by the near-duplication between training objectives and two of the four evaluation metrics, and the theory does not bridge from hidden-state independence to the textual-level faithfulness condition in Definition 1. These issues are load-bearing and need to be addressed before the faithfulness claim can be accepted at face value.

major comments (2)
  1. [Methodology, Eqs. (6)–(7) vs. Appendix A.4 (CIE/CDE)] The two most controlled evaluation metrics are directly trained by the proposed objective. L_X-int (Eq. 6) maximizes log pθ(y'_i|z_i,x'_i) on counterfactual CoT pairs; evaluation CIE generates a counterfactual CoT and measures whether the answer changes. L_Z-int (Eq. 7) trains pθ(y_i|z'_i,x_i) with a biased instruction; evaluation CDE generates a counterfactual question and measures answer invariance. Both use counterfactual data generated by DeepSeek-V3.2 with prompts very similar to those used in training-data construction (Tables S-XIII and S-XIV). Thus the reported CIE and CDE gains are at least partly a consequence of fitting the exact intervention distribution used at evaluation, not evidence of a general X→Y mechanism. Because the G-mean aggregates these metrics, the 37% relative improvement is inflated. Please report FT/EA as the primary metric set, or evaluate with held-out inte
  2. [Theoretical Analysis, Proposition 4 and Problem Formulation, Definition 1] Definition 1 requires pθ(y|z,x)=pθ(y|x) for the visible textual CoT X. Proposition 4 proves only p^AM_θ(y|z,h_X)=p^AM_θ(y|h_X), where H_X is the CoT-side hidden state formed under Z. Since H_X is a function of both Z and X, p^AM_θ(y|z,x)=E_{h_X|z,x}[p^AM_θ(y|h_X)] can still depend on z when x is fixed. The manuscript itself concedes that 'instruction information retained in H_X remains CoT-mediated' (Problem Formulation). Thus the theory establishes removal of the direct token-attention path, not the textual-level conditional independence Y⊥Z|X. The CDE experiments do not close this gap: they use a single greedy decode and counterfactual questions generated by the same model family. A concrete test would be probing h_X for instruction information under fixed x, or evaluating faithfulness with interventions that vary the instruction-dependent component of h_X while keeping the textual CoT
minor comments (4)
  1. [Abstract and Contributions] The abstract reports '37% average per-setting relative improvement' while the contributions section says 'more than 37% average improvement.' Please harmonize the exact number and clarify the averaging procedure across settings.
  2. [Experimental Setup / Tables 1 and 2] All evaluations use 500 test samples, one seed, and nondeterministic decoding (temperature 0.7, top-p 0.9). No confidence intervals or significance tests are reported. Given the emphasis on consistency across settings, reporting standard deviations or paired tests would strengthen the claims.
  3. [Appendix A.6 / StrategyQA main result] The masking ratio β is selected after a sensitivity analysis on StrategyQA and then used in the main StrategyQA results. Please state whether β was tuned on a held-out split and whether the main results are affected by this selection.
  4. [General presentation] The notation I_eP(X;Y|Z) in Proposition 2 and the superscript eP are used without a crisp definition in the main text; please define the augmented distribution eP in the main body, not only in the appendix.

Circularity Check

2 steps flagged

CIE and CDE gains are directly trained by the same interventions used as evaluation metrics; FT/EA gains are largely forced by the attention mask, while Proposition 4 stops at hidden states rather than the visible CoT required by Definition 1.

specific steps
  1. fitted input called prediction [Methodology, Eq. (6) and Appendix A.4 (CIE)]
    "LX-int =E_DB [−log pθ(y′_i | z_i, x′_i)] ... CIE measures whether changing the CoT changes the final answer while keeping the question fixed. For each question zi and the generated CoT xi, we use DeepSeek-V3.2 to generate a counterfactual CoT x′i... CIE = 1/M Σ 1[ŷCIE_i ≠ ŷi]."

    L_X-int is a maximum-likelihood loss on the exact counterfactual-CoT intervention (same z_i, counterfactual CoT x'_i, target y'_i ≠ y_i) that the CIE metric evaluates (replace the CoT with a counterfactual x'_i under the same z_i and count answer changes). Training on D_A and D_B therefore directly optimizes the CIE behavior; reporting high CIE as evidence of faithfulness measures the training objective, not an independent prediction. The counterfactual CoTs are generated by the same DeepSeek-V3.2 protocol in both training and evaluation.

  2. fitted input called prediction [Methodology, Eq. (7) and Appendix A.4 (CDE)]
    "LZ-int =E_DC [−log pθ(yi | z′_i, xi)] ... CDE measures whether changing the question still changes the answer when the CoT is fixed. ... We compare the original prediction ŷi based on (zi, xi) with the prediction ŷCDE_i based on (z′i, xi)."

    L_Z-int trains pθ(y_i | z'_i, x_i) to preserve the answer when the instruction is intervened and the CoT is fixed; CDE is defined as the proportion of answer changes under exactly that intervention (counterfactual z'_i with fixed x_i). A low CDE is thus the direct objective of L_Z-int, and it is further enforced by the inference-time mask that blocks instruction-token attention. Reporting CDE as a faithfulness gain re-states the training target rather than providing independent validation.

full rationale

The central empirical claim—that CASE improves CoT faithfulness—is substantially supported by metrics that are not independent of the method's components. The counterfactual-CoT loss L_X-int (Eq. 6) is trained on (z, x', y') with y' ≠ y, which is precisely the CIE intervention; the biased-instruction loss L_Z-int (Eq. 7) trains the model to keep the same answer under an intervened instruction with fixed CoT, which is precisely what CDE measures. These two components therefore guarantee, by construction, part of the reported G-mean_Faith improvement. The inference-time attention mask further forces answer tokens to rely on CoT-side tokens, so the large FT and EA gains in Table 1 and the ISE-only ablations in Table 3 are substantially structural consequences of the intervention rather than evidence of a learned semantic CoT-to-answer mechanism. Cross-dataset transfer (Table 2) and the TCA-only ablations provide some independent support that the training alignment generalizes, so the circularity is partial rather than complete. The paper also concedes a gap between its theory and its own definition: Definition 1 requires Y⊥Z|X for the visible textual CoT, while Proposition 4 only proves p^AM_θ(y|z,h_X)=p^AM_θ(y|h_X) conditioned on CoT-side hidden states H_X, with the text stating 'instruction information retained in H_X remains CoT-mediated.' That concession is a correctness gap rather than a circular step, but it means the theoretical derivation does not establish the textual-level faithfulness named in Definition 1. Overall, the headline 'faithfulness improvement' partially reduces to the training losses and structural mask used to define the method.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 1 invented entities

The central claim rests on the causal Z→X→Y model, the quality of LLM-generated CoT data, and the assumption that masking direct attention while retaining H_X is sufficient for textual faithfulness. The only tuned free parameter is β, selected on StrategyQA. No new physical entities are introduced; H_X is a theoretical abstraction rather than an empirically anchored mediator.

free parameters (1)
  • β (question-token masking ratio for true/false tasks) = 0.8
    Chosen from a sensitivity sweep on StrategyQA with Llama-3.1-8B (Table S-IV), then applied to all datasets and models. This is a hyperparameter tuned on an evaluation target.
axioms (5)
  • domain assumption Faithful CoT is characterized by the causal chain Z→X→Y and conditional independence Y⊥Z|X (Definition 1).
    Adopted from Bao et al. (2025). If this causal model is wrong, the method's target is misplaced.
  • domain assumption The LLM-generated CoTs and counterfactual CoTs used for training and evaluation are valid and consistent.
    Supported only by an automatic GPT-5.5 audit (Appendix A.3), not by human verification. If generated CoTs are low-quality, the learned X→Y dependence is spurious.
  • domain assumption Attention masking from answer tokens to instruction tokens, while preserving CoT-side hidden states H_X, is a meaningful step toward faithfulness.
    Proposition 4 shows p(y|z,h_X)=p(y|h_X), but h_X is formed under z and contains instruction information; the paper assumes the mediated path is acceptable for the faithfulness definition.
  • domain assumption The perturbation metrics FT, EA, CIE, and CDE are valid measures of CoT faithfulness.
    The paper relies on prior work for metric validity, but these are contested proxies rather than ground-truth causal measurements.
  • standard math Variational lower bounds and KL convexity used in Propositions 1-3.
    Standard information-theoretic inequalities; the proofs in Appendix A.1 are mathematically routine.
invented entities (1)
  • H_X, the CoT-side computational state / mediator no independent evidence
    purpose: Used to argue that attention masking removes only the direct Z→Y path while preserving a mediated path; central to Proposition 4 and the faithfulness argument.
    H_X is a latent construct introduced by the paper. There is no falsifiable handle outside the paper to distinguish predictions made through H_X from predictions made through the visible text X.

pith-pipeline@v1.3.0-alltime-deepseek · 23451 in / 13575 out tokens · 123095 ms · 2026-08-01T14:12:03.206525+00:00 · methodology

0 comments
read the original abstract

Chain-of-thought (CoT) reasoning is widely used to improve both the performance and interpretability of large language models (LLMs), yet the generated reasoning may not faithfully support the final answer. We study this problem from a causal perspective, where a faithful CoT process should follow the chain $Z\rightarrow X\rightarrow Y$, with $Z$, $X$, and $Y$ denoting the instruction, reasoning chain, and final answer, respectively. In this process, the instruction should affect the answer only through the reasoning chain. However, conventional autoregressive LLMs condition answer generation on both the instruction and the CoT, which still allows a direct instruction-to-answer shortcut. To address this issue, we propose CASE, a framework that combines training-time causal alignment and inference-time structural enforcement. During training, CASE builds counterfactual-CoT, biased-instruction, and empty-instruction datasets, and applies selective-loss fine-tuning to strengthen CoT-to-answer dependence while suppressing instruction shortcuts. During inference, CASE masks direct attention from instruction tokens to answer tokens, preventing the model from bypassing the generated CoT. We provide an information-theoretic analysis showing how these components promote faithful chains. Experiments on three models and four benchmarks show that CASE achieves a 37\% average per-setting relative improvement in overall CoT faithfulness over the strongest baselines, exhibits stronger cross-dataset faithfulness transfer, and maintains competitive average accuracy. Code is available at https://github.com/oddwang/CASE.

Figures

Figures reproduced from arXiv: 2607.18820 by Changwu Huang, Ke Tang, Xin Yao, Yinghua Yao, Ziming Wang.

Figure 1
Figure 1. Figure 1: High-level intuition of CASE. CASE improves CoT [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed CASE framework. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] 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

39 extracted references · 6 linked inside Pith

  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=

  2. [2]

    Advances in Neural Information Processing Systems , volume=

    Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting , author=. Advances in Neural Information Processing Systems , volume=

  3. [3]

    Faithfulness tests for natural language explanations

    Atanasova, Pepa and Camburu, Oana-Maria and Lioma, Christina and Lukasiewicz, Thomas and Simonsen, Jakob Grue and Augenstein, Isabelle. Faithfulness tests for natural language explanations. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). 2023

    2023

  4. [4]

    arXiv preprint arXiv:2307.13702 , year=

    Measuring faithfulness in chain-of-thought reasoning , author=. arXiv preprint arXiv:2307.13702 , year=

    Pith/arXiv arXiv

  5. [5]

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

    Measuring chain of thought faithfulness by unlearning reasoning steps , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  6. [6]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    On measuring faithfulness or self-consistency of natural language explanations , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  7. [7]

    Advances in Neural Information Processing Systems , volume=

    A unified approach to interpreting model predictions , author=. Advances in Neural Information Processing Systems , volume=

  8. [8]

    Findings of the Association for Computational Linguistics: NAACL 2024 , pages=

    How interpretable are reasoning explanations from prompting large language models? , author=. Findings of the Association for Computational Linguistics: NAACL 2024 , pages=

    2024

  9. [9]

    Chain-of-probe: Examining the necessity and accuracy of

    Wang, Zezhong and Zeng, Xingshan and Liu, Weiwen and Wang, Yufei and Li, Liangyou and Wang, Yasheng and Shang, Lifeng and Jiang, Xin and Liu, Qun and Wong, Kam-Fai , booktitle=. Chain-of-probe: Examining the necessity and accuracy of

  10. [10]

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

    Towards better chain-of-thought: A reflection on effectiveness and faithfulness , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  11. [11]

    Le and Ed H

    Xuezhi Wang and Jason Wei and Dale Schuurmans and Quoc V. Le and Ed H. Chi and Sharan Narang and Aakanksha Chowdhery and Denny Zhou , title =. The Eleventh International Conference on Learning Representations,

  12. [12]

    Drift: Enhancing

    Li, Jiazheng and Yan, Hanqi and He, Yulan. Drift: Enhancing. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025

    2025

  13. [13]

    The power of scale for parameter-efficient prompt tuning

    Lester, Brian and Al-Rfou, Rami and Constant, Noah. The power of scale for parameter-efficient prompt tuning. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021

    2021

  14. [14]

    Think you have solved question answering?

    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?

  15. [15]

    Did aristotle use a laptop?

    Geva, Mor and Khashabi, Daniel and Segal, Elad and Khot, Tushar and Roth, Dan and Berant, Jonathan , journal=. Did aristotle use a laptop?. 2021 , publisher=

    2021

  16. [16]

    arXiv preprint arXiv:1705.04146 , year=

    Program induction by rationale generation: Learning to solve and explain algebraic word problems , author=. arXiv preprint arXiv:1705.04146 , year=

    Pith/arXiv arXiv

  17. [17]

    2024 , note =

    Introducing. 2024 , note =

    2024

  18. [18]

    arXiv preprint arXiv:2412.19437 , year=

    Deepseek-v3 technical report , author=. arXiv preprint arXiv:2412.19437 , year=

    Pith/arXiv arXiv

  19. [19]

    the Twelfth International Conference on Learning Representations , year=

    Let's verify step by step , author=. the Twelfth International Conference on Learning Representations , year=

  20. [20]

    Rewarding progress: Scaling automated process verifiers for

    Setlur, Amrith and Nagpal, Chirag and Fisch, Adam and Geng, Xinyang and Eisenstein, Jacob and Agarwal, Rishabh and Agarwal, Alekh and Berant, Jonathan and Kumar, Aviral , journal=. Rewarding progress: Scaling automated process verifiers for

  21. [21]

    Achiam, Josh and Adler, Steven and Agarwal, Sandhini and Ahmad, Lama and Akkaya, Ilge and Aleman, Florencia Leoni and Almeida, Diogo and Altenschmidt, Janko and Altman, Sam and Anadkat, Shyamal and others , journal=

  22. [22]

    International Conference on Machine Learning , pages=

    On calibration of modern neural networks , author=. International Conference on Machine Learning , pages=. 2017 , organization=

    2017

  23. [23]

    Towards faithfully interpretable NLP systems: How should we define and evaluate faithfulness?

    Jacovi, Alon and Goldberg, Yoav. Towards faithfully interpretable NLP systems: How should we define and evaluate faithfulness?. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020

    2020

  24. [24]

    Preprint, alphaXiv , pages=

    Chain-of-thought is not explainability , author=. Preprint, alphaXiv , pages=

  25. [25]

    arXiv preprint arXiv:2307.11768 , year=

    Question decomposition improves the faithfulness of model-generated reasoning , author=. arXiv preprint arXiv:2307.11768 , year=

    Pith/arXiv arXiv

  26. [26]

    2026 , url=

    Xu Shen and Song Wang and Zhen Tan and Laura Yao and Xinyu Zhao and Kaidi Xu and Xin Wang and Tianlong Chen , booktitle=. 2026 , url=

    2026

  27. [27]

    arXiv preprint arXiv:2406.10625 , year=

    On the hardness of faithful chain-of-thought reasoning in large language models , author=. arXiv preprint arXiv:2406.10625 , year=

    Pith/arXiv arXiv

  28. [28]

    The 5th Workshop on Mathematical Reasoning and AI at NeurIPS 2025 , year=

    FoCus: Improving Faithfulness in Chain-of-Thoughts by Training on Structured Reasoning Data , author=. The 5th Workshop on Mathematical Reasoning and AI at NeurIPS 2025 , year=

    2025

  29. [29]

    How likely do

    Bao, Guangsheng and Zhang, Hongbo and Wang, Cunxiang and Yang, Linyi and Zhang, Yue , booktitle=. How likely do

  30. [30]

    Causal Learning: Psychology, Philosophy, and Computation , volume=

    Causal reasoning through intervention , author=. Causal Learning: Psychology, Philosophy, and Computation , volume=

  31. [31]

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

    Making reasoning matter: Measuring and improving faithfulness of chain-of-thought reasoning , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    2024

  32. [32]

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

    A causal lens for evaluating faithfulness metrics , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  33. [33]

    Advances in Neural Information Processing Systems , volume=

    Direct preference optimization: Your language model is secretly a reward model , author=. Advances in Neural Information Processing Systems , volume=

  34. [34]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  35. [35]

    Guo, Daya and Yang, Dejian and Zhang, Haowei and Song, Junxiao and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Zhang, Ruoyu and Ma, Shirong and Bi, Xiao and others , journal=

  36. [36]

    Liu, Jian and Cui, Leyang and Liu, Hanmeng and Huang, Dandan and Wang, Yile and Zhang, Yue , booktitle =

  37. [37]

    Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages=

    Can a suit of armor conduct electricity? A new dataset for open book question answering , author=. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages=

    2018

  38. [38]

    Quality Progress , volume=

    A balancing act-optimizing a products properties , author=. Quality Progress , volume=

  39. [39]

    2026 , howpublished =

    GPT-5.5 system card , author =. 2026 , howpublished =

    2026