Pith. sign in

REVIEW 4 major objections 5 minor 74 references

REIN: Bridging the Gap between Reasoning and Reliability via Reflection and Abstention Alignment

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read REIN trains a large reasoning model to think, reflect on its own draft reasoning, and then either answer or say 'I don't know', reducing the fraction of wrong drafts it confidently endorses by 58–72% while keeping 86–91% of questions…

desk verdict REIN is a competent single-pass reflection-and-abstention alignment method whose headline H-Proxy reduction is partly a reward-echo artifact, but whose cross-task transfer and selective accuracy gains are real enough to warrant serious review. read the letter →

arxiv 2608.07931 v1 pith:R6J3FMPH submitted 2026-08-08 cs.AI

Zhengze Huang , Luyang Yu , Di Hong , Xinzhe Huang , Wanyu Lin , Zhixuan Chu , Zhan Qin , Tianhang Zheng This is my paper · ORCID
classification cs.AI
keywords largereasoningmodelshallucinationreflectionabstentionself-correctionreinforcementlearninguncertaintycalibrationselectiveanswering
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Large reasoning models hallucinate through two separable routes: reasoning hallucination, where a flawed inference step propagates to a wrong conclusion, and knowledge hallucination, where the model simply lacks the relevant knowledge. This paper claims that one alignment pass can handle both, by training the model to emit a structured completion then then , and rewarding the reflection for correctly judging whether its own draft conclusion is sound. A second, boundary-aware reward makes the model answer 'I don't know' when none of a group of sampled attempts produced a correct answer, which the paper treats as evidence the question sits beyond the model's current knowledge. On math and commonsense benchmarks across four backbones, REIN endorses 58–72% fewer incorrect draft conclusions than the untrained base models while keeping 86–91% of questions answered, all in a single forward pass with no verifier or retry loop at test time. If this holds, reliability is something that can be trained into the model's own self-assessment rather than bolted on with extra inference-time machinery.

What carries the argument

The load-bearing object is the structured completion <think> → <reflection> → <answer>, parsed into a pre-reflection reasoning trajectory, the draft conclusion it implies, a reflection span, and a final answer that is either substantive or 'I don't know'. Two rewards carry the argument: the reflection-veracity reward aligns the reflection's reliable/unreliable stance with whether the draft conclusion was actually correct, and the boundary-aware IDK reward estimates a policy-relative knowledge boundary from a group of $K=16$ on-policy completions—'within' if any sample succeeded, 'beyond' otherwise—and prices abstention only against that boundary. The rewards are combined into one scalar with accuracy and format rewards and optimized by GRPO, a group-relative policy-gradient update, under a two-phase curriculum: the first phase explicitly advertises 'I don't know' as an available final action, and the second withdraws the cue so the model must infer when to abstain on its own. The boundary indicator is the mechanism that lets a single training signal tell a fixable inference slip apart from genuinely missing knowledge.

What would settle it

Recompute the hallucination proxy for the untrained base models under a free-format protocol, where the endorsement decision is made by an independent judge instead of the model's own forced reflection line; if the base models' proxy under free formatting is much lower than the reported 35–48% range, then part of the 58–72% relative reduction is a measurement artifact. As a second check, test abstention against an oracle boundary built from 64 independent samples: if REIN's IDK precision and recall measured against that oracle do not stay in the paper's reported 78–83% and 66–73% ranges, the learned 'I don't know' behavior is chasing the estimator, not the true knowledge gap.

Watch

Extended reading notes

Core claim

REIN's central claim is that the two failure modes behind hallucination are distinguishable inside one completion and jointly addressable by two aligned reward signals. A reflection-veracity reward gives +1 when the model's reflection stance agrees with the verified correctness of the draft conclusion implied by its reasoning, and −1 when it disagrees, so the model learns to endorse only drafts that are actually right. A boundary-aware IDK reward marks a prompt as within or beyond the policy's sampled knowledge boundary depending on whether any of $K=16$ sampled completions produced a verified-correct final answer, then rewards correct substantive answers inside the boundary, penalizes unnecessary abstention inside it, and rewards 'I don't know' beyond it. Trained with GRPO on GSM8K and evaluated on MATH-500, StrategyQA, and ARC-Challenge, REIN achieves the best selective accuracy and the lowest hallucination proxy (the fraction of incorrect draft conclusions the reflection nonetheless endorses as reliable) in every reported cell, cutting the proxy by 58–72% relative to the base models at 86–91% average coverage.

Load-bearing premise

The headline reduction is measured by forcing every model, trained or not, to print a <reflection> verdict before its answer, and the numbers assume both that an untrained model's forced verdict is a meaningful self-endorsement and that the knowledge boundary estimated from $K=16$ training samples is still the right line to abstain beyond when the model is deployed.

Editorial extensions

If this is right

  • Reliability alignment no longer requires process supervision, inference-time verifiers, majority voting, or multi-round critique loops: the decision to trust or abstain happens inside a single forward pass.
  • Abstention becomes policy-relative and difficulty-sensitive: a prompt with any successful sample is treated as answerable, and only zero-success groups license 'I don't know', which concentrates refusals near the capability frontier (IDK rate rises from 2.1% on two-step GSM8K problems to 14.8% on five-plus-step problems, and from 3.2% on MATH Level 1 to 31.8% on Level 5).
  • Because training on GSM8K alone improves reliability on StrategyQA and ARC-Challenge, the reflect-then-abstain behavior transfers across task families rather than being memorized per format.
  • Reflection acts as a correction mechanism before finalization: across the four benchmarks it repairs 11.0% of initially incorrect drafts while corrupting only 0.9% of initially correct ones.
  • Deeper reflection is counterproductive: adding nested reflection cycles further suppresses hallucination but reduces coverage and effective accuracy through overcorrection, so the single reflection layer is the preferred operating point.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The selective-prediction table shows the trained reflection stance and sixteen-sample self-consistency agreement are complementary reliability signals (the combination gives the best AUROC), which the paper leaves unexploited; a production variant could run the single-pass reflection as the default gate and add sampling only on borderline cases.
  • The paper's confidence bound—no successes in $K$ draws implies the true success probability is at most $1-\alpha^{1/K}$, about 13.4% at $K=16$ and 90% confidence—hands deployments a knob the paper does not turn: choose $K$ by how many mistaken abstentions the application can tolerate, since a 'beyond-boundary' label still leaves a measurable chance the model could have answered.
  • REIN's reflection-veracity reward presupposes an automatic check of whether the draft conclusion is correct, which math and multiple-choice benchmarks provide; moving to open-ended generation would force a learned correctness proxy, in which case the framework is only as reliable as that proxy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. REIN trains a large reasoning model to emit a structured <think> -> <reflection> -> <answer> completion in a single forward pass. During GRPO training, a reflection-veracity reward aligns the reflection stance with the verified correctness of the pre-reflection draft conclusion, and a boundary-aware IDK reward encourages abstention when none of K=16 sampled completions yields a correct final answer. The method is evaluated on GSM8K, MATH-500, StrategyQA, and ARC-Challenge with Qwen2.5-7B and Llama-3.1-8B, training only on GSM8K and reporting selective accuracy, H-Proxy, coverage, and effective accuracy. The paper reports a 58-72% relative reduction in H-Proxy, 86-91% average coverage, and 6.6-14.2% selective-accuracy improvement, plus ablations, a draft-to-final transition analysis, and a robustness appendix.

Significance. If the main results hold, REIN would be a meaningful single-pass reliability alignment method: it makes abstention a trainable decision, connects reflection to verified draft correctness, and does not require process supervision, external search, or test-time controllers. The cross-task transfer design (training only on GSM8K and evaluating on MATH, StrategyQA, and ARC-Challenge) is a genuine strength, as are the component ablation, the K-sensitivity analysis with confidence bounds, the draft-to-final transition matrix, and the robustness appendix that fixes prompt, parser, and decoding choices. The significance is partly conditional, however, because the headline hallucination metric is closely tied to the reflection-veracity training reward and because the baseline comparison for that metric uses a forced reflection template that untrained models were never trained to follow.

major comments (4)
  1. [Appendix A.3, Eq. (24) vs. Section 3.2, Eq. (9)] H-Proxy as defined in Eq. (24) is the conditional probability that an incorrect draft is endorsed by the native reflection, Pr[J(r)=1 | T(x,y0)=0]. This is exactly the quantity that the reflection-veracity reward in Eq. (9) rewards the model to minimize: the reward gives +1 when J(r)=T(x,y0) and -1 when they disagree, so incorrect drafts with J=1 are directly penalized. The reported 58-72% H-Proxy reductions therefore report, to a large degree, the model's success at optimizing its own training objective rather than an independent measurement of hallucination. The cross-task results on MATH, StrategyQA, and ARC-Challenge are less subject to this concern because REIN was trained only on GSM8K, and those results do provide independent support. I recommend reframing H-Proxy as an alignment metric and adding at least one independent hallucination measure that does not appear in any reward term, such as post-hoc verifier disagreement or a factuality benchmark, reported separately for the held-out tasks.
  2. [Section 4 and Appendix A.2] The H-Proxy comparison for untrained baselines is not fully controlled. Section 4 states that for H-Proxy every method, including Base and Instruct, is prompted to generate the same <think>-><reflection>-><answer> structure, and Appendix A.2 defines 'native' solely by whether the reflection is emitted before the final answer, regardless of whether the method was trained with a reflection objective. An untrained base model has never been trained to produce a meaningful reliability stance in that slot, so its 'reflection' is arbitrary text generated under an unfamiliar template. Two concrete consequences follow. First, the denominator N_w^(0) in Eq. (23) counts only parseable drafts; if baselines frequently produce malformed structure, their H-Proxy is computed over a different, smaller population, and the paper reports no per-method draft extraction rates. Second, the stance parser maps wrong, uncertain, and unparsable to non-endorsement, so a baseline that emits 'I am not sure' or no stance is automatically treated as non-endorsing, which can artificially lower or raise its H-Proxy for reasons unrelated to reliability. Please report per-method draft extraction rates and stance distributions, and either restrict H-Proxy to completions with parseable drafts and valid stances or supplement the forced-template comparison with a post-hoc judging protocol that applies the same judge to every method's draft.
  3. [Section 3.4 and Appendix F] The supervised fine-tuning stage that initializes reflection behavior is not specified. Section 3.4 says REIN first applies 'a short supervised fine-tuning stage to initialize the structured output format and the reflection labels,' but the paper does not state how the reflection labels were produced, whether they were written by humans, generated by an LLM, derived from reference answers, or rule-based, nor how many examples were used. This detail is load-bearing because the reflection-veracity reward aligns the model to these initial labels, and because the paper claims the method requires 'no process supervision.' If the SFT labels are generated with an LLM or an oracle, that qualification needs to be stated; if they are rule-based, the construction should be described. Please provide the SFT data construction procedure, label source, and dataset size.
  4. [Abstract, Section 5, Table 1, and Appendix Table 5] The headline relative reduction in H-Proxy is not consistent with the reported tables. In Table 1, the eight backbone-benchmark cells give relative reductions from Base to REIN ranging from about 62.5% (Llama-3.1-8B on MATH, 48.0 to 18.0) to about 75.0% (Qwen2.5-7B on ARC-Challenge, 36.0 to 9.0). Including Appendix Table 5, the Mistral-7B cell on MATH shows a reduction of about 54%, and the DeepSeek-R1-Distill-8B cell on MATH shows about 81%. The Abstract's '58~72%' range does not match either set of numbers. The text in Section 5 also reports a macro-averaged reduction of 28.5 percentage points, which is a different quantity from the 58-72% relative reduction. Please state exactly which subset of cells the 58-72% range refers to, or report a consistent range computed from all evaluated backbone-benchmark pairs.
minor comments (5)
  1. [Table 3] The H-proxy column in Table 3 reports values as 0.11, 0.08, and 0.06, while Table 1 and the rest of the paper report H-Proxy as percentages (e.g., 11.0, 8.0, 6.0). Please use one consistent convention.
  2. [Section 4, Models and Baselines] R-Tuning is cited as 'Cohen et al., 2024,' but the reference list attributes R-Tuning to Zhang et al., 2024; Cohen et al. is the IDK-token work. Please correct the citation.
  3. [Appendix H] The third qualitative example shows 'Wrong. There are two roots ±2; I only listed one.' followed by a final answer of \boxed{2}. This appears to be an example where the reflection identifies the error but the final answer is not corrected; please clarify whether this is an intended failure-mode illustration or a typo in the final answer.
  4. [Section 5, first paragraph] The sentence 'REIN reduces H-Proxy by 28.5%' says '28.5%' but the surrounding text compares percentage-point differences (e.g., 'by 14.0%' and 'by 7.0%' for selective and effective accuracy). Please use 'percentage points' consistently to avoid confusing relative and absolute improvements.
  5. [Appendix B.3] The MCC formula is split across a line break with the equation typeset as 'MCC = TP·TN−FP·FNp' followed by a continuation; the square-root term and the denominator are not readable. Please typeset the full formula on one line.

Circularity Check

1 steps flagged · score 4.0 of 10

H-Proxy is the exact target of the reflection-veracity reward, so the headline hallucination reduction is partly a fitted objective; held-out transfer and oracle-based IDK evals keep the central claim partly independent.

  1. fitted input called prediction [Section 3.2 Eq. (9); Section 4 metric description; Section A.3 H-Proxy definition]
    "The reflection-veracity reward is defined as r_vrcty(x, r, y0) := +1, s(r)≠⊥ ∧J(r) =T(x, y0), −1, s(r)≠⊥ ∧J(r)≠T(x, y0), 0, s(r) =⊥. ... H-Proxy measures the proportion of incorrect pre-reflection draft conclusions y0 that are nevertheless positively endorsed by the corresponding reflection r. Specifically, it estimates Pr[J(r) = 1|T(x, y0) = 0]."

    For every incorrect draft, T(x,y0)=0, so the reward assigns −1 when the reflection endorses the draft (J=1) and +1 when it does not (J=0). GRPO on r_vrcty therefore directly minimizes the count of incorrect drafts endorsed as correct, which is exactly the numerator N_nat_e of H-Proxy, while the denominator N_w is the count of incorrect parseable drafts. The reported 58–72% H-Proxy reduction is thus the fitted value of the training objective rather than an independent measurement of hallucination. The finding is only partly circular because H-Proxy is also reported on held-out benchmarks (MATH, StrategyQA, ARC-C) and against baselines that were not trained with this reward, giving the reduction genuine transfer content.

full rationale

No load-bearing self-citation chain exists: the authors' own prior works appear only in related-work surveys and are not used to justify any theorem, uniqueness claim, or design choice. The framework derivation (structured <think>-><reflection>-><answer> completion, reflection-veracity reward, boundary-aware IDK reward, GRPO) is self-contained. The main circularity concern is that the headline H-Proxy metric coincides by construction with the reflection-veracity reward target: Eq. (9) rewards exactly J(r)=T(x,y0), and H-Proxy counts incorrect drafts (T=0) that are endorsed (J=1), so the training objective directly minimizes the reported metric on the training distribution. This prevents treating the hallucination reduction as an independent validation of the method. However, the central claim retains independent content because the reduction is measured on three benchmarks not used for GRPO training, compared against untrained and other- objective baselines under a unified protocol, and supplemented by IDK precision/recall evaluated against a 64-sample oracle reference. The forced reflection template for baselines is a fairness/correctness concern, not a circularity, and is therefore not scored here.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The 'sampled knowledge boundary' is an empirical estimator derived from the policy's own rollouts, not an independent ontological commitment. The main free parameters are the group size K, the manually chosen reward weights, and the reflection depth d.

free parameters (3)
  • Group size K = 16
    Chosen as a cost/stability trade-off. Table 10 shows boundary accuracy and selective accuracy increase with K, so the headline numbers depend on this choice.
  • Reward weights (w_acc, w_xml, w_pres, w_vrcty, w_idk) = 8.0, 3.0, 1.5, 2.0, 4.0
    Manually set in Section 3.4 and Table 13. No sensitivity analysis over reward weights is reported, so the robustness of this weighting is unverified.
  • Reflection depth d = 1
    Selected from the ablation in Table 3. Deeper reflection lowers H-Proxy but reduces coverage and effective accuracy, so this design choice determines the reported trade-off.
assumptions (5)
  • domain assumption K sampled completions are conditionally independent given the prompt and policy.
    Used in Eq. (26) and Lemma 1 to derive the confidence bound on the knowledge boundary. Sampling with temperature may exhibit dependencies in practice.
  • domain assumption The task verifier T(x,y) is exact and complete for GSM8K, MATH-500, StrategyQA, and ARC-Challenge.
    All rewards and metrics depend on automated verification (numeric equivalence, math equivalence, exact matching). Verifier errors would bias both training and evaluation.
  • ad hoc to paper Forcing untrained baselines to emit a reflection span produces a fair H-Proxy comparison.
    Section 4 prompts every method with the same structured format. If these reflections are arbitrary or uncalibrated for methods not trained with reflection, the reported H-Proxy reduction is not a fair comparison.
  • ad hoc to paper Reflection labels for the SFT initialization can be reliably produced.
    Section 3.4 mentions a short SFT stage to initialize reflection behavior, but does not describe how the reflection supervision labels are created.
  • domain assumption Training only on GSM8K transfers to MATH, StrategyQA, and ARC-Challenge.
    GRPO uses only the GSM8K training set, and the main results assume the learned reliability behavior generalizes across tasks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of REIN: Bridging the Gap between Reasoning and Reliability via Reflection and Abstention Alignment." pith.science (2026). https://pith.science/paper/R6J3FMPH

@misc{pith2026260807931,
  author       = {Pith},
  title        = {Pith review of: REIN: Bridging the Gap between Reasoning and Reliability via Reflection and Abstention Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R6J3FMPH}},
  note         = {Machine review of arXiv:2608.07931}
}
abstract

Large reasoning models (LRMs) are prone to hallucination, which undermines their reliability and poses challenges for safe deployment. Hallucinations in LRMs arise from two distinct failure sources: reasoning hallucination, where flawed inference steps propagate to an incorrect conclusion, and knowledge hallucination, where the model lacks the requisite factual knowledge to answer the query. To address reasoning hallucination, we propose REIN, an alignment framework that trains LRMs to produce a structured reasoning sequence, $\texttt{<think>} $$\rightarrow$ $\texttt{<reflection>} $$\rightarrow$ $\texttt{<answer>}$, enabling explicit self-reflection before committing to a final answer. To address knowledge hallucination, REIN introduces a reward mechanism that encourages explicit abstention (e.g., "I don't know") when none of the sampled reasoning chains yields a correct answer, allowing the model to refrain from unsupported predictions. Extensive evaluations on mathematical and commonsense reasoning benchmarks show that REIN consistently improves selective accuracy, reduces incorrect-but-self-endorsed responses, and maintains high coverage compared with competitive baselines. Notably, REIN achieves these gains within a single forward pass, without requiring process supervision, inference-time controllers, external search, or multi-round critiques. Experiments on multiple backbones show that REIN reduces the hallucination proxy by $58\sim72\%$ relative to the base models while maintaining $86\sim91\%$ average coverage, and improves selective accuracy on attempted questions by $6.6\sim14.2\%$.

Figures

Figures reproduced from arXiv: 2608.07931 by the authors.

Figure 1
Figure 1. Overview of the REIN training pipeline: reflection-augmented completions feed into GRPO with reflection [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Reliability judgment calibration across methods and backbones. We report reflection–answer agreement and [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Comparison of REIN with existing methods. REIN achieves the highest Accuracy, Truthfulness, and [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Reward Composition Dynamics During REIN Training (Qwen2.5-7B) [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Reduction of false endorsement through reflection alignment. False endorsement measures the fraction [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]
Figure 6
Figure 6. Figure 6: Training curves of KL divergence, reward, and loss. KL remains low and stable, while reward steadily [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: Reflection judgments align with answer outcomes across backbones and benchmarks. Each small point [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

74 extracted references · 40 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=

  2. [2]

    arXiv preprint arXiv:2110.14168 , year=

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

    arXiv

  3. [3]

    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

  4. [4]

    Advances in Neural Information Processing Systems , year=

    Training Language Models to Follow Instructions with Human Feedback , author=. Advances in Neural Information Processing Systems , year=

  5. [5]

    Advances in Neural Information Processing Systems , year=

    Direct Preference Optimization: Your Language Model is Secretly a Reward Model , author=. Advances in Neural Information Processing Systems , year=

  6. [6]

    Dai, Josef and Pan, Xuehai and Sun, Ruiyang and Ji, Jiaming and Xu, Xinbo and Liu, Mickel and Wang, Yizhou and Yang, Yaodong , journal=. Safe

  7. [7]

    Rejection Improves Reliability: Training

    Xu, Ziwei and Liu, Sanxing and Guo, Qing and Yao, Liang and Du, Yifan and Xu, Chen and Xu, Kai and Song, Dacheng and Chen, Cheng and Zhang, Jun and Xiong, Hui , journal=. Rejection Improves Reliability: Training

    work page

  8. [8]

    Cheng, Qinyuan and Sun, Tianxiang and Liu, Xiangyang and Zhang, Wenwei and Yin, Zhangyue and Li, Shimin and Li, Linyang and He, Zhengfu and Chen, Kai and Qiu, Xipeng , booktitle=. Can. 2024 , publisher=

    work page 2024

Show all 74 references
  1. [9]

    arXiv preprint arXiv:2602.02136 , year=

    Mitigating Safety Tax via Distribution-Grounded Refinement in Large Reasoning Models , author=. arXiv preprint arXiv:2602.02136 , year=

  2. [10]

    Journal of Machine Learning Research , volume=

    On the Foundations of Noise-Free Selective Classification , author=. Journal of Machine Learning Research , volume=

  3. [11]

    Min, Sewon and Krishna, Kalpesh and Lyu, Xinxi and Lewis, Mike and Yih, Wen-tau and Koh, Pang Wei and Iyyer, Mohit and Zettlemoyer, Luke and Hajishirzi, Hannaneh , journal=

  4. [12]

    arXiv preprint arXiv:2312.10997 , year=

    Retrieval-Augmented Generation for Large Language Models: A Survey , author=. arXiv preprint arXiv:2312.10997 , year=

    arXiv

  5. [13]

    arXiv preprint arXiv:2305.13534 , year=

    How Language Model Hallucinations Can Snowball , author=. arXiv preprint arXiv:2305.13534 , year=

    arXiv

  6. [14]

    arXiv preprint arXiv:2505.23646 , year=

    Are Reasoning Models More Prone to Hallucination? , author=. arXiv preprint arXiv:2505.23646 , year=

    arXiv

  7. [15]

    arXiv preprint arXiv:2505.09388 , year=

    arXiv

  8. [16]

    arXiv preprint arXiv:2501.12948 , year=

    arXiv

  9. [17]

    2024 , howpublished=

    2024

  10. [18]

    Yu, Qiying and others , journal=

  11. [19]

    Chen, Zhenyu and others , journal=

  12. [20]

    Scaling Physical Reasoning with the

    Zheng, Shenghe and Cheng, Qianjia and Yao, Junchi and Wu, Mengsong and He, Haonan and Ding, Ning and Cheng, Yu and Hu, Shuyue and Bai, Lei and Zhou, Dongzhan and Cui, Ganqu and Ye, Peng , journal=. Scaling Physical Reasoning with the

  13. [21]

    Using an

    Nam, Daye and Macvean, Andrew and Hellendoorn, Vincent and Vasilescu, Bogdan and Myers, Brad , journal=. Using an

  14. [22]

    Advances in Neural Information Processing Systems , volume=

    Self-Refine: Iterative Refinement with Self-Feedback , author=. Advances in Neural Information Processing Systems , volume=. 2023 , url=

    2023

  15. [23]

    Advances in Neural Information Processing Systems , volume=

    Reflexion: Language Agents with Verbal Reinforcement Learning , author=. Advances in Neural Information Processing Systems , volume=. 2023 , url=

    2023

  16. [24]

    Advances in Neural Information Processing Systems , volume=

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models , author=. Advances in Neural Information Processing Systems , volume=. 2023 , url=

    2023

  17. [25]

    arXiv preprint arXiv:2211.14275 , year=

    Solving Math Word Problems with Process- and Outcome-based Feedback , author=. arXiv preprint arXiv:2211.14275 , year=

    arXiv

  18. [26]

    2024 , url=

    Lightman, Hunter and Kosaraju, Vineet and Burda, Yuri and Edwards, Harri and Baker, Bowen and Lee, Teddy and Leike, Jan and Schulman, John and Sutskever, Ilya and Cobbe, Karl , booktitle=. 2024 , url=

    2024

  19. [27]

    arXiv preprint arXiv:2207.05221 , year=

    Language Models (Mostly) Know What They Know , author=. arXiv preprint arXiv:2207.05221 , year=

    arXiv

  20. [28]

    Proceedings of the 36th International Conference on Machine Learning , pages=

    SelectiveNet: A Deep Neural Network with an Integrated Reject Option , author=. Proceedings of the 36th International Conference on Machine Learning , pages=. 2019 , url=

    2019

  21. [29]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics , pages=

    TruthfulQA: Measuring How Models Mimic Human Falsehoods , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics , pages=. 2022 , url=

    2022

  22. [30]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=. 2023 , url=

    2023

  23. [31]

    arXiv preprint arXiv:2310.01798 , year=

    Large Language Models Cannot Self-Correct Reasoning Yet , author=. arXiv preprint arXiv:2310.01798 , year=

    arXiv

  24. [32]

    When Can

    Kamoi, Ryo and Zhang, Yusen and Zhang, Nan and Han, Jiawei and Zhang, Rui , journal=. When Can. 2024 , url=

    2024

  25. [33]

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

    When Hindsight is Not 20/20: Testing Limits on Reflective Thinking in Large Language Models , author=. Findings of the Association for Computational Linguistics: NAACL 2024 , pages=. 2024 , url=

    2024

  26. [34]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , pages=

    Selective Question Answering under Domain Shift , author=. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , pages=. 2020 , url=

    2020

  27. [35]

    Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics , pages=

    The Art of Abstention: Selective Prediction and Error Regularization for Natural Language Processing , author=. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics , pages=. 2021 , url=

    2021

  28. [36]

    Advances in Neural Information Processing Systems , volume=

    Deep Gamblers: Learning to Abstain with Portfolio Theory , author=. Advances in Neural Information Processing Systems , volume=. 2019 , url=

    2019

  29. [37]

    2024 , url=

    Zhang, Hanning and Chen, Shizhe and Wang, Yaowei and Chen, Chao and Wang, Li and Liu, Zhiyuan and Sun, Maosong , booktitle=. 2024 , url=

    2024

  30. [38]

    Don't Hallucinate, Abstain: Identifying

    Feng, Shangbin and Shi, Weijia and Wang, Yizhong and Ding, Yifei and Balachandran, Vidhisha and Tsvetkov, Yulia , booktitle=. Don't Hallucinate, Abstain: Identifying. 2024 , url=

    2024

  31. [39]

    I Don't Know: Explicit Modeling of Uncertainty with an

    Cohen, Roi and Biran, Eden and Yoran, Ori and Geva, Mor and Globerson, Amir , journal=. I Don't Know: Explicit Modeling of Uncertainty with an. 2024 , url=

    2024

  32. [40]

    Shao, Zhihong and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Song, Junxiao and Bi, Xiao and Zhang, Haowei and Zhang, Mingchuan and Li, Y. K. and Wu, Y. and Guo, Daya , journal=. 2024 , url=

    2024

  33. [41]

    arXiv preprint arXiv:2308.01825 , year=

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

    arXiv

  34. [42]

    arXiv preprint arXiv:2305.04388 , year=

    Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting , author=. arXiv preprint arXiv:2305.04388 , year=

    arXiv

  35. [43]

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

    Tyen, Gladys and Mansoor, Hassan and C. Findings of the Association for Computational Linguistics: ACL 2024 , pages =. 2024 , url =

    2024

  36. [44]

    International Conference on Learning Representations , year =

    Training Language Models to Self-Correct via Reinforcement Learning , author =. International Conference on Learning Representations , year =

  37. [45]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages =

    Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback , author =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages =. 2023 , url =

    2023

  38. [46]

    Math-Shepherd: Verify and Reinforce

    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 =. Math-Shepherd: Verify and Reinforce. 2024 , url =

    2024

  39. [47]

    1970 , doi =

    On Optimum Recognition Error and Reject Tradeoff , author =. 1970 , doi =

    1970

  40. [48]

    Nature , volume =

    Detecting Hallucinations in Large Language Models Using Semantic Entropy , author =. Nature , volume =. 2024 , doi =

    2024

  41. [49]

    Teaching Large Language Models to Express

    Chen, Lida and Liang, Zujie and Wang, Xintao and Liang, Jiaqing and Xiao, Yanghua and Wei, Feng and Chen, Jinglei and Hao, Zhenghong and Han, Bing and Wang, Wei , booktitle =. Teaching Large Language Models to Express. 2025 , address =. doi:10.18653/v1/2025.knowllm-1.3 , url =

    2025 doi

  42. [50]

    Transactions on Machine Learning Research , year =

    Pich. Transactions on Machine Learning Research , year =

  43. [51]

    2026 , url =

    Zhai, Skylar and Liang, Jingcheng and Kang, Dongyeop , journal =. 2026 , url =

    2026

  44. [52]

    arXiv preprint arXiv:2604.12002 , year =

    Self-Distillation Zero: Self-Revision Turns Binary Rewards into Dense Supervision , author =. arXiv preprint arXiv:2604.12002 , year =

    arXiv

  45. [53]

    2026 , eprint=

    BARREL: Boundary-Aware Reasoning for Factual and Reliable LRMs , author=. 2026 , eprint=

    2026

  46. [54]

    International Conference on Learning Representations , year=

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

  47. [55]

    2021 , doi=

    Geva, Mor and Khashabi, Daniel and Segal, Elad and Khot, Tushar and Roth, Dan and Berant, Jonathan , journal=. 2021 , doi=

    2021

  48. [56]

    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. 2018 , url=

    2018

  49. [57]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle=

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle=. 2022 , url=

    2022

  50. [58]

    Mou, Yutao and Luo, Yuxiao and Zhang, Shikun and Ye, Wei , journal=

  51. [59]

    Xiong, Miao and Hu, Zhiyuan and Lu, Xinyang and Li, Yifei and Fu, Jie and He, Junxian and Hooi, Bryan , journal=. Can

  52. [60]

    The Internal State of an

    Azaria, Amos and Mitchell, Tom , journal=. The Internal State of an

  53. [61]

    arXiv preprint arXiv:2412.15115 , year=

    arXiv

  54. [62]

    Dubey, Abhimanyu and Jauhri, Abhinav and Pandey, Abhinav and Kadian, Abhishek and Al-Dahle, Ahmad and Letman, Aiesha and Mathur, Akhil and Schelten, Alan and Yang, Amy and Fan, Angela and others , journal=. The. 2024 , url=

    2024

  55. [63]

    Matthews, B. W. , journal=. Comparison of the Predicted and Observed Secondary Structure of. 1975 , doi=

    1975

  56. [64]

    2025 , eprint=

    Empowering LLMs with Logical Reasoning: A Comprehensive Survey , author=. 2025 , eprint=

    2025

  57. [65]

    I Don't Know

    Beyond "I Don't Know": Evaluating LLM Self-Awareness in Discriminating Data and Model Uncertainty , author=. 2026 , eprint=

    2026

  58. [66]

    2025 , eprint=

    Self-rewarding correction for mathematical reasoning , author=. 2025 , eprint=

    2025

  59. [67]

    2026 , eprint=

    TruthRL: Incentivizing Truthful LLMs via Reinforcement Learning , author=. 2026 , eprint=

    2026

  60. [68]

    2026 , eprint=

    KnowRL: Exploring Knowledgeable Reinforcement Learning for Factuality , author=. 2026 , eprint=

    2026

  61. [69]

    2025 , eprint=

    AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions , author=. 2025 , eprint=

    2025

  62. [70]

    2026 , eprint=

    BAPO: Boundary-Aware Policy Optimization for Reliable Agentic Search , author=. 2026 , eprint=

    2026

  63. [71]

    arXiv preprint arXiv:2504.18564 , year=

    DualBreach: Efficient Dual-Jailbreaking via Target-Driven Initialization and Multi-Target Optimization , author=. arXiv preprint arXiv:2504.18564 , year=

  64. [72]

    arXiv preprint arXiv:2607.15081 , year=

    DataShield: Uncovering Risky Fine-Tuning Data Across LLMs Through Consensus Subspace Alignment , author=. arXiv preprint arXiv:2607.15081 , year=

    arXiv

  65. [73]

    2025 , eprint=

    From Topology to Behavioral Semantics: Enhancing BGP Security by Understanding BGP's Language with LLMs , author=. 2025 , eprint=

    2025

  66. [74]

    arXiv , primaryClass=:2510.02422 , year=

    Dynamic Jailbreaking Attack , author=. arXiv , primaryClass=:2510.02422 , year=

    arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.