Pith. sign in

REVIEW 4 major objections 5 minor 31 references

Hide and Seek with LLMs: An Adversarial Game for Sneaky Error Generation and Self-Improving Diagnosis

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read An adversarial hide-and-seek game between two LLM roles lifts error-correction success by up to 31.4 percent.

desk verdict A genuinely new two-role adversarial training recipe for LLM error diagnosis, but the headline evaluation is self-referential and missing control baselines, so the diagnostic gains are not yet established. read the letter →

arxiv 2508.03396 v1 pith:KQU55627 submitted 2025-08-05 cs.AI

Rui Zou , Mengqi Wei , Yutao Zhu , Jirong Wen , Xin Zhao , Jing Chen This is my paper
classification cs.AI
keywords adversarialtrainingerrordiagnosislargelanguagemodelsmathematicalreasoningreinforcementlearningstealthyerrorsself-improvementhideandseekgame
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

This paper claims that large language models can learn to find and explain reasoning errors much better when a deliberately sneaky error generator and a diagnostician are trained against each other in a loop, rather than when either is trained alone or against a static error set. The proposed Hide and Seek Game (HSG) alternates reinforcement-learning updates between a Sneaky role, rewarded for producing wrong answers that are hard to catch, and a Diagnosis role, rewarded for spotting them and for making corrections succeed. On three math reasoning benchmarks, diagnostics produced by the trained Diagnosis role raise correction success rates by 16.8% to 31.4% over each model's own untrained diagnostics, across model sizes from Qwen3-4B to GPT-4o. The paper also reports that HSG-generated errors resist correction more than errors from an LLM-rater baseline or from plain RL. If this claim holds, adversarial co-evolution is a practical route to stronger error analysis and correction in LLMs.

What carries the argument

The load-bearing mechanism is the adversarial training loop itself, organized as a hide-and-seek game between two roles played by the same underlying LLM (parameter-shared, prompt-driven). Sneaky samples wrong answers and is rewarded by the hierarchical function $R(r_{\mathrm{main}},r_{\mathrm{secondary}}) = \max(r_{\mathrm{main}}, \tau)[\beta + (1-\beta)r_{\mathrm{secondary}}]$, where the main reward favors incorrectness and the auxiliary reward penalizes diagnoses and corrections that succeed. Diagnosis is rewarded for accurately judging whether a solution is wrong and for producing reports that let a fixed correction model fix it. Training alternates between the two roles under Group Relative Policy Optimization (GRPO), and the reward to each role is fed with the other role's performance so that harder-to-catch errors and sharper diagnostics emerge together.

What would settle it

Run the trained Diagnosis role on fixed, independently collected reasoning errors and compare correction success against an untrained diagnostician; if D* shows no advantage on errors it did not co-evolve with, the claim of general diagnostic improvement collapses.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that error stealth and diagnostic precision can be improved together through adversarial co-evolution. A Sneaky generator and a Diagnosis model are optimized in alternating GRPO steps, with a hierarchical reward function that makes the generator's reward depend on the diagnostician's failure and on failed correction, and the diagnostician's reward depend on both correct detection and successful correction. The trained Diagnosis role D* produces diagnostic reports that, when given to five different LLMs, improve average correction success from 35.61%-53.21% (using each model's own diagnostics) to 65.84%-70.01% (using D*), an absolute gain of 16.8-31.4 percentage points. The trained Sneaky role generates errors that five baseline models fail to correct 56.17% of the time, versus 38.51% for a reward-model-based adversarial baseline and 27.43% for RL alone. The paper further reports that 92.6% of HSG's errors contain neither a shifted final answer nor an explicitly contaminated reasoning chain, the two trivial error types that dominate the baselines.

Load-bearing premise

The evaluation of diagnostic quality uses only errors generated by HSG's own Sneaky role, so the measured improvement may reflect specialization to that co-adapted error distribution rather than a general improvement in error diagnosis.

Editorial extensions

If this is right

  • A diagnostic report produced by the trained Diagnosis role improves correction accuracy for every model tested, including models larger and stronger than the trained diagnostician itself.
  • Errors produced by the Sneaky role are substantially harder for five different LLMs to correct than errors from RL-only or LLM-rater adversarial training.
  • Adversarial training filters out the two trivial error patterns (changed final answer, explicit reasoning mistake) that dominate non-adversarial baselines.
  • The released dataset of sneaky errors with diagnostic annotations provides a benchmark for measuring diagnostic quality beyond correctness of the final answer.

Reading between the lines

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

  • A natural next test is whether D* also outperforms baseline diagnosticians on independent static error sets; that would separate general diagnostic skill from specialization to the co-adapted error distribution.
  • The same two-role loop could extend to domains where a corrected output is checkable, such as code that must compile or pass tests, using correction success as the shared reward signal.
  • The Sneaky role could be reused as an adversarial data generator to harden verifiers and reward models, since it produces errors that current raters judge as plausible.
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. The paper introduces Hide and Seek Game (HSG), an adversarial training framework in which a Sneaky role generates deceptive mathematical reasoning errors and a Diagnosis role produces diagnostic reports, with rewards that include correction success of a fixed correction model. The authors train Qwen3-4B with GRPO and evaluate on GSM8K, MATH, and NuminaMATH-TIR, reporting that HSG-trained diagnostics (D*) improve correction success by 16.8%-31.4% over the same models' untrained diagnostics (D), and that HSG-generated errors are harder to correct than errors from RL-only or LLM-rater baselines. The paper also announces a dataset of stealthy errors and diagnostic annotations.

Significance. If the central claim is established, the HSG framework would be a useful contribution: it offers a concrete adversarial training recipe for error generation and diagnosis, and the released benchmark could support future work on diagnostic quality. The paper is commendably explicit about its reward definitions, training details, and baseline training methods, and the stealthiness comparison across five base models is a useful empirical probe. However, the evaluation as presented does not yet separate genuine diagnostic skill from answer leakage and co-adaptation to a single error generator. Because the headline improvement is the paper's main contribution, the needed controls are substantive rather than cosmetic; the framework itself is plausible and worth further validation.

major comments (4)
  1. [§Framework, Eq. (7)-(9); §Diagnosis Performance Study, Table 1] The diagnostic role D is conditioned on the reference answer atruth (sampling step: aD^(i) ~ πθ_D(·|atruth, aS^(i))), and the collaborative reward R_collaborative_D = R[R_D, rcorr(aC)] rewards D whenever the corrected answer aC matches atruth. Because aD is the only channel through which D influences aC, the reward-maximizing policy can simply restate the reference answer or a full correct solution, without localizing the faulty step. Table 1 compares D* against each model's own diagnostic D, but the paper does not state whether D also receives atruth when generating its report. Please add an ablation in which atruth is withheld from D at diagnosis time, or report error-localization metrics on the faulty step, and include a trained-diagnostic baseline trained with identical rewards on static errors.
  2. [§Diagnosis Performance Study; §Stealthiness Analysis] The diagnostic evaluation is performed on errors produced by the same HSG Sneaky policy that co-evolved with D*, so the reported 16.8%-31.4% improvement may reflect specialization to the co-adapted error distribution rather than a general improvement in error diagnosis. The paper reports no evaluation on independent error benchmarks (e.g., MathClean, CriticBench, or human-annotated error sets) or on errors generated by the RL-only and LLM-rater baselines. Please add cross-generator and static-error evaluations; this is necessary to support the abstract's claim of improved diagnostic ability rather than improved fit to a single generator.
  3. [§Experimental Setup, Training Details] The final checkpoint is selected based on the lowest correction success rate ACCcorr, i.e., the most deceptive model, but the paper does not state whether this selection is performed on a held-out validation split or on the test set. If the selection criterion uses the same distribution as the test sets in Table 1 and Figure 4, the reported stealthiness and correction-failure numbers are optimistically biased. Please describe the selection procedure, report the validation split, and provide standard deviations across independent runs.
  4. [§Diagnosis Performance Study, Table 2] The GPT-4o win-rate validation is weak as presented: win rates range from 9.57% to 26.60%, with ties in the majority of cases, and the text states that the trends 'closely match' the correction-accuracy improvements. With such low win rates and no reported sample size, the preference signal does not robustly corroborate the correction-accuracy gains. Please report the number of samples, the preference prompt, and inter-judge agreement, or replace this with a more discriminative diagnostic-quality metric such as error-step localization F1.
minor comments (5)
  1. [§Reinforcement Learning, Eq. (8)] Equation (8) is typeset incompletely; the definition of rcorr(aC) needs to be written out explicitly rather than split across lines and the equation number.
  2. [§Experimental Setup, Training Details] The hyperparameters τ, β, Lmin/Lmax, and the two KL regularization coefficients are given without sensitivity analysis; a short ablation or a reference to a sensitivity study would help establish that the results are not artifacts of these choices.
  3. [Abstract; Conclusion] The paper announces a public dataset of stealthy errors and diagnostics, but no URL, repository, or access instructions are provided; please include the release mechanism.
  4. [§Framework, Role D Sampling] In the Role D Sampling paragraph, the correction model is written as πcorrect(aD^(i)) without the dependence on aS, whereas earlier it is defined as πcorrect(aS, aD); this notation should be made consistent.
  5. [§Diagnosis Performance Study, Table 2] The table does not report the number of test examples used for the win-rate evaluation; please include sample counts so the reader can gauge the precision of the reported percentages.

Circularity Check

2 steps flagged · score 6.0 of 10

The headline diagnostic gain is measured by the same correction-success reward used to train D*, and D* is conditioned on the reference answer, so Table 1 may reflect answer leakage and in-distribution specialization rather than general diagnostic skill.

  1. fitted input called prediction [Section 'Feedback Rewards' (Eq. 9) and 'Diagnosis Performance Study']
    "Collaborative Feedback for D ... Auxiliary reward: whether the diagnosis enables successful correction. The collaborative feedback reward is: RcollaborativeD = R [RD, rcorr(aC)]"

    D* is trained by GRPO to maximize RcollaborativeD, whose secondary term is rcorr(aC), i.e., correction success. The paper's headline evaluation metric is ACCcorr, also correction success. Thus the reported 16.8%–31.4% 'improvement in diagnosis' is the training objective itself, evaluated on the same HSG-generated error distribution. No independent diagnostic metric (e.g., localization accuracy on static errors or errors from other generators) is reported, so the 'prediction' is forced by construction.

  2. self definitional [Section 'Reward for D' and 'Role D Sampling']
    "D compares atruth and aS, outputting a diagnostic report aD identifying errors. ... a(i)D ∼ πθD(·|atruth, a(i)S)"

    The diagnostic model is conditioned on the reference answer atruth and is rewarded by rcorr(aC) for any report that allows the fixed corrector to reach the gold answer. The simplest policy maximizing this reward is to restate atruth (or a full correct solution) rather than to pinpoint the faulty step. Table 1 credits such answer-bearing reports as 'diagnostics', so the reported gain can be reproduced without any improvement in error localization. The paper does not state whether the untrained baseline D also receives atruth, making the comparison asymmetric.

full rationale

The central claim that HSG 'significantly boosts error diagnosis' rests on Table 1, which compares correction success using HSG-trained diagnostics (D*) against each model's own untrained diagnostics (D). Two reductions make this claim circular rather than merely imperfect. First, the training reward for D* explicitly includes correction success (Eq. 9: RcollaborativeD = R[RD, rcorr(aC)]), and the evaluation metric is also correction success (ACCcorr). The reported improvement is therefore the training objective evaluated on the same adversarial error distribution; it is not an independent measurement of diagnostic localization. Second, D* is sampled conditioned on the reference answer atruth, so a report that simply contains or restates atruth maximizes the correction reward. The paper never establishes that D* produces error-localizing diagnoses rather than answer-bearing copies, and it does not disclose whether the baseline D has the same access to atruth. The paper's own statement that 'D∗ is trained on stealthier errors generated by the S role under HSG' confirms that both the training and test errors come from the same co-adapted S*–D* pair, so the comparison is in-distribution for D* and out-of-distribution for the untrained D. The stealthiness analysis (Fig. 4) uses external baseline correctors and is less circular, which prevents a score of 8 or higher. No load-bearing self-citation or uniqueness-imported-from-authors pattern is present. Overall, the primary diagnostic claim is substantially reduced by construction, though the framework's error-generation results retain some independent content.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No invented physical entities. The framework introduces two model roles (Sneaky, Diagnosis) and a reward shaping function, all within standard ML training. Free parameters are hand-set hyperparameters. Key dependencies are GRPO, the fixed corrector's reliability, and the proxy metric.

free parameters (4)
  • tau (hierarchical reward lower bound) = 0.05
    Chosen by hand to keep exploration when the main reward is zero; affects the reward scale for both roles.
  • beta (reward weighting coefficient) = 0.6
    Chosen by hand to balance main and auxiliary rewards; directly shapes the training signal.
  • Lmin/Lmax (answer length bounds) = 50/600
    Hand-set to penalize overly short or long answers; affects the format rewards.
  • KL regularization coefficients = 0.01 (S), 0.04 (D)
    Hand-set coefficients for KL divergence in GRPO; standard but not optimized.
assumptions (4)
  • standard math GRPO provides a valid policy-gradient estimator with the group-relative advantage.
    The training objectives in Eqs. 11-15 rely on Shao et al. 2024 without modification.
  • domain assumption The fixed correction model p_correct produces reliable correction-success labels.
    Both the S and D rewards depend on rcorr(aC); the identity and quality of p_correct are not disclosed.
  • domain assumption Correction success rate is a valid proxy for diagnostic quality.
    ACCcorr is the headline metric in Table 1; diagnosing an error well should enable correction, but this conflates diagnosis quality with the corrector's ability.
  • ad hoc to paper The hierarchical reward form R = max(rmain, tau)[beta + (1-beta)rsecondary] is a sensible, stable training signal.
    This function is introduced without derivation or comparison to other reward shapes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hide and Seek with LLMs: An Adversarial Game for Sneaky Error Generation and Self-Improving Diagnosis." pith.science (2026). https://pith.science/paper/KQU55627

@misc{pith2026250803396,
  author       = {Pith},
  title        = {Pith review of: Hide and Seek with LLMs: An Adversarial Game for Sneaky Error Generation and Self-Improving Diagnosis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KQU55627}},
  note         = {Machine review of arXiv:2508.03396}
}
read the original abstract

Large Language Models (LLMs) excel in reasoning and generation across domains, but still struggle with identifying and diagnosing complex errors. This stems mainly from training objectives that prioritize correct answers, limiting exposure to and learning from errors. While recent studies have begun to address this by introducing error signals, most rely on shallow, static errors, restricting improvement in deep diagnostic ability. To overcome this, we propose Hide and Seek Game (HSG), a dynamic adversarial framework for error generation and diagnosis, and evaluate it on mathematical problem-solving. HSG involves two adversarial roles: Sneaky, which "hides" by generating subtle, deceptive reasoning errors, and Diagnosis, which "seeks" to accurately detect them. Through adversarial co-evolution, both error stealth and diagnostic precision are enhanced. Experiments on several math reasoning tasks show that HSG significantly boosts error diagnosis, achieving 16.8\%--31.4\% higher accuracy than baselines like GPT-4o. We also release a challenging dataset of deceptive errors and diagnostic annotations as a benchmark for future research.

Figures

Figures reproduced from arXiv: 2508.03396 by the authors.

Figure 1
Figure 1. Comparison of three error utilization strategies. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Training pipeline of HSG. swers containing stealthy errors, and D, which detects these errors. The overall training pipeline is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Prompts for roles S and D. Reward for D D compares atruth and aS, outputting a di￾agnostic report aD identifying errors. As aS may be correct early in training, we define a diagnosis recognition function: Γdiagnosis (aD, rcorr(aS)) =  1, if match 0, otherwise (6) “if match” denotes that aD correctly reflects the correctness of aS. The reward for D is: RD = R (Γdiagnosis (aD, rcorr(aS)), R[rformat(aD), rlength(aD)])… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Correction difficulty (stealthiness) for sneaky an [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Examples of sneaky answers generated by the three methods. The first two display Type A and Type B errors. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Error type distribution in sneaky answers from [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 16 canonical work pages

  1. [1]

    S.; Kazi, S.; Sourabh, V.; et al

    Bao, F.; Li, M.; Qu, R.; Luo, G.; Wan, E.; Tang, Y.; Fan, W.; Tamber, M. S.; Kazi, S.; Sourabh, V.; et al. 2025. FaithBench: A Diverse Hallucination Benchmark For Summarization By Modern LLMs. In Proceedings of Association For Computational Linguistics, 448--461

    work page 2025

  2. [2]

    Bewersdorff, A.; Se ler, K.; Baur, A.; Kasneci, E.; and Nerdel, C. 2023. Assessing Student Errors In Experimentation Using Artificial Intelligence And Large Language Models: A Comparative Study With Human Raters. Computers and Education: Artificial Intelligence, 5: 100177

    work page 2023

  3. [3]

    Chang, Y.; Wang, X.; Wang, J.; Wu, Y.; Yang, L.; Zhu, K.; Chen, H.; Yi, X.; Wang, C.; Wang, Y.; et al. 2024. A Survey On Evaluation Of Large Language Models. ACM Transactions On Intelligent Systems And Technology, 15(3): 1--45

    work page 2024

  4. [4]

    Cobbe, K.; Kosaraju, V.; Bavarian, M.; Chen, M.; Jun, H.; Kaiser, L.; Plappert, M.; Tworek, J.; Hilton, J.; Nakano, R.; et al. 2021. Training Verifiers To Solve Math Word Problems. arXiv Preprint arXiv:2110.14168

    arXiv 2021

  5. [5]

    Du, D.; Liu, S.; Yang, T.; Chen, S.; and Li, Y. 2025. UloRL: An Ultra-Long Output Reinforcement Learning Approach For Advancing Large Language Models' Reasoning Abilities. arXiv Preprint arXiv:2507.19766

    arXiv 2025

  6. [6]

    Fan, Y.; Jiang, F.; Li, P.; and Li, H. 2023. Grammargpt: Exploring Open-Source LLMs For Native Chinese Grammatical Error Correction With Supervised Fine-Tuning. In Proceedings Of Natural Language Processing And Chinese Computing, 69--80

    work page 2023

  7. [7]

    Gou, Z.; Shao, Z.; Gong, Y.; Shen, Y.; Yang, Y.; Duan, N.; and Chen, W. 2024. CRITIC: Large Language Models Can Self-Correct With Tool-Interactive Critiquing. In Proceedings of International Conference on Learning Representations

    work page 2024

  8. [8]

    L.; Liu, Y.; Shang, N.; Sun, Y.; Zhu, Y.; Yang, F.; and Yang, M

    Guan, X.; Zhang, L. L.; Liu, Y.; Shang, N.; Sun, Y.; Zhu, Y.; Yang, F.; and Yang, M. 2025. rStar-Math: Small LLMs Can Master Math Reasoning With Self-Evolved Deep Thinking. In Proceedings Of International Conference On Machine Learning

    work page 2025

Show all 31 references
  1. [9]

    Gulati, A.; Miranda, B.; Chen, E.; Xia, E.; Fronsdal, K.; Dumont, B. d. M.; and Koyejo, S. 2025. Putnam-AXIOM: A Functional And Static Benchmark For Measuring Higher Level Mathematical Reasoning In LLMs. In Proceedings Of International Conference On Machine Learning

    2025

  2. [10]

    Hendrycks, D.; Burns, C.; Kadavath, S.; Arora, A.; Basart, S.; Tang, E.; Song, D.; and Steinhardt, J. 2021. Measuring Mathematical Problem Solving With The Math Dataset. arXiv Preprint arXiv:2103.03874

    2021 arXiv

  3. [11]

    S.; Yu, A.; Song, X.; and Zhou, D

    Huang, J.; Chen, X.; Mishra, S.; Zheng, H. S.; Yu, A.; Song, X.; and Zhou, D. 2024. Large Language Models Cannot Self-Correct Reasoning Yet. In Proceedings of the International Conference on Learning Representations

    2024

  4. [12]

    H.; Chen, Y.; Edwards, H.; Leike, J.; McAleese, N.; and Burda, Y

    Kirchner, J. H.; Chen, Y.; Edwards, H.; Leike, J.; McAleese, N.; and Burda, Y. 2024. Prover-Verifier Games Improve Legibility Of Llm Outputs. arXiv Preprint arXiv:2407.13692

    2024 arXiv

  5. [13]

    Q.; Shen, Z.; et al

    Li, J.; Beeching, E.; Tunstall, L.; Lipkin, B.; Soletskyi, R.; Huang, S.; Rasul, K.; Yu, L.; Jiang, A. Q.; Shen, Z.; et al. 2024. Numinamath: The Largest Public Dataset In Ai4maths With 860K Pairs Of Competition Math Problems And Solutions. Hugging Face Repository, 13(9): 9

    2024

  6. [14]

    Liang, H.; Qiang, M.; Li, Y.; He, Z.; Guo, Y.; Zhu, Z.; Zhang, W.; and Cui, B. 2025. Mathclean: A Benchmark For Synthetic Mathematical Data Cleaning. arXiv Preprint arXiv:2502.19058

    2025 arXiv

  7. [15]

    Lin, Z.; Gou, Z.; Liang, T.; Luo, R.; Liu, H.; and Yang, Y. 2024. CriticBench: Benchmarking LLMs For Critique-Correct Reasoning. In Proceedings of Association For Computational Linguistics, 1552--1587

    2024

  8. [16]

    Y.; Zeng, L.; Xiao, Y.; He, J.; Liu, J.; Wang, C.; Yan, R.; Shen, W.; Zhang, F.; Xu, J.; et al

    Liu, C. Y.; Zeng, L.; Xiao, Y.; He, J.; Liu, J.; Wang, C.; Yan, R.; Shen, W.; Zhang, F.; Xu, J.; et al. 2025. Skywork-Reward-V2: Scaling Preference Data Curation Via Human-Ai Synergy. arXiv Preprint arXiv:2507.01352

    2025 arXiv

  9. [17]

    Maity, S.; and Saikia, M. J. 2025. Large Language Models In Healthcare And Medical Applications: A Review. Bioengineering, 12(6): 631

    2025

  10. [18]

    Ouyang, L.; Wu, J.; Jiang, X.; Almeida, D.; Wainwright, C.; Mishkin, P.; Zhang, C.; Agarwal, S.; Slama, K.; Ray, A.; et al. 2022. Training Language Models To Follow Instructions With Human Feedback. Advances In Neural Information Processing Systems, 35: 27730--27744

    2022

  11. [19]

    D.; Ermon, S.; and Finn, C

    Rafailov, R.; Sharma, A.; Mitchell, E.; Manning, C. D.; Ermon, S.; and Finn, C. 2023. Direct Preference Optimization: Your Language Model Is Secretly A Reward Model. Advances In Neural Information Processing Systems, 36: 53728--53741

    2023

  12. [20]

    Schulman, J.; Wolski, F.; Dhariwal, P.; Radford, A.; and Klimov, O. 2017. Proximal Policy Optimization Algorithms. arXiv Preprint arXiv:1707.06347

    2017 arXiv

  13. [21]

    Shalev-Shwartz, S.; Shammah, S.; and Shashua, A. 2017. On A Formal Model Of Safe And Scalable Self-Driving Cars. arXiv Preprint arXiv:1708.06374

    2017 arXiv

  14. [22]

    Shao, Z.; Wang, P.; Zhu, Q.; Xu, R.; Song, J.; Bi, X.; Zhang, H.; Zhang, M.; Li, Y.; et al. 2024. Deepseekmath: Pushing The Limits Of Mathematical Reasoning In Open Language Models. arXiv Preprint arXiv:2402.03300

    2024 arXiv

  15. [23]

    P.; and Mak, T

    Tyen, G.; Mansoor, H.; C a rbune, V.; Chen, Y. P.; and Mak, T. 2024. LLMs Cannot Find Reasoning Errors, But Can Correct Them Given The Error Location. In Proceedings of Association For Computational Linguistics, 13894--13908

    2024

  16. [24]

    Wang, L.; Lyu, C.; Ji, T.; Zhang, Z.; Yu, D.; Shi, S.; and Tu, Z. 2023. Document-Level Machine Translation With Large Language Models. In Proceedings of Empirical Methods In Natural Language Processing, 16646--16661

    2023

  17. [25]

    S.; and Wen, Q

    Wang, S.; Xu, T.; Li, H.; Zhang, C.; Liang, J.; Tang, J.; Yu, P. S.; and Wen, Q. 2024. Large Language Models For Education: A Survey And Outlook. arXiv Preprint arXiv:2403.18105

    2024 arXiv

  18. [26]

    V.; Zhou, D.; et al

    Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Xia, F.; Chi, E.; Le, Q. V.; Zhou, D.; et al. 2022. Chain-Of-Thought Prompting Elicits Reasoning In Large Language Models. Advances In Neural Information Processing Systems, 35: 24824--24837

    2022

  19. [27]

    Ying, Z.; Liu, A.; Liu, X.; and Tao, D. 2024. Unveiling The Safety Of Gpt-4o: An Empirical Study Using Jailbreak Attacks. arXiv Preprint arXiv:2406.06302

    2024 arXiv

  20. [28]

    Zhang, Y.; Li, M.; Long, D.; Zhang, X.; Lin, H.; Yang, B.; Xie, P.; Yang, A.; Liu, D.; Lin, J.; et al. 2025. Qwen3 Embedding: Advancing Text Embedding And Reranking Through Foundation Models. arXiv Preprint arXiv:2506.05176

    2025 arXiv

  21. [29]

    Zhu, Z.; Yang, Y.; and Sun, Z. 2024. Halueval-Wild: Evaluating Hallucinations Of Language Models In The Wild. arXiv Preprint arXiv:2403.04307

    2024 arXiv

  22. [30]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  23. [31]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.