Pith. sign in

REVIEW 5 major objections 6 minor 4 cited by

Revisiting Overthinking in Long Chain-of-Thought from the Perspective of Self-Doubt

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper shows that self-doubt, not length alone, drives overthinking in reasoning LLMs, and that a one-line prompt asking the model to check input validity before answering cuts token use by about 37% while keeping or improving accuracy.

desk verdict Useful prompt-based token reduction with a plausible but unvalidated self-doubt mechanism; the causal claim needs a be-concise control and a human-checked judge before it supports the framing. read the letter →

arxiv 2505.23480 v1 pith:QBAYXLYY submitted 2025-05-29 cs.CL

Keqin Peng , Liang Ding , Yuanxin Ouyang , Meng Fang , Dacheng Tao This is my paper · ORCID
classification cs.CL
keywords overthinkingself-doubtlongchain-of-thoughtreasoninglargelanguagemodelspromptingmathematicalmissingpremisesLLM-as-a-judge
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

The paper sets out to show that a large fraction of overthinking in long chain-of-thought reasoning is self-doubt: the model re-verifies steps it already knows are correct, spending tokens after the answer is settled. It quantifies this by having a judge model classify reasoning paths into three categories, finding overthinking in over 80% of responses and self-doubt in nearly 60% of the hardest math set. It then tests a one-line prompt that asks the model to first check whether the question has all needed information and to answer with the minimum tokens required. Across four reasoning models, this cuts average output length by about 37% on well-defined math tasks while slightly improving accuracy, and sharply improves the model's willingness to abstain when premises are missing. If the claim is right, a cheap intervention rather than retraining can reduce a major source of wasted computation in reasoning models.

What carries the argument

The load-bearing object is the three-way classification of a generated reasoning path into self-doubt, overthinking without self-doubt, and non-overthinking, assigned by a judge model to the second half of each response split at paragraph boundaries. The intervention itself is a prompt template: before reasoning deeply, the model must check whether all necessary information is available; if anything is missing or ambiguous, it must say so first, otherwise it should answer with the minimum number of tokens required. This prompt is what carries the causal claim, because the paper attributes the length reduction to the reduction in re-verification rather than to the brevity instruction alone.

What would settle it

A decisive experiment would be a brevity-only control: replace the prompt's validity-check clause with a neutral sentence while keeping 'minimum number of tokens.' If the length reduction survives, the paper's claim that reduced self-doubt causes the saving is unsupported. A second check would compare the judge's self-doubt labels to human labels on a random sample of 200 responses per dataset.

Watch

Extended reading notes

Core claim

The central discovery is that self-doubt is a measurable, dominant component of overthinking in reasoning LLMs, and that it can be reduced by a prompt that makes the model critically evaluate the input before committing to a long chain of thought. The paper supports this with a three-way classification of responses (self-doubt, overthinking without self-doubt, non-overthinking) and shows that self-doubt is especially prevalent on harder problems, reaching 59.5% on Math-500. The proposed prompt consistently reduces response length by roughly a third, maintains or improves accuracy on GSM8K, GSM8K-Zero, and Math-500, and improves abstain rates on four missing-premise datasets. The authors interpret the result as evidence that apparent sycophancy toward the user input causes the model to keep checking its own work.

Load-bearing premise

The load-bearing premise is that the automated judge, applied to the second half of each response after splitting at paragraph breaks, correctly identifies self-doubt; the paper reports no human agreement check.

Editorial extensions

If this is right

  • On well-defined math tasks, the prompt cuts response length by about 37% on average across the four tested reasoning models while keeping accuracy roughly flat or better (e.g., +16.7 and +14.7 points on GSM8K-Zero for the 32B and 70B DeepSeek models).
  • On missing-premise datasets, the prompt raises abstain rates by up to nearly 40 percentage points and cuts token use by more than half on most settings, so models stop guessing when the question is under-specified.
  • The number of reasoning steps drops on all three math datasets, indicating the intervention changes the reasoning process itself, not just the wording of the final response.
  • Because the method is a zero-shot prompt addition, it can be deployed on existing reasoning models without fine-tuning or architectural changes.

Reading between the lines

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

  • Not tested here: a brevity-only control prompt (minimum tokens without the validity check) would isolate whether the effect comes from reducing self-doubt or just from instructing brevity.
  • The paper's math-only scope leaves open whether the same prompt curbs self-doubt in commonsense reasoning, planning, or multimodal tasks; a transfer test would settle it.
  • If the judge labels are correct, training-time rewards could penalize re-verification of already-correct steps, making the efficiency gain a permanent model property rather than a prompt effect.
  • Human annotation of a small sample of classified responses would show whether 'self-doubt' is a stable, recognizable behavior across models or a byproduct of the paragraph-splitting heuristic.
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

5 major / 6 minor

Summary. The paper quantifies overthinking in long chain-of-thought (CoT) reasoning for reasoning LLMs by labeling reasoning paths as self-doubt, overthinking without self-doubt, or non-overthinking using Qwen2.5-72B-Instruct as an LLM judge. It reports that self-doubt is a major component of overthinking (e.g., 59.5% on MATH-500 for DeepSeek-R1-Distill-Qwen-32B) and proposes a simple prompting intervention: ask the model to check whether the input question has all necessary information, then answer with the minimum number of tokens. Experiments on GSM8K, GSM8K-Zero, MATH-500, and four missing-premise datasets across four RLLMs show large response-length reductions and improved abstain rates, with an average length reduction of 37.1% and an average accuracy change of +3.6% on well-defined tasks.

Significance. If the mechanistic claims were established, the paper would be practically valuable: a one-line prompt that reduces token usage by about one-third while preserving accuracy on math tasks and sharply improving missing-premise abstention could reduce inference cost in deployed RLLMs. The paper is also useful as a reproducible starting point because it specifies models, datasets, prompt templates, and judge templates. However, the central causal claim—that self-doubt drives overthinking and that the prompt reduces overthinking by reducing self-doubt—currently rests on an unvalidated LLM-based metric and is confounded by the prompt's explicit token-minimization instruction. The practical length-reduction result is credible, but the mechanistic interpretation and the "significant improvement" language are not yet supported.

major comments (5)
  1. [Section 2 / Appendix A, Table 5] The self-doubt metric is computed by Qwen2.5-72B-Instruct on only the second half of each response split by "\n\n", with no human validation, inter-annotator agreement, or alternative metric reported. The Limitations section concedes that the judge "may misunderstand nuanced reasoning patterns or miss subtle verification steps." Because Figure 1 and Table 4 are the only quantitative evidence that self-doubt is a major driver of overthinking and that the prompt reduces self-doubt, this unvalidated measurement is load-bearing. Please report human agreement on a sample and validate the paragraph-splitting heuristic.
  2. [Section 3 / Table 2] The treatment prompt explicitly instructs the model to "answer it with the minimum number of tokens required." The large length reductions in Tables 1 and 3, and the drop in self-doubt in Table 4 as judged by a rubric that flags "unnecessarily long reasoning," can be explained purely as compliance with a brevity instruction. A control with a conciseness-only prompt (no premise-validity check) is needed to attribute the effect to the critique step or to reduced self-doubt rather than to the explicit token-minimization directive.
  3. [Section 5 / Table 1] The abstract and Section 5 claim "significant improvements across nearly all datasets," but no significance tests are reported and several accuracy numbers decrease (e.g., DS-14B MATH-500 92.8 to 92.2, DS-70B MATH-500 92.4 to 91.6, Qwen3-32B MATH-500 97.2 to 96.8). The term "significant" should be supported by paired tests or confidence intervals, or the claim should be revised to reflect mixed accuracy changes with consistent length reduction.
  4. [Section 5 / Table 3] The missing-premise abstain-rate gains are partly a direct effect of the instruction "If any key data is missing or ambiguous, explicitly state that first," so the large abstain-rate improvements do not by themselves demonstrate improved critique ability or reduced self-doubt. Please include a control that gives the same missing-premise clue without the conciseness instruction, and report non-abstaining answer quality so that the MiP results are not evaluated only on the instructed behavior.
  5. [Section 3 / Section 6, Table 4] The causal narrative that self-doubt drives overthinking and that the prompt works by reducing self-doubt is not directly tested. Table 4 shows the self-doubt ratio increases on GSM8K (5.9 to 17.2), which the paper attributes to "early verification" without supporting evidence; this inconsistency, together with the lack of any manipulation of model confidence or user-deference, weakens the mechanistic claim. At minimum, report per-dataset examples and a confidence-related metric that is not defined by the same brevity criterion as the intervention.
minor comments (6)
  1. [Section 4 / Section 5] Section 5 contains "Results on 1" instead of "Table 1," and Section 4 has "statistic metric" instead of "statistical metric."
  2. [Figure 1] The caption misspells "Deepseek" and "Non-Overtinking."
  3. [Section 3 / Table 2] The text describes a two-stage process ("first prompt the model... then instruct"), but Table 2 shows a single user message containing both instructions; please clarify whether the intervention is one stage or two.
  4. [Table 4] The table does not define the denominator or report confidence intervals for the self-doubt ratio; specify whether it is computed over all examples or only over overthinking examples.
  5. [Appendix A] Decoding details for the Qwen2.5-72B-Instruct judge (temperature, number of samples, version) are omitted; since the judge is a core measurement instrument, these details should be provided for reproducibility.
  6. [Table 6] The baseline and "Ours" responses are nearly identical in their opening sentences; please indicate whether the example is representative and how it was selected.

Circularity Check

2 steps flagged · score 6.0 of 10

Self-doubt reduction is measured by a rubric whose definition the prompt explicitly enforces, so the central mechanistic claim is partly circular; the engineering result remains self-contained.

  1. self definitional [Section 2 (SD definition) and Section 3 / Table 2 (prompt)]
    ""Overthinking with Self-Doubt (SD): The model's answer includes unnecessarily long reasoning and repeatedly verifies information that is already correct." ... "otherwise, answer it with the minimum number of tokens required.""

    The construct 'self-doubt' is defined by the presence of 'unnecessarily long reasoning' plus repeated verification of already-correct content. The intervention prompt explicitly instructs the model to 'answer it with the minimum number of tokens required,' which directly removes the first defining condition. Therefore the reductions in response length and judge-labeled self-doubt reported in Tables 1, 3, and 4 are entailed by the prompt instruction itself; they cannot serve as independent evidence that the mechanism 'reduced self-doubt' is what drives the improvement. The causal attribution 'alleviate overthinking by avoiding self-doubt' restates the instruction rather than testing an independent mechanism.

  2. self definitional [Section 6, Table 4 paragraph]
    ""As shown in Table 4, our method reduces the self-doubt ratio in two out of three datasets, with a particularly notable 23.8% reduction on the complex dataset MATH-500, demonstrating our method can effectively reduce self-doubt.""

    This 'self-doubt ratio' is computed with the LLM judge whose rubric is the Section 2 definition of self-doubt. Because the treated prompt explicitly demands the minimum number of tokens, the judge's criterion 'unnecessarily long reasoning' will almost mechanically flag fewer treated responses as self-doubt. The reported reduction therefore reflects the prompt's length constraint, not an independently measured change in a distinct psychological or mechanistic state. The paper's own Limitations section concedes the judge 'may misunderstand nuanced reasoning patterns or miss subtle verification steps,' and no human-validated alternative is provided, so the claim that self-doubt was the operative cause is not secured.

full rationale

The practical engineering result—that a prompt containing an explicit 'minimum number of tokens required' instruction reduces output length while often preserving or improving accuracy and abstention rates on the tested benchmarks—is self-contained and credible. The paper also correctly separates well-defined from missing-premise tasks and reports consistent gains. However, the central mechanistic claim (self-doubt significantly contributes to overthinking, and the proposed method works by reducing self-doubt) is partially circular: self-doubt is operationalized with a judge rubric that defines it as unnecessarily long reasoning with repeated verification, and the intervention directly instructs the model to use the minimum number of tokens. The measured reduction in self-doubt is thus aligned with the prompt by construction. No load-bearing self-citation or imported uniqueness theorem is present; the descriptive proportions in Figure 1 are not themselves circular, though the causal language around them is under-supported. Overall, the paper contains a significant definitional reduction in its key interpretative step, but the independent benchmark results prevent a higher score.

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

The paper introduces no new theoretical entities or fitted constants. Its central quantitative claims rest on an unvalidated LLM judge, a paragraph-splitting heuristic, and an analogy-based motivation. The only 'free parameter' in a loose sense is the prompt wording itself, which is chosen by hand and directly forces shorter outputs.

assumptions (3)
  • domain assumption The LLM judge Qwen2.5-72B-Instruct reliably classifies reasoning paths into self-doubt, overthinking-without-self-doubt, and non-overthinking categories.
    Section 2 uses this judge to produce Figure 1 and Table 4; no human validation, inter-annotator agreement, or alternative metric is reported.
  • domain assumption Splitting responses at the paragraph delimiter and scoring only the second segment isolates self-doubt.
    Appendix A states this heuristic; it is unvalidated and may miss self-doubt appearing in the first half of the response.
  • ad hoc to paper Social comparison theory applies to LLM behavior and explains self-doubt as excessive deference to the user.
    Section 3 motivates the method by analogy to Festinger (1957); the paper provides no evidence that models defer to users in the way the theory describes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Revisiting Overthinking in Long Chain-of-Thought from the Perspective of Self-Doubt." pith.science (2026). https://pith.science/paper/QBAYXLYY

@misc{pith2026250523480,
  author       = {Pith},
  title        = {Pith review of: Revisiting Overthinking in Long Chain-of-Thought from the Perspective of Self-Doubt},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QBAYXLYY}},
  note         = {Machine review of arXiv:2505.23480}
}
read the original abstract

Reasoning Large Language Models (RLLMs) have demonstrated impressive performance on complex tasks, largely due to the adoption of Long Chain-of-Thought (Long CoT) reasoning. However, they often exhibit overthinking -- performing unnecessary reasoning steps even after arriving at the correct answer. Prior work has largely focused on qualitative analyses of overthinking through sample-based observations of long CoTs. In contrast, we present a quantitative analysis of overthinking from the perspective of self-doubt, characterized by excessive token usage devoted to re-verifying already-correct answer. We find that self-doubt significantly contributes to overthinking. In response, we introduce a simple and effective prompting method to reduce the model's over-reliance on input questions, thereby avoiding self-doubt. Specifically, we first prompt the model to question the validity of the input question, and then respond concisely based on the outcome of that evaluation. Experiments on three mathematical reasoning tasks and four datasets with missing premises demonstrate that our method substantially reduces answer length and yields significant improvements across nearly all datasets upon 4 widely-used RLLMs. Further analysis demonstrates that our method effectively minimizes the number of reasoning steps and reduces self-doubt.

Figures

Figures reproduced from arXiv: 2505.23480 by the authors.

Figure 1
Figure 1. The proportion of overthinking and self-doubt across three different datasets with Deepseek￾Deepseek-R1-Distill-Qwen-32B. SD: Self-Doubt, OT w/o SD: Overthinking without Self-Doubt, NOT: Non￾Overtinking. put question, thereby engaging its critique abilities and boosting its confidence; We then instruct the model to provide a concise response based on the outcome of this evaluation. Experiments on three mathematical … view at source ↗
Figure 2
Figure 2. The average thinking steps across three mathematical reasoning tasks with DeepSeek-R1-Distill￾Qwen-32B. 6 Analysis Our method can significantly reduce the number of reasoning steps in Long Chain-of-Thought. Self-doubt often causes the model to generate more steps in its thought process. To validate whether our method can reduce self-doubt, we first com￾pare our method with the baseline from thinking patterns. Specif… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Diagnosing and Mitigating Thinking Collapse in On-Policy Self-Distillation

    cs.CL 2026-07 conditional novelty 6.5 of 10

    Thinking Collapse in reasoning OPSD is driven by teacher gradients at high-entropy forks; AD-OPSD’s dual-perspective soft gate recovers thinking density and up to +4.1% average accuracy.

  2. A Temporal Reasoning Benchmarking Framework for LRMs via Difficulty-controlled and Dynamic Test Generation

    cs.SE 2026-07 conditional novelty 6.5 of 10

    TRACEBench shows LRM temporal accuracy falls almost linearly with a controllable Allen-algebra difficulty score, while ~28% of mid-sized model successes are spurious guesses.

  3. BLADE: Boundary-Expanded and Layer-Adaptive Dynamic Exit for Efficient LLM Reasoning

    cs.CL 2026-07 conditional novelty 5.0 of 10

    BLADE stops LLM reasoning early at sentence and self-doubt boundaries using a layer-adaptive probe, cutting tokens ~25% with near-matched accuracy on five math benchmarks.

  4. Better Starts, Better Ends: Bootstrapped Iterative Self-Reasoning Distillation for Compressed Reasoning

    cs.CL 2026-07 conditional novelty 5.0 of 10

    Warm-starting on-policy self-distillation with a correctness-filtered, prompt-switched SFT bootstrap improves compression and accuracy of reasoning traces across math benchmarks.

Reference graph

Works this paper leans on

22 extracted references · 8 canonical work pages · cited by 4 Pith papers

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Qiguang Chen, Libo Qin, Jinhao Liu, Dengyun Peng, Jiannan Guan, Peng Wang, Mengkang Hu, Yuhang Zhou, Te Gao, and Wanxiang Che. 2025. https://arxiv.org/abs/2503.09567 Towards reasoning era: A survey of long chain-of-thought for reasoning large language models . arXiv preprint arXiv:2503.09567

    arXiv 2025

  4. [4]

    Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, and 1 others. 2024. https://arxiv.org/abs/2412.21187 Do not think that much for 2+ 3=? on the overthinking of o1-like llms . arXiv preprint arXiv:2412.21187

    arXiv 2024

  5. [5]

    Cheng-Han Chiang and Hung-Yi Lee. 2024. https://aclanthology.org/2024.eacl-short.15.pdf Over-reasoning and redundant calculation of large language models . In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers), pages 161--169

    work page 2024

  6. [6]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, and 1 others. 2021. https://arxiv.org/abs/2110.14168 Training verifiers to solve math word problems . arXiv preprint arXiv:2110.14168

    arXiv 2021

  7. [7]

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, Xu Sun, Lei Li, and Zhifang Sui. 2024. https://aclanthology.org/2024.emnlp-main.64/ A survey on in-context learning . In EMNLP

    work page 2024

  8. [8]

    Chenrui Fan, Ming Li, Lichao Sun, and Tianyi Zhou. 2025. https://arxiv.org/abs/2504.06514 Missing premise exacerbates overthinking: Are reasoning models losing critical thinking skill? arXiv preprint arXiv:2504.06514

    arXiv 2025

Show all 22 references
  1. [9]

    Leon Festinger. 1957. https://journals.sagepub.com/doi/10.1177/001872675400700202 Social comparison theory . Selective Exposure Theory

    1957 doi

  2. [10]

    Yichao Fu, Junda Chen, Yonghao Zhuang, Zheyu Fu, Ion Stoica, and Hao Zhang. 2025. https://openreview.net/forum?id=wpK4IMJfdX&referrer= In ICLR 2025 Workshop on Foundation Models in the Wild

    2025

  3. [11]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. https://arxiv.org/abs/2501.12948 Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning . arXiv preprint arXiv...

    2025 arXiv

  4. [12]

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others. 2024. https://arxiv.org/abs/2412.16720 Openai o1 system card . arXiv preprint arXiv:2412.16720

    2024 arXiv

  5. [13]

    https://openreview.net/forum?id=v8L0pN6EOi Let's verify step by step

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. https://openreview.net/forum?id=v8L0pN6EOi Let's verify step by step . In The Twelfth International Conference on Learning Representations

  6. [14]

    Keqin Peng, Liang Ding, Yancheng Yuan, Xuebo Liu, Min Zhang, Yuanxin Ouyang, and Dacheng Tao. 2024. https://aclanthology.org/2024.acl-long.492/ Revisiting demonstration selection strategies in in-context learning . In ACL

    2024

  7. [15]

    Keqin Peng, Liang Ding, Qihuang Zhong, Li Shen, Xuebo Liu, Min Zhang, Yuanxin Ouyang, and Dacheng Tao. 2023. https://aclanthology.org/2023.findings-emnlp.373/ Towards making the most of C hat GPT for machine translation . In Findings of EMNLP 2023

    2023

  8. [16]

    Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Hanjie Chen, and 1 others. 2025. https://arxiv.org/abs/2503.16419 Stop overthinking: A survey on efficient reasoning for large language models . arXiv preprin...

    2025 arXiv

  9. [17]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/hash/9d5609613524ecf4f15af0f7b31abca4-Abstract-Conference.html Chain-of-thought prompting elicits reasoning...

    2022

  10. [18]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. https://arxiv.org/abs/2505.09388 Qwen3 technical report . arXiv preprint arXiv:2505.09388

    2025 arXiv

  11. [19]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, and 22 others. 2024. https://arxiv.org/abs/...

    2024 arXiv

  12. [20]

    Yuqi Zhang, Liang Ding, Lefei Zhang, and Dacheng Tao. 2025. https://aclanthology.org/2025.coling-main.199/ Intention analysis makes llms a good jailbreak defender . In Proceedings of the 31st International Conference on Computational Linguistics, pages 2947--2968

    2025

  13. [21]

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, and 1 others. 2023. https://arxiv.org/abs/2306.05685 Judging llm-as-a-judge with mt-bench and chatbot arena . NIPS, 36:46595--46623

    2023 arXiv

  14. [22]

    Qihuang Zhong, Kang Wang, Ziyang Xu, Juhua Liu, Liang Ding, and Bo Du. 2024. https://arxiv.org/abs/2404.14963 Achieving> 97\ arXiv preprint arXiv:2404.14963

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.