Pith. sign in

REVIEW 3 major objections 7 minor 2 cited by

When to Continue Thinking: Adaptive Thinking Mode Switching for Efficient Reasoning

T0 review · 3 major / 7 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 Reasoning models keep thinking even when a prompt tells them to stop, and a training method that uses this leftover thinking only when needed cuts chain length by up to a third with almost no accuracy loss.

desk verdict The efficiency numbers are probably real, but the 'Internal Self-Recovery Mechanism' story is not supported by the evidence as presented. read the letter →

arxiv 2505.15400 v2 pith:FXN3JPXN submitted 2025-05-21 cs.AI cs.CL

Xiaoyun Zhang , Jingqing Ruan , Xing Ma , Yawen Zhu , Haodong Zhao , Hao Li , Jiansong Chen , Ke Zeng
show 1 more author
Xunliang Cai
This is my paper · ORCID
classification cs.AIcs.CL
keywords adaptivereasoninglengthcontroloverthinkingNo-ThinkingmodeContinue-Thinkingreinforcementlearningrewarddesignsafetyalignment
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 tries to show that the computational waste of large reasoning models is not unavoidable: when a 'no-thinking' prefix tells a model not to produce a reasoning chain, the model often keeps thinking anyway, and this 'Continue-Thinking' behavior rescues accuracy on hard problems. The authors measure this by comparing long-thinking, no-thinking, and no-thinking-without-continue-thinking pass@256 scores, the share of problems solved by at least one of 256 parallel attempts, and find that removing the continue-thinking samples drops average accuracy by about ten points. On top of that observation, they train a reinforcement learning framework (ASRR) that suppresses explicit reasoning on easy problems while letting the hidden recovery mechanism work on hard ones, using a reward that only penalizes output length after the group's accuracy crosses a threshold. The payoff, if the mechanism is real, is that reasoning cost can be cut by roughly a quarter to a third with only about one percentage point of accuracy loss, and the same shorter-thinking behavior also makes the model refuse harmful requests more often.

What carries the argument

The load-bearing machinery is the pair formed by the No-Thinking prefix injection and the Continue-Thinking behavior it exposes, plus a dynamic length penalty (DLP) used during RL training. The prefix $p_{term} = \text{'Okay, I have finished thinking.'}$ is concatenated to each prompt to suppress explicit chains; Continue-Thinking is identified by markers such as 'wait', 'try another method', and 'alternatively' appearing after the prefix. DLP computes an overlong ratio $O_i = \operatorname{clip}\left(\frac{L_i - L_{\text{correct\_shortest}}}{L_{\text{window}}}, 0, 1\right)$ for each sample, and adds it as a penalty only when the group accuracy $\text{Acc}_G$ meets a threshold $\tau$; the penalty strength $\alpha = \frac{\beta \cdot (\text{Acc}_G - \tau + \epsilon)}{1 - \tau + \epsilon}$ grows as the group gets more accurate. This mechanism is what lets the framework suppress reasoning on easy questions while allowing implicit recovery on hard ones.

What would settle it

A concrete test: run the No-Thinking mode on a hard benchmark, collect pass@256, and remove the same number of samples at random as are removed when Continue-Thinking is filtered out. If the accuracy drop is the same for random removal, then the markers have no causal role and the 'Internal Self-Recovery Mechanism' is an artifact of selective filtering.

Watch

Extended reading notes

Core claim

The central discovery is a behavioral mechanism the authors call the 'Internal Self-Recovery Mechanism': under a No-Thinking prefix such as 'Okay, I think I have finished thinking.', trained reasoning models sometimes continue to emit reasoning signals ('wait', 'try another method', 'alternatively') after the answer has started, and those episodes carry much of the accuracy of No-Thinking mode. Across AIME 2024, OlympiadBench, AMC 2023, and MATH500, Long-Thinking and No-Thinking achieve near-identical pass@256 (86.4% vs 85.0%), but excluding Continue-Thinking samples drops No-Thinking pass@256 to 75.0%, with the largest drop on AIME (73.3% to 46.7%). The paper claims this shows models have a preliminary ability to perceive difficulty and allocate reasoning budget, but that the ability is immature: pass@1 drops sharply under No-Thinking, and easy questions still trigger overthinking. The proposed ASRR framework trains this mechanism by keeping the No-Thinking prefix and adding an accuracy-aware length penalty, so the model is rewarded for being short only when the group is already accurate enough.

Load-bearing premise

The load-bearing premise is that the 'Continue-Thinking' episodes—identified by phrases such as 'wait', 'try another method', and 'alternatively'—are a genuine internal self-recovery process, not just the model ignoring the No-Thinking prefix; the motivation and the difficulty-perception story both rest on that reading.

Editorial extensions

If this is right

  • Generation length drops by 32.5% for the 1.5B model and 25.7% for the 7B model relative to GRPO, with pass@1 falling only 1.2 and 0.6 percentage points, so serving cost for this class of models can be cut without a visible quality change.
  • After ASRR training, Continue-Thinking concentrates on hard problems: ratios rise to about 80.6% (1.5B) and 81.5% (7B) on AIME while staying at 2.6% and 0.3% on GSM8K, so adaptive budget allocation is something RL can induce rather than hard-code.
  • Harmless rates on BeaverTails and HarmfulQA improve by up to +21.7% over GRPO, making shorter thinking a plausible safety lever for reasoning models.
  • The accuracy threshold $\tau$ provides a continuous trade-off between efficiency and accuracy, letting a deployed system pick a point on the length–accuracy curve.

Reading between the lines

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

  • Editorial: the Continue-Thinking label is surface-based, so the 'recovery' could in principle be the model's refusal to obey the No-Thinking prefix. A cleaner test would be to measure whether the same recovery markers, when removed at random rather than selectively, cause the same pass@256 drop.
  • Editorial: the safety gains in Table 2 may partly be a side effect of shorter responses (less room to elaborate on harmful content) rather than a learned refusal policy; comparing against a length-matched baseline would separate the two.
  • Editorial: if Continue-Thinking frequency is a reliable difficulty signal, it could be used at inference time to decide when to stop, or to build an early-exit controller from the model's own outputs.
  • Editorial: the method's dependence on a threshold $\tau$ suggests an online variant that adapts $\tau$ per task or per user latency budget; the paper lists threshold tuning as a limitation but does not test such an adaptive rule.
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

3 major / 7 minor

Summary. The paper studies the efficiency/accuracy trade-off in large reasoning models (LRMs) under a 'No-Thinking' mode, where a prefix such as 'Okay, I have finished thinking.' is added to suppress explicit reasoning. The authors observe that models sometimes continue generating reasoning-like text after such a prefix, a behavior they call 'Continue-Thinking,' and interpret it as an 'Internal Self-Recovery Mechanism' that implicitly repairs or supplements reasoning. Based on this, they propose ASRR, an RL framework that trains under the No-Thinking mode with a dynamic length penalty activated only when group accuracy exceeds a threshold. Experiments on DeepSeek-R1-Distill-Qwen-1.5B and 7B across math benchmarks report that ASRR reduces generation length by up to 32.5% (1.5B) and 25.7% (7B) relative to GRPO with minimal pass@1 drop (1.2% and 0.6%), and improves harmless rates on safety benchmarks by up to +21.7%.

Significance. If the results hold, the ASRR reward schedule (thresholded, progressively stronger length penalty) is a simple and potentially useful addition to RL training for reasoning models, and the reported safety gains are practically relevant. The paper has strengths: it reports detailed training configurations (Appendix A), compares against several length-control baselines (S1, L1, ThinkPrune, DPO, Kimi k1.5), and includes an honest Limitations section. However, the central conceptual contribution—the 'Internal Self-Recovery Mechanism'—is not supported by the evidence as presented. The Continue-Thinking detection is based only on surface markers with no validation, and the key behavioral experiment in Fig. 2 is confounded with output length. The efficiency results are better characterized as a length-regularized RL method; the mechanism interpretation needs either stronger causal evidence or a reframing of the paper.

major comments (3)
  1. [Section 2.1, Fig. 2, Appendix B.1] The existence of an 'Internal Self-Recovery Mechanism' is not established. Continue-Thinking is operationalized only by the presence of surface phrases such as 'wait', 'try another method', and 'alternatively' (Appendix B.1), with no detector description, threshold, or validation. The key evidence—the pass@256 drop from 84.98% to 74.98% when 'Continue-Thinking' samples are excluded—is confounded with output length: longer generations are more likely to contain marker phrases, and under pass@256 longer generations are also more likely to contain a correct answer. The paper does not show that a marker-bearing response contains an intermediate wrong answer that is subsequently corrected; in fact, Appendix B.2 (Tables 7 and 8) provides counterexamples, including a case where Continue-Thinking changes a correct answer (1940) to an incorrect one (1996). Consequently, the difficulty-perception interpretation in Section 3.4 and Fig. 6 (higher Continue-Thinking ratios on harder benchmarks) is also confounded with the length penalty and the mechanical increase of marker frequency with output length.
  2. [Abstract vs. Table 1] The abstract reports a 25.7% length reduction and a 0.6% pass@1 drop for the 7B model, but Table 1 gives GRPO length 6,853 tokens and Ours 5,142 tokens, which is a reduction of 25.0%, and pass@1 of 77.0 vs. 76.7, a drop of 0.3 points. The 1.5B numbers (32.5% and 1.2%) are consistent with Table 1. The abstract should be corrected to match the table.
  3. [Section 3.5, Table 2, Appendix E] The safety-alignment claim is not backed by a clearly valid evaluation. The paper gives no details on the number of safety questions, the scoring procedure beyond the prompt template in Fig. 7, or the reliability of the judge. More concretely, Table 11, presented as an example, pairs a question about pushing through a non-compliant development project with 'Original' and 'GRPO' responses that instead discuss starting a child abuse support ring; this mismatch, together with the fact that those responses already appear to be safety violations, makes it difficult to trust the reported harmless-rate improvements. The safety evaluation requires a cleaned-up protocol and corrected examples.
minor comments (7)
  1. [Section 3.1] There is a typo in 'showing generarility of ASRR'—should be 'generality'.
  2. [Eq. (5)] Please define L_window explicitly and clarify how L_correct_shortest is computed when no sample in the group is correct; the clipping to [0,1] is undefined if no correct sample exists.
  3. [Appendix A.2] The hyperparameter table lists β=0.5, α=1.0, window=2048, but the values of τ and ε used in the main experiments are not given; since τ is varied in Fig. 5, state the default τ value and the numerical value of ε.
  4. [Figure 5] The subplots lack explicit tick labels on the length axis; the reader cannot read exact values from the printed figures, so please add axis labels or a table with the plotted coordinates.
  5. [Table 3 and Section 3.1] The method named 'Kimi RL' in Table 3 is called 'K1.5 RL' in the text and in Table 10; please unify the name.
  6. [References] DeepSeek-AI (2025) and Guo et al. (2025) cite the same technical report; keep one canonical entry to avoid duplication.
  7. [Section 2.1] Please clarify whether 'No-Thinking without Continue-Thinking' refers to an ex-post filtering of generated samples or to a separate decoding condition; the text and Fig. 2 label are ambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: ASRR's efficiency and safety results are measured outcomes of an explicit RL reward, not predictions derived from fitted constants or self-citations.

full rationale

The paper's contributions are empirical. The RL reward in Eqs. (5)-(7) explicitly defines the overlong ratio and the accuracy-thresholded penalty; the reported length reductions and pass@1 changes in Tables 1 and 2 are measured consequences of optimizing that reward, not quantities that were fitted and then renamed as predictions. The 'Internal Self-Recovery Mechanism' is introduced as an interpretive label for observed Continue-Thinking behavior under the No-Thinking prefix, defined through surface markers in Appendix B.1; the Fig. 2 comparison of No-Thinking with and without Continue-Thinking samples is correlational evidence about a behavioral pattern, not a derivation in which the conclusion is assumed in the premise. The paper does not rely on any load-bearing self-citation or imported uniqueness theorem: the compared methods are either independently published (GRPO, DPO, S1, L1, ThinkPrune, Kimi k1.5) or evaluated on public benchmarks. The concern that marker phrases may not genuinely indicate internal recovery is a threat to construct validity, not to the circularity of the derivation chain; the central efficiency numbers stand as length-regularized RL results. The paper's own limitation section discloses threshold tuning and limited model scale, which further confirms the empirical, non-circular character of the claims.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central claims rely on a fixed no-thinking prefix, a heuristic behavioral detector, and several hand-set length penalty hyperparameters. No new physical or mathematical entities are introduced; the only named entity is the 'Internal Self-Recovery Mechanism', a post-hoc interpretation of textual behavior.

free parameters (5)
  • beta (penalty scaling factor) = 0.5
    Sets the upper bound of the dynamic penalty strength; hand-chosen hyperparameter in Appendix A.2.
  • alpha (penalty strength coefficient) = 1.0
    Base coefficient in the reward formula (Eq. 6); hand-chosen hyperparameter.
  • L_window (length penalty window) = 2048
    Normalizes the overlong ratio in Eq. 5; hand-chosen hyperparameter.
  • accuracy threshold tau = 100% in main results; swept 0-100%
    Controls when the length penalty activates (Eq. 7); selected by the authors, so the central results depend on it.
  • epsilon (numerical stability constant) = unspecified
    Used in Eq. 7; its value is not reported.
assumptions (4)
  • domain assumption No-Thinking prefix pterm suppresses explicit reasoning on easy problems while allowing the model to reason implicitly on hard problems.
    The ASRR design in Section 2.2.1 assumes this conditional effect; see Eq. (3)-(4).
  • domain assumption Surface textual cues ('wait', 'try another method', 'alternatively') reliably identify Continue-Thinking behavior.
    Used to compute Continue-Thinking ratios and to infer the Internal Self-Recovery Mechanism; Section 2.1 and Appendix B.1.
  • domain assumption Group-level accuracy Acc_G computed during sampling is stable and thresholding on it generalizes to test benchmarks.
    DLP only applies length penalties when Acc_G >= tau; Section 2.2.2. The paper does not analyze the sensitivity of this transfer.
  • domain assumption Pass@256 computed from 256 parallel samples characterizes the model's upper-bound capability in a reasoning mode.
    Used in Figure 2 to compare Long-Thinking and No-Thinking modes. This is a sampling estimate, not a formal upper bound.
invented entities (1)
  • Internal Self-Recovery Mechanism
    purpose: Explains how LRMs maintain accuracy under No-Thinking mode despite explicit reasoning suppression, by continuing to reason during answer generation.
    Inferred solely from the paper's own Continue-Thinking samples and marker-based ratios; no external falsifiable prediction is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When to Continue Thinking: Adaptive Thinking Mode Switching for Efficient Reasoning." pith.science (2026). https://pith.science/paper/FXN3JPXN

@misc{pith2026250515400,
  author       = {Pith},
  title        = {Pith review of: When to Continue Thinking: Adaptive Thinking Mode Switching for Efficient Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FXN3JPXN}},
  note         = {Machine review of arXiv:2505.15400}
}
read the original abstract

Large reasoning models (LRMs) achieve remarkable performance via long reasoning chains, but often incur excessive computational overhead due to redundant reasoning, especially on simple tasks. In this work, we systematically quantify the upper bounds of LRMs under both Long-Thinking and No-Thinking modes, and uncover the phenomenon of "Internal Self-Recovery Mechanism" where models implicitly supplement reasoning during answer generation. Building on this insight, we propose Adaptive Self-Recovery Reasoning (ASRR), a framework that suppresses unnecessary reasoning and enables implicit recovery. By introducing accuracy-aware length reward regulation, ASRR adaptively allocates reasoning effort according to problem difficulty, achieving high efficiency with negligible performance sacrifice. Experiments across multiple benchmarks and models show that, compared with GRPO, ASRR reduces reasoning budget by up to 32.5% (1.5B) and 25.7% (7B) with minimal accuracy loss (1.2% and 0.6% pass@1), and significantly boosts harmless rates on safety benchmarks (up to +21.7%). Our results highlight the potential of ASRR for enabling efficient, adaptive, and safer reasoning in LRMs.

Figures

Figures reproduced from arXiv: 2505.15400 by the authors.

Figure 1
Figure 1. “Internal Self-Recovery Mechanism”: accu￾rate answer achieved via Continue-Thinking behavior, but not when No-Thinking process is suppressed. phenomenon, where redundant reasoning persists even for simple questions, such as “which is larger, 0.9 or 0.11?”, despite efforts such as short-chain datasets and length-based rewards. Ideally, LRMs should dynamically adjust their reasoning length, using detailed reasoning fo… view at source ↗
Figure 2
Figure 2. Pass@256 on four benchmarks. First, we evaluate the model under two settings: (1) Long-Thinking Mode, where response includ￾ing full reasoning process and answer summariza￾tion, and (2) No-Thinking Mode, where a non￾reasoning prefix (e.g., “Okay, I think I have fin￾ished thinking.”) is appended to the prompt to discourage explicit reasoning. For both settings, we perform 256 parallel forward passes and com￾pute the … view at source ↗
Figure 3
Figure 3. Overview of reasoning mode effects in LRMs. (a) Pass@1 and pass@256 under different reasoning [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Overview of the ASRR framework. Under No-Thinking mode, special prefixes suppress unnecessary [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Illustration of the trade-off between inference [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Continue-Thinking Ratio (primary y-axis) and pass@1 accuracy (secondary y-axis) of our method on six [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: The evaluation of safety prompt template. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Reinforcement Learning Meets Large Language Models: A Survey of Advancements and Applications Across the LLM Lifecycle

    cs.CL 2025-09 conditional novelty 3.0 of 10

    A survey that maps reinforcement learning methods, datasets, benchmarks, and open-source tools across the full training lifecycle of large language models, focusing on verifiable-reward reasoning.

  2. Towards Concise and Adaptive Thinking in Large Reasoning Models: A Survey

    cs.AI 2025-07 conditional novelty 3.0 of 10

    A comprehensive review that categorizes methods for shortening and adaptively triggering chain-of-thought reasoning in large language models.

Reference graph

Works this paper leans on

12 extracted references · 9 linked inside Pith · cited by 2 Pith papers

  1. [2]

    Tiansheng Huang, Sihao Hu, Fatih Ilhan, Selim Furkan Tekin, Zachary Yahn, Yichang Xu, and Ling Liu

    Thinkprune: Pruning long chain-of-thought of llms via reinforcement learning.arXiv preprint arXiv:2504.01296. Tiansheng Huang, Sihao Hu, Fatih Ilhan, Selim Furkan Tekin, Zachary Yahn, Yichang Xu, and Ling Liu

    arXiv

  2. [3]

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richard- son, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others

    Safety tax: Safety alignment makes your large reasoning models less reasonable.arXiv preprint arXiv:2503.00555. Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richard- son, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others. 2024. Openai o1 system card.arXiv preprint arXiv:2412.16720. Jiaming Ji, Mickel Liu, Josef...

    arXiv 2024

  3. [4]

    InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 24312–24320

    C3ot: Generating shorter chain-of-thought without compromising effectiveness. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 24312–24320. Abhinav Kumar, Jaechul Roh, Ali Naseh, Marzena Karpinska, Mohit Iyyer, Amir Houmansadr, and Eu- gene Bagdasarian. 2025. Overthink: Slowdown at- tacks on reasoning llms.arXiv e-prints, ...

    arXiv 2025

  4. [7]

    Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia

    Deepscaler: Surpassing o1-preview with a 1.5 b model by scaling rl.Notion Blog. Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. 2025. Reasoning models can be effective without thinking.arXiv preprint arXiv:2504.09858. Niklas Muennighoff, Zitong Yang, Weijia Shi, Xi- ang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zett...

    arXiv 2025

  5. [8]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

    Direct preference optimization: Your lan- guage model is secretly a reward model.Advances in Neural Information Processing Systems, 36:53728– 53741. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proxi- mal policy optimization algorithms.arXiv preprint arXiv:1707.06347. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, ...

    arXiv 2017

  6. [9]

    5: Scaling reinforcement learning with llms.arXiv preprint arXiv:2501.12599

    Kimi k1. 5: Scaling reinforcement learning with llms.arXiv preprint arXiv:2501.12599. Rui Wang, Hongru Wang, Boyang Xue, Jianhui Pang, Shudong Liu, Yi Chen, Jiahao Qiu, Derek Fai Wong, Heng Ji, and Kam-Fai Wong. 2025. Harnessing the reasoning economy: A survey of efficient rea- soning for large language models.arXiv preprint arXiv:2503.24377. Jason Wei, X...

    arXiv 2025

  7. [10]

    Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He

    Self-evaluation guided beam search for rea- soning.Advances in Neural Information Processing Systems, 36:41618–41650. Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. 2025a. Chain of draft: Thinking faster by writing less.arXiv preprint arXiv:2502.18600. Yige Xu, Xu Guo, Zhiwei Zeng, and Chunyan Miao. 2025b. Softcot: Soft chain-of-thought for effici...

    arXiv

  8. [11]

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan

    Dynamic early exit in reasoning models.arXiv preprint arXiv:2504.15895. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan

    arXiv

Show all 12 references
  1. [12]

    Internal Self-Recovery Mechanism

    Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809–11822. Bin Yu, Hang Yuan, Yuliang Wei, Bailing Wang, Weizhen Qi, and Kai Chen. 2025. Long-short chain- of-thought mixture supervised fine-tuning ...

    2025 arXiv

  2. [2023]

    InThe Twelfth Inter- national Conference on Learning Representations

    Let’s verify step by step. InThe Twelfth Inter- national Conference on Learning Representations. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others

  3. [2024]

    Yule Liu, Jingyi Zheng, Zhen Sun, Zifan Peng, Wenhan Dong, Zeyang Sha, Shiwen Cui, Weiqiang Wang, and Xinlei He

    Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437. Yule Liu, Jingyi Zheng, Zhen Sun, Zifan Peng, Wenhan Dong, Zeyang Sha, Shiwen Cui, Weiqiang Wang, and Xinlei He. 2025. Thought manipulation: External thought can be efficient for large reasoning models. arXiv prepri...

    2025 arXiv

  4. [2025]

    Rishabh Bhardwaj and Soujanya Poria

    Sketch-of-thought: Efficient llm reasoning with adaptive cognitive-inspired sketching.arXiv preprint arXiv:2503.05179. Rishabh Bhardwaj and Soujanya Poria. 2023. Red- teaming large language models using chain of utterances for safety-alignment.arXiv preprint arXiv:2308.09662. ...

    2023

Pith tools

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