pith. sign in

arxiv: 2607.00862 · v1 · pith:4UJBU2UGnew · submitted 2026-07-01 · 💻 cs.CL · cs.AI

CAT: Confidence-Adaptive Thinking for Efficient Reasoning of Large Reasoning Models

Qizhi Jiang , Shuo Wang , Pei Ke , Yuhang Song , Ke Qin This is my paper

Pith reviewed 2026-07-02 13:20 UTC · model grok-4.3

classification 💻 cs.CL cs.AI
keywords confidence-adaptive thinkinglarge reasoning modelschain-of-thought compressionpreference optimizationself-certainty signalsreasoning efficiency
0
0 comments X

The pith

CAT trains reasoning models to shorten their chain-of-thought on easy problems by feeding their own certainty estimates into preference optimization.

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

The paper presents Confidence-Adaptive Thinking as a way for large reasoning models to decide how long to reason based on their internal assessment of certainty about each query. Prior compression techniques apply the same length reduction everywhere or rely on external difficulty labels, which often lowers accuracy on the hardest cases. CAT instead inserts the model's self-generated certainty scores directly into the preference optimization step so the model learns to produce shorter trajectories when confident and longer ones when uncertain. A sympathetic reader would care because this targets the token waste from overthinking simple questions without the accuracy penalty that uniform shortening produces.

Core claim

CAT incorporates the model's intrinsic self-certainty signals as confidence into the preference optimization process, which autonomously modulates reasoning lengths based on problem difficulty. Experimental results show that CAT consistently outperforms state-of-the-art baselines on reasoning accuracy across multiple benchmarks on different base models. The framework enables LRMs to effectively compress confident responses while deliberating on uncertain ones.

What carries the argument

Self-certainty signals inserted as confidence scores inside preference optimization to produce difficulty-adaptive reasoning lengths.

If this is right

  • Reasoning length decreases automatically on problems where the model reports high certainty.
  • Accuracy on difficult problems stays at or above the level of uncompressed models.
  • The same training procedure transfers across multiple base reasoning models without task-specific tuning.

Where Pith is reading between the lines

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

  • Production systems could route simple queries to shorter paths and reserve full computation for uncertain ones without external difficulty classifiers.
  • The same certainty signal might be reused at inference time to decide early stopping rather than only during training.
  • Combining CAT with token-level pruning could produce further efficiency gains on the already-shortened trajectories.

Load-bearing premise

The model's own certainty estimates reliably mark which problems are easy enough for shorter reasoning without hurting final accuracy on harder problems.

What would settle it

A controlled test on the hardest quartile of each benchmark where CAT-trained models show lower accuracy than the uncompressed baseline would falsify the central claim.

Figures

Figures reproduced from arXiv: 2607.00862 by Ke Qin, Pei Ke, Qizhi Jiang, Shuo Wang, Yuhang Song.

Figure 1
Figure 1. Figure 1: Overview of the CAT framework. path-level self-certainty scores as confidence via a dedicated forward pass (Section 3.2.1). Secondly, we construct preference pairs based on the confi￾dence and lengths, and apply dynamic selection to prioritize more informative supervision (Section 3.2.1). Finally, we fine-tune the base LRM with a confidence-weighted preference optimization ob￾jective, which incorporates co… view at source ↗
Figure 2
Figure 2. Figure 2: Analysis of Self-Certainty (SC) distributions regarding response correctness and robustness to length on [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Case study on DeepSeek-R1-Distill-Qwen-7B. All the three methods solve the problem correctly. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Difficulty and length distributions illustrating the diversity of the question set. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Hyperparameter analysis. Level 1 Level 2 Level 3 Level 4 Level 5 75 80 85 90 95 100 Accuracy (%) Origin ConCISE DAST CAT (a) Accuracy Performance on Difficulty Levels. Level 1 Level 2 Level 3 Level 4 Level 5 0 1000 2000 3000 4000 5000 6000 Average Token Length Origin ConCISE DAST CAT (b) Average Token Length on Difficulty Levels [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance comparison on MATH-500 across different difficulty levels. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Average Self-Certainty (SC) across difficulty [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
read the original abstract

Large Reasoning Models (LRMs) have achieved remarkable success on complex tasks by leveraging long chain-of-thought (CoT) trajectories, yet they frequently exhibit overthinking on simple queries, resulting in significant token overhead and reduced inference efficiency. However, existing compression methods predominantly apply uniform length reduction or rely on coarse-grained difficulty estimation, often leading to performance degradation on difficult problems. To address this limitation, we propose Confidence-Adaptive Thinking (CAT), a framework that incorporates the model's intrinsic self-certainty signals as confidence into the preference optimization process, which autonomously modulates reasoning lengths based on problem difficulty. Experimental results show that CAT consistently outperforms state-of-the-art baselines on reasoning accuracy across multiple benchmarks on different base models. Our work enables LRMs to effectively compress confident responses while deliberating on uncertain ones, offering a potentially robust solution for balancing accuracy and latency in practical industrial scenarios.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes Confidence-Adaptive Thinking (CAT), a framework that incorporates a large reasoning model's intrinsic self-certainty signals into preference optimization to autonomously modulate chain-of-thought length according to estimated problem difficulty. It claims that this yields consistent accuracy gains over state-of-the-art compression baselines across multiple reasoning benchmarks and base models while reducing token overhead on easy problems.

Significance. If the self-certainty signals are shown to be calibrated and the optimization demonstrably preserves accuracy on hard cases, the method could provide a practical route to efficient inference for LRMs by avoiding both uniform over-compression and external difficulty estimators. The use of intrinsic signals is a conceptual strength relative to prior coarse-grained approaches.

major comments (3)
  1. [Abstract and §4] Abstract and §4 (Experiments): the headline claim of 'consistent outperformance' is asserted without any reported accuracy deltas, token counts, baseline definitions, or statistical tests; the experiments section must supply these quantitative results (including per-benchmark tables and error analysis) for the central claim to be evaluable.
  2. [§3 and §4.3] §3 (Method) and §4.3 (Ablations): the framework's validity hinges on self-certainty reliably indicating difficulty and on preference optimization shortening high-certainty trajectories without degrading low-certainty performance. No calibration plots, correlation with ground-truth difficulty, or ablation that removes the confidence term from the objective are described; these checks are load-bearing for the 'no degradation on difficult problems' assertion.
  3. [§4.2] §4.2 (Results): if a per-difficulty or per-certainty stratified breakdown exists, it must demonstrate that accuracy on hard/uncertain items is preserved; its absence leaves open the possibility that aggregate gains mask losses on difficult cases, directly undermining the modulation claim.
minor comments (2)
  1. [§3] Clarify the exact definition and computation of the self-certainty score (e.g., whether it is token-averaged log-probability, entropy, or another quantity) with an equation in the method section.
  2. [§5] Add a short discussion of failure modes, such as cases where self-certainty is miscalibrated on out-of-distribution problems.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback, which highlights important areas for strengthening the quantitative presentation and empirical validation of CAT. We address each major comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (Experiments): the headline claim of 'consistent outperformance' is asserted without any reported accuracy deltas, token counts, baseline definitions, or statistical tests; the experiments section must supply these quantitative results (including per-benchmark tables and error analysis) for the central claim to be evaluable.

    Authors: We agree that explicit quantitative details are required to substantiate the central claims. The experiments section contains comparative results, but we will revise §4 to include per-benchmark tables reporting accuracy deltas, token counts, explicit baseline definitions, and statistical significance tests, along with error analysis. The abstract will be updated to reference key quantitative improvements. revision: yes

  2. Referee: [§3 and §4.3] §3 (Method) and §4.3 (Ablations): the framework's validity hinges on self-certainty reliably indicating difficulty and on preference optimization shortening high-certainty trajectories without degrading low-certainty performance. No calibration plots, correlation with ground-truth difficulty, or ablation that removes the confidence term from the objective are described; these checks are load-bearing for the 'no degradation on difficult problems' assertion.

    Authors: We acknowledge that these supporting analyses are necessary. In the revision we will add calibration plots for the self-certainty signals, their correlation with ground-truth difficulty, and an ablation that removes the confidence term from the preference optimization objective. These additions will directly validate that high-certainty trajectories can be shortened without degrading performance on low-certainty cases. revision: yes

  3. Referee: [§4.2] §4.2 (Results): if a per-difficulty or per-certainty stratified breakdown exists, it must demonstrate that accuracy on hard/uncertain items is preserved; its absence leaves open the possibility that aggregate gains mask losses on difficult cases, directly undermining the modulation claim.

    Authors: We agree that aggregate results alone are insufficient. We will incorporate a per-difficulty and per-certainty stratified breakdown in the revised §4.2 to explicitly demonstrate that accuracy on hard or uncertain items is preserved while compression occurs on easier cases. revision: yes

Circularity Check

0 steps flagged

No circularity: framework relies on external preference optimization and empirical validation without self-referential reductions

full rationale

The paper proposes incorporating intrinsic self-certainty signals into preference optimization to modulate CoT length, with claims of accuracy gains supported by experimental results on benchmarks. No equations, fitted parameters, or derivations are presented in the provided text that reduce a claimed prediction or result back to the inputs by construction. The central mechanism is described as an autonomous modulation process rather than a mathematical identity or self-citation chain that forces the outcome. This is a standard empirical proposal whose validity rests on external benchmarks rather than definitional equivalence, qualifying for score 0.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no information on free parameters, background axioms, or newly postulated entities.

pith-pipeline@v0.9.1-grok · 5683 in / 894 out tokens · 31297 ms · 2026-07-02T13:20:29.323567+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 28 canonical work pages · 9 internal anchors

  1. [1]

    OpenAI o1 System Card

    OpenAI , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2412.16720 , eprinttype =. 2412.16720 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.16720 2024

  2. [2]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    DeepSeek. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning , journal =. 2025 , url =. doi:10.48550/ARXIV.2501.12948 , eprinttype =. 2501.12948 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2501.12948 2025

  3. [3]

    Chi and Quoc V

    Jason Wei and Xuezhi Wang and Dale Schuurmans and Maarten Bosma and Brian Ichter and Fei Xia and Ed H. Chi and Quoc V. Le and Denny Zhou , editor =. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , booktitle =. 2022 , url =

    2022

  4. [4]

    Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs

    Xingyu Chen and Jiahao Xu and Tian Liang and Zhiwei He and Jianhui Pang and Dian Yu and Linfeng Song and Qiuzhi Liu and Mengfei Zhou and Zhuosheng Zhang and Rui Wang and Zhaopeng Tu and Haitao Mi and Dong Yu , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2412.21187 , eprinttype =. 2412.21187 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.21187 2024

  5. [5]

    CoRR , volume =

    Yue Liu and Jiaying Wu and Yufei He and Hongcheng Gao and Hongyu Chen and Baolong Bi and Jiaheng Zhang and Zhiqi Huang and Bryan Hooi , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2503.23077 , eprinttype =. 2503.23077 , timestamp =

    work page doi:10.48550/arxiv.2503.23077 2025

  6. [6]

    Stop Overthinking:

    Yang Sui and Yu. Stop Overthinking:. Trans. Mach. Learn. Res. , volume =. 2025 , url =

    2025

  7. [7]

    Sicheng Feng and Gongfan Fang and Xinyin Ma and Xinchao Wang , title =. Trans. Mach. Learn. Res. , volume =. 2025 , url =

    2025

  8. [8]

    A Survey of Efficient Reasoning for Large Reasoning Models: Language, Multimodality, and Beyond , journal =

    Xiaoye Qu and Yafu Li and Zhaochen Su and Weigao Sun and Jianhao Yan and Dongrui Liu and Ganqu Cui and Daizong Liu and Shuxian Liang and Junxian He and Peng Li and Wei Wei and Jing Shao and Chaochao Lu and Yue Zhang and Xian. A Survey of Efficient Reasoning for Large Reasoning Models: Language, Multimodality, and Beyond , journal =. 2025 , url =. doi:10.4...

    work page doi:10.48550/arxiv.2503.21614 2025

  9. [9]

    T oken S kip: Controllable Chain-of-Thought Compression in LLM s

    Xia, Heming and Leong, Chak Tou and Wang, Wenjie and Li, Yongqi and Li, Wenjie. T oken S kip: Controllable Chain-of-Thought Compression in LLM s. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.165

    work page doi:10.18653/v1/2025.emnlp-main.165 2025

  10. [10]

    CoT-Valve: Length-Compressible Chain-of-Thought Tuning , booktitle =

    Xinyin Ma and Guangnian Wan and Runpeng Yu and Gongfan Fang and Xinchao Wang , editor =. CoT-Valve: Length-Compressible Chain-of-Thought Tuning , booktitle =. 2025 , url =

    2025

  11. [11]

    Self-Training Elicits Concise Reasoning in Large Language Models , booktitle =

    Tergel Munkhbat and Namgyu Ho and Seo Hyun Kim and Yongjin Yang and Yujin Kim and Se. Self-Training Elicits Concise Reasoning in Large Language Models , booktitle =. 2025 , url =

    2025

  12. [12]

    CoRR , volume =

    Zhiyuan Zeng and Qinyuan Cheng and Zhangyue Yin and Bo Wang and Shimin Li and Yunhua Zhou and Qipeng Guo and Xuanjing Huang and Xipeng Qiu , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2412.14135 , eprinttype =. 2412.14135 , timestamp =

    work page doi:10.48550/arxiv.2412.14135 2024

  13. [13]

    s1: Simple test-time scaling , booktitle =

    Niklas Muennighoff and Zitong Yang and Weijia Shi and Xiang Lisa Li and Li Fei. s1: Simple test-time scaling , booktitle =. 2025 , url =. doi:10.18653/V1/2025.EMNLP-MAIN.1025 , timestamp =

    work page doi:10.18653/v1/2025.emnlp-main.1025 2025

  14. [14]

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

    Yi Shen and Jian Zhang and Jieyun Huang and Shuming Shi and Wenjing Zhang and Jiangze Yan and Ning Wang and Kai Wang and Zhaoxiang Liu and Shiguo Lian , title =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,. 2025 , url =

    2025

  15. [15]

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

    Ziqing Qiao and Yongheng Deng and Jiali Zeng and Dong Wang and Lai Wei and Guanbo Wang and Fandong Meng and Jie Zhou and Ju Ren and Yaoxue Zhang , title =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,. 2025 , url =. doi:10.18653/V1/2025.EMNLP-MAIN.405 , timestamp =

    work page doi:10.18653/v1/2025.emnlp-main.405 2025

  16. [16]

    Zhao and Kelvin Guu and Adams Wei Yu and Brian Lester and Nan Du and Andrew M

    Jason Wei and Maarten Bosma and Vincent Y. Zhao and Kelvin Guu and Adams Wei Yu and Brian Lester and Nan Du and Andrew M. Dai and Quoc V. Le , title =. The Tenth International Conference on Learning Representations,. 2022 , url =

    2022

  17. [17]

    Yu Meng and Mengzhou Xia and Danqi Chen , title =. Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024 , year =

    2024

  18. [18]

    Scalable Best-of-N Selection for Large Language Models via Self-Certainty,

    Zhewei Kang and Xuandong Zhao and Dawn Song , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2502.18581 , eprinttype =. 2502.18581 , timestamp =

    work page doi:10.48550/arxiv.2502.18581 2025

  19. [19]

    Deep Think with Confidence

    Yichao Fu and Xuewei Wang and Yuandong Tian and Jiawei Zhao , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2508.15260 , eprinttype =. 2508.15260 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2508.15260 2025

  20. [20]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers),

    Jiahui Geng and Fengyu Cai and Yuxia Wang and Heinz Koeppl and Preslav Nakov and Iryna Gurevych , title =. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers),. 2024 , url =. doi:10.18653/V1/2024.NAACL-LONG.366 , timestamp =

    work page doi:10.18653/v1/2024.naacl-long.366 2024

  21. [21]

    Fact-Checking the Output of Large Language Models via Token-Level Uncertainty Quantification

    Fadeeva, Ekaterina and Rubashevskii, Aleksandr and Shelmanov, Artem and Petrakov, Sergey and Li, Haonan and Mubarak, Hamdy and Tsymbalov, Evgenii and Kuzmin, Gleb and Panchenko, Alexander and Baldwin, Timothy and Nakov, Preslav and Panov, Maxim. Fact-Checking the Output of Large Language Models via Token-Level Uncertainty Quantification. Findings of the A...

    work page doi:10.18653/v1/2024.findings-acl.558 2024

  22. [22]

    Qwen3 Technical Report

    An Yang and Anfeng Li and Baosong Yang and Beichen Zhang and Binyuan Hui and Bo Zheng and Bowen Yu and Chang Gao and Chengen Huang and Chenxu Lv and Chujie Zheng and Dayiheng Liu and Fan Zhou and Fei Huang and Feng Hu and Hao Ge and Haoran Wei and Huan Lin and Jialong Tang and Jian Yang and Jianhong Tu and Jianwei Zhang and Jian Yang and Jiaxi Yang and Ji...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.09388 2025

  23. [23]

    Measuring Mathematical Problem Solving With the

    Dan Hendrycks and Collin Burns and Saurav Kadavath and Akul Arora and Steven Basart and Eric Tang and Dawn Song and Jacob Steinhardt , editor =. Measuring Mathematical Problem Solving With the. Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS Datasets and Benchmarks 2021, December 2021, virtual , year =

    2021

  24. [24]

    The Twelfth International Conference on Learning Representations,

    Hunter Lightman and Vineet Kosaraju and Yuri Burda and Harrison Edwards and Bowen Baker and Teddy Lee and Jan Leike and John Schulman and Ilya Sutskever and Karl Cobbe , title =. The Twelfth International Conference on Learning Representations,. 2024 , url =

    2024

  25. [25]

    GPQA: A Graduate-Level Google-Proof Q&A Benchmark

    David Rein and Betty Li Hou and Asa Cooper Stickland and Jackson Petty and Richard Yuanzhe Pang and Julien Dirani and Julian Michael and Samuel R. Bowman , title =. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2311.12022 , eprinttype =. 2311.12022 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2311.12022 2023

  26. [26]

    CoRR , volume =

    Hugh Zhang and Daniel Duckworth and Daphne Ippolito and Arvind Neelakantan , title =. CoRR , volume =. 2020 , url =. 2004.10450 , timestamp =

    work page arXiv 2020

  27. [27]

    Varshney , title =

    Sourya Basu and Govardana Sachitanandam Ramachandran and Nitish Shirish Keskar and Lav R. Varshney , title =. 9th International Conference on Learning Representations,. 2021 , url =

    2021

  28. [28]

    CoRR , volume =

    Yuyang Wu and Yifei Wang and Tianqi Du and Stefanie Jegelka and Yisen Wang , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2502.07266 , eprinttype =. 2502.07266 , timestamp =

    work page doi:10.48550/arxiv.2502.07266 2025

  29. [29]

    CoRR , volume =

    Yue Wang and Qiuzhi Liu and Jiahao Xu and Tian Liang and Xingyu Chen and Zhiwei He and Linfeng Song and Dian Yu and Juntao Li and Zhuosheng Zhang and Rui Wang and Zhaopeng Tu and Haitao Mi and Dong Yu , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2501.18585 , eprinttype =. 2501.18585 , timestamp =

    work page doi:10.48550/arxiv.2501.18585 2025

  30. [30]

    Findings of the Association for Computational Linguistics,

    Yingqian Cui and Pengfei He and Jingying Zeng and Hui Liu and Xianfeng Tang and Zhenwei Dai and Yan Han and Chen Luo and Jing Huang and Zhen Li and Suhang Wang and Yue Xing and Jiliang Tang and Qi He , title =. Findings of the Association for Computational Linguistics,. 2025 , url =

    2025

  31. [31]

    2025 , issn =

    Toward large reasoning models: A survey of reinforced reasoning with large language models , journal =. 2025 , issn =. doi:https://doi.org/10.1016/j.patter.2025.101370 , url =

    work page doi:10.1016/j.patter.2025.101370 2025

  32. [32]

    L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning

    Pranjal Aggarwal and Sean Welleck , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2503.04697 , eprinttype =. 2503.04697 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.04697 2025

  33. [33]

    ArXiv , year=

    O1-Pruner: Length-Harmonizing Fine-Tuning for O1-Like Reasoning Pruning , author=. ArXiv , year=

  34. [34]

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale

    Qiying Yu and Zheng Zhang and Ruofei Zhu and Yufeng Yuan and Xiaochen Zuo and Yu Yue and Tiantian Fan and Gaohong Liu and Lingjun Liu and Xin Liu and Haibin Lin and Zhiqi Lin and Bole Ma and Guangming Sheng and Yuxuan Tong and Chi Zhang and Mofan Zhang and Wang Zhang and Hang Zhu and Jinhua Zhu and Jiaze Chen and Jiangjie Chen and Chengyi Wang and Hongli ...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.14476 2025

  35. [35]

    Token-Budget-Aware

    Tingxu Han and Zhenting Wang and Chunrong Fang and Shiyu Zhao and Shiqing Ma and Zhenyu Chen , editor =. Token-Budget-Aware. Findings of the Association for Computational Linguistics,. 2025 , url =

    2025

  36. [36]

    2nd International Conference on Foundation and Large Language Models,

    Matthew Renze and Erhan Guven , title =. 2nd International Conference on Foundation and Large Language Models,. 2024 , url =. doi:10.1109/FLLM63129.2024.10852493 , timestamp =

    work page doi:10.1109/fllm63129.2024.10852493 2024

  37. [37]

    Concise thoughts: Impact of output length on llm reasoning and cost.arXiv preprint arXiv:2407.19825,

    Sania Nayab and Giulio Rossolini and Giorgio C. Buttazzo and Nicolamaria Manes and Fabrizio Giacomelli , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2407.19825 , eprinttype =. 2407.19825 , timestamp =

    work page doi:10.48550/arxiv.2407.19825 2024

  38. [38]

    Confident or Seek Stronger: Exploring Uncertainty-Based On-device

    Yu. Confident or Seek Stronger: Exploring Uncertainty-Based On-device. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2502.04428 , eprinttype =. 2502.04428 , timestamp =

    work page doi:10.48550/arxiv.2502.04428 2025

  39. [39]

    RouteLLM: Learning to Route LLMs from Preference Data , booktitle =

    Isaac Ong and Amjad Almahairi and Vincent Wu and Wei. RouteLLM: Learning to Route LLMs from Preference Data , booktitle =. 2025 , url =

    2025

  40. [40]

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

    Zhenyi Shen and Hanqi Yan and Linhai Zhang and Zhanghao Hu and Yali Du and Yulan He , editor =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,. 2025 , url =. doi:10.18653/V1/2025.EMNLP-MAIN.36 , timestamp =

    work page doi:10.18653/v1/2025.emnlp-main.36 2025

  41. [41]

    Training Large Language Models to Reason in a Continuous Latent Space

    Shibo Hao and Sainbayar Sukhbaatar and DiJia Su and Xian Li and Zhiting Hu and Jason Weston and Yuandong Tian , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2412.06769 , eprinttype =. 2412.06769 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.06769 2024

  42. [42]

    Bartlett and Andrea Zanette , editor =

    Hanshi Sun and Momin Haider and Ruiqi Zhang and Huitao Yang and Jiahao Qiu and Ming Yin and Mengdi Wang and Peter L. Bartlett and Andrea Zanette , editor =. Fast Best-of-N Decoding via Speculative Rejection , booktitle =. 2024 , url =

    2024

  43. [43]

    Chain of draft: Thinking faster by writing less.arXiv preprint arXiv:2502.18600, 2025a

    Silei Xu and Wenhao Xie and Lingxiao Zhao and Pengcheng He , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2502.18600 , eprinttype =. 2502.18600 , timestamp =

    work page doi:10.48550/arxiv.2502.18600 2025

  44. [44]

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

    Jiaxin Zhang , title =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,. 2025 , url =

    2025

  45. [45]

    American Invitational Mathematics Examination-AIME 2024 , author=

    2024

  46. [46]

    Open R1: A fully open reproduction of DeepSeek-R1 , url =

  47. [47]

    Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen

    Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen. LoRA: Low-Rank Adaptation of Large Language Models , booktitle =. 2022 , url =

    2022

  48. [48]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  49. [49]

    Publications Manual , year = "1983", publisher =

    1983

  50. [50]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    work page doi:10.1145/322234.322243 1981

  51. [51]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  52. [52]

    Dan Gusfield , title =. 1997

    1997

  53. [53]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  54. [54]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =