Pith. sign in

REVIEW 5 major objections 5 minor 8 cited by

BudgetThinker: Empowering Budget-aware LLM Reasoning with Control Tokens

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that periodically inserting eight control tokens at fractions of a user-specified token budget, then training with SFT and GRPO, makes reasoning models follow the budget precisely and improves accuracy by 4.9% on average.

desk verdict A genuinely new control-token mechanism for budget-aware reasoning, but the central accuracy claim rests on a data-hygiene question that must be answered before trusting the numbers. read the letter →

arxiv 2508.17196 v2 pith:MI3IKDKL submitted 2025-08-24 cs.LG cs.AI

Hao Wen , Xinrui Wu , Yi Sun , Feifei Zhang , Liye Chen , Jie Wang , Yunxin Liu , Yunhao Liu
show 2 more authors
Ya-Qin Zhang Yuanchun Li
This is my paper
classification cs.LGcs.AI
keywords budget-awarereasoningcontroltokenschain-of-thoughttokenbudgetreinforcementlearningGRPOsupervisedfine-tuningmathematical
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 tries to establish that a language model can be trained to reason for a user-specified number of tokens, rather than thinking as long as it likes or stopping early. The method inserts a fixed set of control tokens at fractions of the requested budget during generation, so the model continuously sees how much budget remains. Supervised fine-tuning teaches the meaning of the tokens, and then reinforcement learning with a length-aware reward sharpens both accuracy and adherence. If this works, applications with latency or cost limits can set an explicit reasoning budget and get predictable behavior.

What carries the argument

The central object is the control-token set C = {c1,...,c8}, inserted deterministically at budget-relative positions t = k·floor(B/8). Token c_k is a learned embedding that tells the model the generation has consumed k/8 of the budget. Because the insertion positions scale with B, the same eight tokens serve any budget, unlike fixed-interval signaling which would need a new token per interval. The second mechanism is the asymmetric length reward used during GRPO: reward includes k3·max(1 - gamma·((B-|y|)/B)^2, 0), with gamma=16 for overshoot and gamma=1 for undershoot, which makes exact budget adherence the optimized behavior.

What would settle it

Use the released code to evaluate the final model at budgets not present in the training curriculum, such as 2500, 3500, or 8000 tokens. If the budget-following ratio and accuracy at these budgets drop substantially below what interpolation from the trained budgets (2000, 3000, 4000, 6000, 10000) predicts, then control-token semantics do not transfer to arbitrary budgets. A cleaner causal probe: swap the order of the eight control tokens during a fixed-length generation; if the model's stopping point shifts with the token rather than the position, the tokens carry the claimed fraction semantic

Watch

Extended reading notes

Core claim

On its own terms, BudgetThinker claims that budget following is a trainable skill and that the training signal must be repeated during generation, not only stated once in the prompt. Eight special tokens are inserted at positions k·floor(B/8); each one signals that k/8 of the budget has been spent. The model is first fine-tuned on reasoning traces rewritten to include these tokens, then optimized with GRPO using a reward that penalizes overshooting the budget 16 times more heavily than undershooting it. On MATH-500 and AMC 2023, across budgets from 500 to 10,000 tokens, the method improves pass@1 accuracy by 4.9% on average over the base model and the efficient-reasoning baseline while termi

Load-bearing premise

The load-bearing premise is that the model reads the control tokens as fractions of the current budget rather than as cues tied to the absolute token positions where the tokens appeared during training; if it latches onto positions, budgets outside the training range would stop working.

Editorial extensions

If this is right

  • Operators can set a hard token budget per query and expect the model to stop near it, making latency and per-query cost more predictable.
  • At a fixed budget, the trained models score higher on MATH-500 and AMC 2023 than either the untrained base model or the RL-only efficient-reasoning baseline; the reported average gain is 4.9% across budgets.
  • The method inserts only eight tokens and uses standard SFT plus GRPO, so it can be layered onto existing long-CoT training recipes without changing the base model architecture.
  • Training through progressively smaller budgets makes the model more concise even on large budgets, and a final mixed-budget phase recovers accuracy on large budgets.
  • On AIME 2024 the three methods perform comparably, so the benefit is not uniform across all problem types.

Reading between the lines

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

  • The paper only tests token budgets; the same ratio-based signaling idea could be applied to other measurable budgets such as wall-clock latency or API cost if a position-based progress signal can be attached.
  • The curriculum always goes from large to small budgets, assuming conciseness is the harder skill; the reverse order or a random order would test whether the direction matters.
  • The control tokens are effectively a progress bar; one could use the final fraction to trigger a verification mode, a behavioral switch the paper does not explore.
  • Transfer beyond the training range is evaluated at 500 and 10,000 tokens, but the mechanism by which the model generalizes to arbitrary intermediate budgets is not analyzed.
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 / 5 minor

Summary. BudgetThinker proposes a framework for budget-aware LLM reasoning in which a fixed set of K special control tokens is deterministically inserted at budget-relative positions during generation, continuously signaling the remaining token budget. The method trains a backbone model (DeepSeek-R1-distilled Qwen-2.5-1.5B/7B) in two stages: SFT on budget-reconstructed examples, followed by curriculum-based GRPO with a length-aware reward that penalizes deviation from the target budget. The paper reports that BudgetThinker improves accuracy by an average of 4.9% over baselines and adheres more precisely to specified token budgets on MATH-500 and AMC 2023, while being comparable to baselines on AIME 2024. Ablations examine control-token insertion strategies, iterative training, and the contribution of RL.

Significance. If the empirical claims are validated, this is a useful and practical contribution to controllable test-time computation. The method is simple, plug-and-play, and uses a fixed token set that is intended to scale to arbitrary budgets. The paper ships source code and includes ablations that isolate the control tokens, the curriculum, and the RL phase. However, the central empirical claim currently depends on unresolved data-hygiene and statistical-reporting questions, so the significance can only be assessed after those are addressed.

major comments (5)
  1. [§3.1 (Training Details / Evaluation Details)] The RL training corpus is stated to include “the numina_amc_aime labeled data in the PRIME-RL/Eurus-2-RL-Data dataset,” while the evaluation benchmarks are AMC 2023 and AIME 2024. The paper never reports whether AMC 2023 / AIME 2024 problems were excluded from this training split, nor does it provide any overlap or deduplication analysis. If the training split contains exact or near-duplicate evaluation problems, then Table 1 and Figure 2 accuracy deltas—and the headline 4.9% average—are contaminated and cannot be attributed to BudgetThinker. Please provide a deduplication check and, if overlap exists, re-evaluate on clean subsets.
  2. [§3.2, Figure 2, Table 1] No error bars, confidence intervals, or significance tests are reported. AMC 2023 and AIME 2024 evaluations use 64 samples at temperature 0.6, so pass@1 estimates have sampling variance; MATH-500 is a 500-problem set. Many reported differences are small (e.g., AIME at B=10000: 28.90 vs 28.88; MATH-500 several 1–2 point gaps). The claim of consistent, average +4.9% improvement needs statistical support such as multiple seeds, bootstrap confidence intervals, or paired significance tests. Without this, the central empirical claim is underdetermined.
  3. [Abstract and §3.2] The abstract claims the method “improves accuracy by an average of 4.9% across all tested budgets,” but Section 3.2 states that on AIME 2024 all methods are comparable, and Table 1 shows BudgetThinker slightly below Original-1.5B at B=6000 (25.63 vs 25.71) and essentially tied at B=10000. The 4.9% figure appears to be an average over MATH-500 and AMC 2023 only, not over all tested benchmarks. The claim should be scoped precisely to the benchmarks and budgets where the improvement actually holds.
  4. [§2.1 and §3.3] The budget-ratio insertion strategy is explicitly motivated by scalability to arbitrary budgets (e.g., B=500 and B=10000 outside the RL training set {2000,3000,4000,6000}). However, the paper does not analyze whether the model learns budget-relative semantics or exploits absolute position cues: for B=6000, c_1 appears around token 750, while for B=500 it appears around token 62. If the model relies on absolute positions, transfer to unseen budgets would not be a property of the method. The empirical results suggest transfer works, but a mechanistic analysis (probing token representations, ablating insertion positions, testing intermediate budgets) is needed to support the design claim.
  5. [§3.2, Baselines] The only trained baseline is ThinkPrune, and only at the 1.5B scale because the official repository does not provide a 7B model. The “original model” baseline has not been trained for budget control. The paper’s phrase “surpasses strong baselines” is therefore overstated. Missing comparisons with other length-control RL methods (e.g., L1, DAPO, TokenSkip, and other budget-aware approaches) limit external validity. At minimum, the comparison scope should be stated clearly and the claim softened.
minor comments (5)
  1. [Eq. (1), §2.1] The indexing is unclear: for k in {0,...,K−1}, the insertion rule places c_1 at t=0 and c_K at t=(K−1)·floor(B/K), leaving the final budget fraction without a signal. Clarify whether c_{K+1} exists or whether the intended positions are k·floor(B/K) for k=1,...,K.
  2. [§2.2] The Insert function is described as adding a token “when t = k·I,” but I is not defined in this section. It should be floor(B/K) or a budget-ratio-based interval to match the method.
  3. [§3.1] For MATH-500, evaluation uses “temperature of 0 and top-p of 1.” Top-p is irrelevant for greedy decoding; please clarify the sampling configuration.
  4. [§3.5 and Figure 4b] The text says the length reward will “decade to 0” when the budget is exceeded by 1/4; this should be “decay.”
  5. [Table 2] The checkpoint labels are inconsistent: “BudgetThinker 6k” should be “6k” or “6k-only” to parallel the later rows. Also, the darker-cell highlight is not explained in the caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the derivation is self-contained and the empirical claims are tested on held-out budgets.

full rationale

BudgetThinker's derivation chain is a standard train-then-evaluate pipeline. The control-token insertion rule (Sec. 2.1) is defined independently of the evaluation metrics: token positions are fixed fractions of budget B and the token set C is fixed (K=8). The SFT targets are constructed from source datasets with budgets B = T*ceil(|y|/T) (Sec. 2.2), and the RL reward in Sec. 2.3 includes a length term that directly penalizes deviation from B. This means improved budget adherence is partly a training objective, but the paper does not present adherence as an unprompted prediction; moreover, the evaluation includes budgets B=500 and B=10000 that never appear in the RL curriculum {6000,4000,3000,2000} (Sec. 3.1), so the generalization claim is not forced by construction. The only overlapping-author citations (Sun et al. 2025; Wen et al. 2023; Jiang et al. 2021; Han et al. 2021) are background/latency citations and are not load-bearing. The reviewer's AMC/AIME data-overlap concern is a data-hygiene/correctness issue, not a circularity: even if training and evaluation sets overlap, that would be leakage, not an equivalence between the paper's equations and its conclusions. No equation is defined in terms of the result it is used to establish, and no fitted constant is renamed as a prediction. Therefore no circular step can be exhibited.

Assumptions & free parameters 7 free parameters · 4 assumptions · 1 invented entities

The central claim rests on a set of hand-chosen hyperparameters (K, T, reward coefficients, curriculum) and on several domain assumptions about the validity of GRPO, exact-match rewards, dataset transfer, and the learnability of control-token semantics. No physical entities or external predictions are introduced; the control tokens are engineered inputs, not explanatory constructs.

free parameters (7)
  • K (number of control intervals) = 8
    Hand-chosen; determines how many control tokens are inserted during generation and training.
  • T (budget granularity) = 50
    Used to round up answer lengths to the nearest multiple of T when assigning training budgets.
  • k1 (correctness reward weight) = 0.7
    Reward coefficient for exact-match correctness; hand-selected.
  • k2 (format reward weight) = 0.15
    Reward coefficient for format check; hand-selected.
  • k3 (length reward weight) = 0.15
    Reward coefficient for length adherence; hand-selected.
  • gamma (over-budget penalty multiplier) = 16
    Asymmetric penalty applied when generated length exceeds budget; chosen to make the length reward decay to zero when exceeding by 1/4.
  • Curriculum budget sequence = 6000, 4000, 3000, 2000
    Hand-selected schedule for the iterative RL curriculum; the paper does not justify these specific values.
assumptions (4)
  • domain assumption GRPO is a valid and effective RL algorithm for fine-tuning LLMs to follow token-budget constraints.
    The paper relies on GRPO (Shao et al., 2024) for the RL phase; no alternative or analysis is provided for its appropriateness.
  • domain assumption Exact-match correctness (with format check) is a reliable proxy for mathematical solution correctness.
    The reward function uses 1{y = ygold}; this ignores partial credit and may mis-score paraphrased but correct answers.
  • domain assumption Training on the selected SFT mix (s1k, LIMO, Bespoke-Stratos, NuminaMath, MATH) transfers to the evaluation benchmarks (MATH-500, AMC 2023, AIME 2024).
    The paper assumes cross-dataset generalization without measuring distribution shift.
  • domain assumption LLMs can learn the semantics of inserted control tokens from SFT targets and use them at inference.
    The entire mechanism depends on the model internalizing control tokens as budget-relative signals, which is shown empirically but not proven.
invented entities (1)
  • Control tokens (c_1...c_K)
    purpose: Inserted during generation to signal remaining budget fractions to the model, and included in SFT targets to teach budget adherence.
    These are new vocabulary tokens introduced by the method. They have no empirical handle outside the paper's training and evaluation; their effectiveness is only demonstrated within the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BudgetThinker: Empowering Budget-aware LLM Reasoning with Control Tokens." pith.science (2026). https://pith.science/paper/MI3IKDKL

@misc{pith2026250817196,
  author       = {Pith},
  title        = {Pith review of: BudgetThinker: Empowering Budget-aware LLM Reasoning with Control Tokens},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MI3IKDKL}},
  note         = {Machine review of arXiv:2508.17196}
}
read the original abstract

Recent advancements in Large Language Models (LLMs) have leveraged increased test-time computation to enhance reasoning capabilities, a strategy that, while effective, incurs significant latency and resource costs, limiting their applicability in real-world time-constrained or cost-sensitive scenarios. This paper introduces BudgetThinker, a novel framework designed to empower LLMs with budget-aware reasoning, enabling precise control over the length of their thought processes. We propose a methodology that periodically inserts special control tokens during inference to continuously inform the model of its remaining token budget. This approach is coupled with a comprehensive two-stage training pipeline, beginning with Supervised Fine-Tuning (SFT) to familiarize the model with budget constraints, followed by a curriculum-based Reinforcement Learning (RL) phase that utilizes a length-aware reward function to optimize for both accuracy and budget adherence. We demonstrate that BudgetThinker significantly surpasses strong baselines in maintaining performance across a variety of reasoning budgets on challenging mathematical benchmarks. Our method provides a scalable and effective solution for developing efficient and controllable LLM reasoning, making advanced models more practical for deployment in resource-constrained and real-time environments.

Discussion (0). Sign in to comment.

Forward citations

Cited by 8 Pith papers

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

  1. Budget-Aware LLM Discovery via Cost-Calibrated Frontier Utility

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Cost-blind LLM discovery controllers can waste almost all attainable quality under heterogeneous action costs; CostAda’s cost-calibrated frontier utility recovers budget efficiency and final quality across eight benchmarks.

  2. Budget-Aware LLM Discovery via Cost-Calibrated Frontier Utility

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Cost-calibrated frontier utility — progress divided by realized token cost and weighted by remaining budget — lets CostAda reach full-budget quality at half the budget in 12 of 16 benchmark–backbone pairs and achieve ...

  3. SeqRoute: Global Budget-Aware Sequential LLM Routing via Offline Reinforcement Learning

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    SeqRoute applies offline RL with CQL and Hindsight Budget Relabeling to sequential LLM routing under global budgets, claiming 6.0-73.5% cost reduction, maintained or improved quality, and under 1% bankruptcy rate.

  4. Nice Fold or Hero Call: Learning Budget-Efficient Thinking for Adaptive Reasoning

    cs.AI 2026-05 unverdicted novelty 6.0 of 10

    BET reduces reasoning tokens by about 55% on average while improving performance across benchmarks by learning to short-solve easy queries, fold early on unsolvable ones, and preserve budget for hard solvable queries.

  5. Behavior Cue Reasoning: Monitorable Reasoning Improves Efficiency and Safety through Oversight

    cs.AI 2026-05 conditional novelty 6.0 of 10

    Behavior Cue Reasoning trains LLMs to emit special tokens before behaviors, enabling monitors to cut up to 50% wasted reasoning tokens and recover safe actions from 80% of unsafe traces, more than doubling success rat...

  6. Behavior Cue Reasoning: Monitorable Reasoning Improves Efficiency and Safety through Oversight

    cs.AI 2026-05 unverdicted novelty 6.0 of 10

    Behavior Cue Reasoning trains LLMs to emit special tokens before behaviors, enabling monitors to prune up to 50% of wasted tokens and recover safe actions from 80% of unsafe traces, more than doubling success rates wi...

  7. ParaThinker: Native Parallel Thinking as a New Paradigm to Scale LLM Test-time Compute

    cs.CL 2025-08 conditional novelty 6.0 of 10

    ParaThinker trains LLMs for native parallel reasoning and reports 7 to 12 percent higher accuracy on math benchmarks over sequential thinking with modest latency overhead.

  8. Beyond Penalizing Mistakes: Stabilizing Efficiency Training in Large Reasoning Models via Adaptive Correct-Only Rewards

    cs.AI 2026-06 unverdicted novelty 5.0 of 10

    ACOER applies adaptive correct-only efficiency rewards in GRPO to avoid reward collapse, yielding higher accuracy and over 60% fewer tokens on math reasoning benchmarks.

Reference graph

Works this paper leans on

51 extracted references · 6 canonical work pages · cited by 6 Pith papers

  1. [1]

    L1: Controlling how long a reasoning model thinks with reinforcement learning, 2025

    Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning, 2025. URL https://arxiv.org/abs/2503.04697

    arXiv 2025

  2. [2]

    Training language models to reason efficiently

    Daman Arora and Andrea Zanette. Training language models to reason efficiently. arXiv preprint arXiv:2502.04463, 2025

    arXiv 2025

  3. [3]

    Typefly: Flying drones with large language model

    Guojun Chen, Xiaojing Yu, Neiwen Ling, and Lin Zhong. Typefly: Flying drones with large language model. arXiv preprint arXiv:2312.14950, 2023

    arXiv 2023

  4. [4]

    Timelynet: Adaptive neural architecture for autonomous driving with dynamic deadline

    Jiale Chen, Duc Van Le, Yuanchun Li, Yunxin Liu, and Rui Tan. Timelynet: Adaptive neural architecture for autonomous driving with dynamic deadline. In Proceedings of the International Conference on Embedded Software (EMSOFT), 2025 a

    work page 2025

  5. [5]

    Do not think that much for 2+3=? on the overthinking of o1-like llms, 2025 b

    Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, and Dong Yu. Do not think that much for 2+3=? on the overthinking of o1-like llms, 2025 b . URL https://arxiv.org/abs/2412.21187

    arXiv 2025

  6. [6]

    Process reinforcement through implicit rewards, 2025

    Ganqu Cui, Lifan Yuan, Zefan Wang, Hanbin Wang, Wendi Li, Bingxiang He, Yuchen Fan, Tianyu Yu, Qixin Xu, Weize Chen, Jiarui Yuan, Huayu Chen, Kaiyan Zhang, Xingtai Lv, Shuo Wang, Yuan Yao, Xu Han, Hao Peng, Yu Cheng, Zhiyuan Liu, Maosong Sun, Bowen Zhou, and Ning Ding. Process reinforcement through implicit rewards, 2025. URL https://arxiv.org/abs/2502.01456

    arXiv 2025

  7. [7]

    Learning how hard to think: Input-adaptive allocation of lm computation, 2024

    Mehul Damani, Idan Shenfeld, Andi Peng, Andreea Bobu, and Jacob Andreas. Learning how hard to think: Input-adaptive allocation of lm computation, 2024. URL https://arxiv.org/abs/2410.04707

    arXiv 2024

  8. [8]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025

    DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025. URL https://arxiv.org/abs/2501.12948

    arXiv 2025

Show all 51 references
  1. [9]

    Efficiently scaling llm reasoning with certaindex, 2025

    Yichao Fu, Junda Chen, Siqi Zhu, Zheyu Fu, Zhongdongming Dai, Yonghao Zhuang, Yian Ma, Aurick Qiao, Tajana Rosing, Ion Stoica, and Hao Zhang. Efficiently scaling llm reasoning with certaindex, 2025. URL https://arxiv.org/abs/2412.20993

    2025 arXiv

  2. [10]

    Bartoldson, Bhavya Kailkhura, Abhinav Bhatele, and Tom Goldstein

    Jonas Geiping, Sean McLeish, Neel Jain, John Kirchenbauer, Siddharth Singh, Brian R. Bartoldson, Bhavya Kailkhura, Abhinav Bhatele, and Tom Goldstein. Scaling up test-time compute with latent reasoning: A recurrent depth approach. CoRR, abs/2502.05171, February 2025. URL https...

    2025 doi arXiv

  3. [11]

    Gemini 2.5: Our most intelligent ai model

    Google. Gemini 2.5: Our most intelligent ai model. https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/#gemini-2-5-thinking, March 2025

    2025

  4. [13]

    Token-budget-aware llm reasoning

    Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. Token-budget-aware llm reasoning. arXiv preprint arXiv:2412.18547, 2024

    2024 arXiv

  5. [14]

    Measuring mathematical problem solving with the math dataset

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. NeurIPS, 2021

    2021

  6. [15]

    Thinkprune: Pruning long chain-of-thought of llms via reinforcement learning, 2025

    Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu, Kaizhi Qian, Jacob Andreas, and Shiyu Chang. Thinkprune: Pruning long chain-of-thought of llms via reinforcement learning, 2025. URL https://arxiv.org/abs/2504.01296

    2025 arXiv

  7. [16]

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Kai Dang, et al. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186, 2024

    2024 arXiv

  8. [17]

    Flexible high-resolution object detection on edge devices with tunable latency

    Shiqi Jiang, Zhiqi Lin, Yuanchun Li, Yuanchao Shu, and Yunxin Liu. Flexible high-resolution object detection on edge devices with tunable latency. In Proceedings of the 27th Annual International Conference on Mobile Computing and Networking, MobiCom '21, pp.\ 559–572, New York...

    2021

  9. [18]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems...

    2023

  10. [19]

    Bespoke-stratos: The unreasonable effectiveness of reasoning distillation

    Bespoke Labs. Bespoke-stratos: The unreasonable effectiveness of reasoning distillation. https://www.bespokelabs.ai/blog/bespoke-stratos-the-unreasonable-effectiveness-of-reasoning-distillation, 2025. Accessed: 2025-01-22

    2025

  11. [20]

    Numinamath

    Jia LI, Edward Beeching, Lewis Tunstall, Ben Lipkin, Roman Soletskyi, Shengyi Costa Huang, Kashif Rasul, Longhui Yu, Albert Jiang, Ziju Shen, Zihan Qin, Bin Dong, Li Zhou, Yann Fleureau, Guillaume Lample, and Stanislas Polu. Numinamath. [https://huggingface.co/AI-MO/NuminaMath...

    2024

  12. [21]

    Selfbudgeter: Adaptive token allocation for efficient llm reasoning

    Zheng Li, Qingxiu Dong, Jingyuan Ma, Di Zhang, and Zhifang Sui. Selfbudgeter: Adaptive token allocation for efficient llm reasoning. arXiv preprint arXiv:2505.11274, 2025

    2025 arXiv

  13. [22]

    Let's verify step by step, 2023

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step, 2023. URL https://arxiv.org/abs/2305.20050

    2023 arXiv

  14. [23]

    The architectural implications of autonomous driving: Constraints and acceleration

    Shih-Chieh Lin, Yunqi Zhang, Chang-Hong Hsu, Matt Skach, Md E Haque, Lingjia Tang, and Jason Mars. The architectural implications of autonomous driving: Constraints and acceleration. In Proceedings of the twenty-third international conference on architectural support for progr...

    2018

  15. [24]

    s1: Simple test-time scaling

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Cand \`e s, and Tatsunori Hashimoto. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393, 2025

    2025 arXiv

  16. [25]

    Openai o1 system card

    OpenAI. Openai o1 system card. 2024. URL https://arxiv.org/abs/2412.16720

    2024 arXiv

  17. [26]

    Gpt-5 system card

    OpenAI. Gpt-5 system card. https://cdn.openai.com/gpt-5-system-card.pdf, August 2025. Accessed: 2025-08-21

    2025

  18. [27]

    Thoughtterminator: Benchmarking, calibrating, and mitigating overthinking in reasoning models

    Xiao Pu, Michael Saxon, Wenyue Hua, and William Yang Wang. Thoughtterminator: Benchmarking, calibrating, and mitigating overthinking in reasoning models. arXiv preprint arXiv:2504.13367, 2025

    2025 arXiv

  19. [28]

    Yuxiao Qu, Matthew Y. R. Yang, Amrith Setlur, Lewis Tunstall, Edward Emanuel Beeching, Ruslan Salakhutdinov, and Aviral Kumar. Optimizing test-time compute via meta reinforcement fine-tuning, 2025. URL https://arxiv.org/abs/2503.07572

    2025 arXiv

  20. [29]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. URL https://arxiv.org/abs/2402.03300

    2024 arXiv

  21. [30]

    Vlm-r1: A stable and generalizable r1-style large vision-language model

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, et al. Vlm-r1: A stable and generalizable r1-style large vision-language model. arXiv preprint arXiv:2504.07615, 2025

    2025 arXiv

  22. [31]

    An empirical study of llm reasoning ability under strict output length constraint

    Yi Sun, Han Wang, Jiaqiang Li, Jiacheng Liu, Xiangyu Li, Hao Wen, Yizhen Yuan, Huiwen Zheng, Yan Liang, Yuanchun Li, and Yunxin Liu. An empirical study of llm reasoning ability under strict output length constraint. arXiv preprint arXiv:2504.14350, April 2025. doi:10.48550/arX...

    2025 doi arXiv

  23. [32]

    Reason-rft: Reinforcement fine-tuning for visual reasoning

    Huajie Tan, Yuheng Ji, Xiaoshuai Hao, Minglan Lin, Pengwei Wang, Zhongyuan Wang, and Shanghang Zhang. Reason-rft: Reinforcement fine-tuning for visual reasoning. arXiv preprint arXiv:2503.20752, 2025

    2025

  24. [33]

    Kimi k2: Open agentic intelligence, 2025 a

    Kimi Team. Kimi k2: Open agentic intelligence, 2025 a . URL https://arxiv.org/abs/2507.20534

    2025 arXiv

  25. [34]

    Kimi k1.5: Scaling reinforcement learning with llms, 2025

    Kimi Team et al. Kimi k1.5: Scaling reinforcement learning with llms, 2025. URL https://arxiv.org/abs/2501.12599

    2025 arXiv

  26. [35]

    Sky-t1: Train your own o1 preview model within \ 450

    NovaSky Team. Sky-t1: Train your own o1 preview model within \ 450. https://novasky-ai.github.io/posts/sky-t1, 2025 b . Accessed: 2025-01-09

    2025

  27. [36]

    Stepwise informativeness search for efficient and effective llm reasoning, 2025 a

    Siyuan Wang, Enda Zhao, Zhongyu Wei, and Xiang Ren. Stepwise informativeness search for efficient and effective llm reasoning, 2025 a . URL https://arxiv.org/abs/2502.15335

    2025 arXiv

  28. [37]

    Make every penny count: Difficulty-adaptive self-consistency for cost-efficient reasoning, 2025 b

    Xinglin Wang, Shaoxiong Feng, Yiwei Li, Peiwen Yuan, Yueqi Zhang, Chuyi Tan, Boyuan Pan, Yao Hu, and Kan Li. Make every penny count: Difficulty-adaptive self-consistency for cost-efficient reasoning, 2025 b . URL https://arxiv.org/abs/2408.13457

    2025 arXiv

  29. [38]

    Adaptivenet: Post-deployment neural architecture adaptation for diverse edge environments

    Hao Wen, Yuanchun Li, Zunshuai Zhang, Shiqi Jiang, Xiaozhou Ye, Ye Ouyang, Yaqin Zhang, and Yunxin Liu. Adaptivenet: Post-deployment neural architecture adaptation for diverse edge environments. In Proceedings of the 29th Annual International Conference on Mobile Computing and...

    2023

  30. [39]

    Lapo: Internalizing reasoning efficiency via length-adaptive policy optimization

    Xingyu Wu, Yuchen Yan, Shangke Lyu, Linjuan Wu, Yiwen Qiu, Yongliang Shen, Weiming Lu, Jian Shao, Jun Xiao, and Yueting Zhuang. Lapo: Internalizing reasoning efficiency via length-adaptive policy optimization. arXiv preprint arXiv:2507.15758, 2025

    2025 arXiv

  31. [40]

    Tokenskip: Controllable chain-of-thought compression in llms, 2025

    Heming Xia, Chak Tou Leong, Wenjie Wang, Yongqi Li, and Wenjie Li. Tokenskip: Controllable chain-of-thought compression in llms, 2025. URL https://arxiv.org/abs/2502.12067

    2025

  32. [41]

    Redstar: Does scaling long-cot data unlock better slow-reasoning systems?, 2025 a

    Haotian Xu, Xing Wu, Weinong Wang, Zhongzhi Li, Da Zheng, Boyuan Chen, Yi Hu, Shijia Kang, Jiaming Ji, Yingying Zhang, Zhijiang Guo, Yaodong Yang, Muhan Zhang, and Debing Zhang. Redstar: Does scaling long-cot data unlock better slow-reasoning systems?, 2025 a . URL https://arx...

    2025 arXiv

  33. [42]

    Chain of draft: Thinking faster by writing less

    Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. Chain of draft: Thinking faster by writing less. arXiv preprint arXiv:2502.18600, 2025 b

    2025 arXiv

  34. [43]

    Qwen3 technical report, 2025

    An Yang, Anfeng Li, Baosong Yang, et al. Qwen3 technical report, 2025. URL https://arxiv.org/abs/2505.09388

    2025 arXiv

  35. [44]

    Limo: Less is more for reasoning

    Yixin Ye, Zhen Huang, Yang Xiao, Ethan Chern, Shijie Xia, and Pengfei Liu. Limo: Less is more for reasoning. arXiv preprint arXiv:2502.03387, 2025

    2025 arXiv

  36. [45]

    Dapo: An open-source llm reinforcement learning system at scale, 2025

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, et al. Dapo: An open-source llm reinforcement learning system at scale, 2025. URL https://arxiv. org/abs/2503.14476, 2025 a

    2025 arXiv

  37. [46]

    Think smarter not harder: Adaptive reasoning with inference aware optimization, 2025 b

    Zishun Yu, Tengyu Xu, Di Jin, Karthik Abinav Sankararaman, Yun He, Wenxuan Zhou, Zhouhao Zeng, Eryk Helenowski, Chen Zhu, Sinong Wang, Hao Ma, and Han Fang. Think smarter not harder: Adaptive reasoning with inference aware optimization, 2025 b . URL https://arxiv.org/abs/2501.17974

    2025 arXiv

  38. [47]

    Pruning the unsurprising: Efficient code reasoning via first-token surprisal, 2025

    Wenhao Zeng, Yaoning Wang, Chao Hu, Yuling Shi, Chengcheng Wan, Hongyu Zhang, and Xiaodong Gu. Pruning the unsurprising: Efficient code reasoning via first-token surprisal, 2025. URL https://arxiv.org/abs/2508.05988

    2025

  39. [48]

    Towards concise and adaptive thinking in large reasoning models: A survey, 2025

    Jason Zhu and Hongyu Li. Towards concise and adaptive thinking in large reasoning models: A survey, 2025. URL https://arxiv.org/abs/2507.09662

    2025 arXiv

  40. [49]

    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...

  41. [50]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  42. [51]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  43. [52]

    thinking

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.