Pith. sign in

REVIEW 4 major objections 6 minor 47 references

Can Past Experience Accelerate LLM Reasoning?

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read LLMs can reason up to 56% faster on familiar questions when equipped with memory and adaptive early stopping.

desk verdict The empirical question is worth asking, but the paper's theoretical guarantee rests on a proof step that is simply wrong: a larger expected score does not imply a larger probability of exceeding a fixed threshold. read the letter →

arxiv 2505.20643 v1 pith:2CWSQ43T submitted 2025-05-27 cs.LG cs.AI

Bo Pan , Liang Zhao This is my paper
classification cs.LGcs.AI
keywords LLMreasoningspeeduptest-timescalingadaptivecomputebudgetmemorymechanismsin-contextlearningquestionsimilarityearlystopping
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 asks whether LLM reasoning, like human skill acquisition, becomes faster with repeated exposure to similar problems. It formalizes reasoning speedup and proposes SpeedupLLM, a framework that combines adaptive compute budget allocation with memory mechanisms. Experiments show compute reductions of up to 56% on similar question streams, with speed gains generally accompanied by accuracy improvements.

What carries the argument

The key mechanism is the combination of adaptive compute budget allocation (early stopping once a generated candidate exceeds a quality threshold $\tau$) and a memory mechanism that accumulates correct question-answer pairs as in-context examples, reflections, or parameter updates. Theorems 1 and 2 formalize how non-decreasing answer quality translates into non-increasing expected compute cost.

What would settle it

Run a sequence of S4-dissimilar questions where the memory buffer is forced to contain only irrelevant or misleading examples and check whether the early-stopping policy produces a lower-accuracy answer than the no-memory baseline; the paper reports such cases exist, which would falsify the universal speedup claim.

Watch

Extended reading notes

Core claim

The paper claims that an LLM's inference cost on a stream of related questions can be reduced without sacrificing answer quality by pairing a test-time scaling method with an early-stopping threshold and a memory of past question-answer pairs. The theoretical corollary states that if accumulating relevant memory does not degrade performance, the expected compute budget is non-increasing over time while the probability of producing a satisfying answer is maintained. Empirically, the speedup appears across four scaling methods and five memory methods, with the largest reductions for highly similar questions.

Load-bearing premise

The central guarantee requires that adding correct memory from relevant questions never makes the model worse; the paper's own low-similarity (S4) results show this assumption can fail.

Editorial extensions

If this is right

  • On streams of near-duplicate or similar questions, LLM services could cut inference compute by up to 56% while maintaining or improving accuracy.
  • In-context episodic memory generally outperforms reflection-based semantic memory for early speedup, suggesting that raw past examples are more immediately useful than distilled rules.
  • Parametric memory (e.g., SFT) keeps improving beyond the context window limit, unlike text-based memory which plateaus once the context is full.
  • The speedup is not tied to a single scaling method; the best combination depends on the memory type, and the correlation between speed and accuracy suggests compute savings do not come at the cost of correctness.
  • The framework provides a unified benchmark for measuring and comparing reasoning speedup across different memory and scaling methods.

Reading between the lines

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

  • The theoretical guarantee relies on memory always helping or being neutral; in real deployments, memory retrieval that pulls in irrelevant or misleading examples could break the monotonic improvement and invalidate the compute-reduction guarantee.
  • The compute metric counts method-specific units (answers, nodes, tokens) rather than wall-clock latency, so the reported 56% reduction may not translate directly to end-to-end speedup on systems with fixed batch or overhead costs.
  • A natural extension is to make memory selection adaptive: instead of always using the last few examples, retrieve past question-answer pairs by similarity to the current query, which could extend the benefit to lower-similarity settings where the paper observes degradation.
  • The benchmark's question sets are generated from MATH problems; testing on broader reasoning domains with noisy or partially correct memory would reveal how sensitive the speedup is to memory quality.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper investigates whether LLMs can reason faster after repeated exposure to similar questions, formalizes this as 'reasoning speedup', and proposes SpeedupLLM, a framework combining memory mechanisms (SFT, in-context, reflection) with adaptive compute-budget allocation (early stopping with a score threshold). The authors prove two theorems and a corollary intended to guarantee non-increasing compute cost with accumulating relevant memory, and they report experiments across four similarity levels, five memory methods, and four test-time scaling methods, claiming up to a 56% compute-cost reduction. The paper concludes that reasoning speedup generally emerges with appropriate memory and higher question similarity, while also acknowledging limitations for low similarity (S4) and long-CoT settings.

Significance. If the theoretical guarantee were valid, the paper would offer a principled framework for a practically important problem: reducing inference cost on streams of similar questions without sacrificing accuracy. The empirical study covers a useful combination of memory and scaling methods and reports a notable negative correlation between compute budget and accuracy, suggesting that adaptive early stopping with memory can improve both efficiency and correctness in some settings. The manuscript also provides a clear taxonomy of question similarity levels, which may be reusable by future benchmarks. However, the core theoretical claim is mathematically flawed, and the empirical evidence is narrow, so the contribution as currently stated cannot be accepted.

major comments (4)
  1. [Section 3.2.2, Theorem 2 and Appendix A.2] Theorem 2 is not proven as stated. The assumption is non-decreasing expected score, but the conclusion is non-decreasing probability of crossing a fixed threshold τ. The proof in Appendix A.2 uses Lemma 1 to lower-bound the tail probability by (E[S]−τ)/(1−τ), but a rising lower bound does not imply a rising actual tail probability: for example, a distribution A with P(A=0.9)=0.5 and P(A=0)=0.5 has E[A]=0.45, while B is deterministic at 0.6 with E[B]=0.6, yet P(A≥0.7)=0.5 > P(B≥0.7)=0. Thus the step from (B0) to the stochastic-dominance claim is invalid. In the independent-candidate argument, the proof simply asserts p_{t+1} ≥ p_t, which is exactly the conclusion that needs to be derived. Consequently, Corollary 1 and the abstract's 'theoretically guaranteed' claim do not follow from the stated conditions.
  2. [Appendix A.2, assumptions A1, A2, B3] The generalized proof introduces assumptions (A1) a fixed DAG, (A2) monotone conditional reliance, and (B3) normalization to [0,1] that are not stated in Theorem 2 nor derived from the experimental setup. The theorem's conclusion is therefore conditional on unstated structural assumptions about the generation process, which is not communicated in the main text. The normalization to the unit interval is also not justified for arbitrary score functions, and the threshold τ is not required to lie in (0,1) in the theorem statement but is used in Lemma 1.
  3. [Section 5.2, Finding 4 vs. abstract] The paper's own Finding 4 shows that at similarity level S4, memory mechanisms can increase compute cost and degrade accuracy. Since S4 is included in the definition of 'past experience' in the problem formulation, the general assertion in the abstract and introduction that 'LLMs can generally reason faster with past experience' is not supported across the full range of similarity levels. The 'Generally' should be qualified to the tested conditions, and the theoretical framework does not predict the S4 failure because the monotonicity premise itself is empirically violated there.
  4. [Section 4 and Appendix D] The empirical evidence is narrow and not reported with uncertainty. The dataset uses only 10 MATH backbones and 20 variants each, with 4 runs per question, yet Figures 1–5 show no error bars or variance. The memory update policy stores only correct answers that exceed the score threshold, which introduces a selection bias toward successful cases and may overstate the speedup. No code or data is released. These limitations are not discussed in the main text; they appear only implicitly in Appendix D. To support the 'generally exists' claim, the authors need either substantially broader evaluation or explicit caveats about the scope.
minor comments (6)
  1. [Appendix B] The text contains a placeholder 'introduced in Section X' that should be replaced with the correct section number.
  2. [Section 3.2.1, Eq. (1)] The notation min over f_{s'}(q) subset of R_s is circular because f_{s'} is defined as the solution to the minimization; please clarify that the minimization is over stopping rules, not over the function mapping.
  3. [Section 4, metrics] Please specify how accuracy is computed (e.g., exact match, LLM judge) and whether the threshold τ is the same as the value_thresh reported in Appendix D (0.9).
  4. [Figure 1] The color scale for compute budget and accuracy changes is not defined; consider adding a color bar or numeric annotations for each cell.
  5. [Section 5.2, Finding 3] The text states the largest reductions are at S1 (16.0%) and S2 (15.4%), but Fig. 1 shows a 56% reduction at S3 for In-Context+DFS; clarify that the reported numbers are averages over all settings, while the 56% is a single cell.
  6. [Appendix D, Memory Update Policy] Please also state how many questions per backbone were used to update memory and whether the same memory state is carried across all four repetitions of a question; otherwise the variance across runs is difficult to interpret.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 2's proof assumes the tail-probability monotonicity it is meant to prove, so the 'theoretically guaranteed' speedup is not derived from first principles.

  1. other [Appendix A.2, proof of Theorem 2 (independent-sampling case)]
    "Let p_t := P(score(r^{(t)}_1) ≥ τ). Because memory grows monotonically with t, we have p_{t+1} ≥ p_t."

    Theorem 2 is stated as a derivation: from the premise that additional correct memory does not degrade the expected score, it must follow that P(max_{1≤i≤k} score ≥ τ) is non-decreasing in t. In the proof, the first step asserts p_{t+1} ≥ p_t, i.e., the per-candidate threshold-exceedance probability is non-decreasing, and justifies it only by 'memory grows monotonically.' That assertion is exactly the k=1 case of the theorem's conclusion; it is not a consequence of mean non-degradation, and memory growth alone does not imply it. The remainder of the independent proof (1-(1-p)^k) merely propagates this assumed p_t monotonicity to general k.

full rationale

The paper's empirical benchmarking is self-contained: it directly measures compute budgets and accuracy across memory methods, test-time scaling methods, and question-similarity levels, and the headline 56% reduction is an observed quantity, not a fitted parameter. There is no load-bearing self-citation: the references point to prior work by other research groups, and no uniqueness theorem is imported from the authors' own earlier papers. However, the theoretical 'guarantee' attached to SpeedupLLM is circular at its key step: Theorem 2's proof assumes the single-candidate threshold probability is non-decreasing with memory, which is exactly the conclusion to be proved. The generalized proof does not fix this; its mean-to-tail step is mathematically invalid, and its extra DAG and monotone-reliance assumptions are not part of the theorem's stated conditions. Consequently, Corollary 1 does not provide an independent first-principles derivation of reasoning speedup; modulo an invalid inequality, it restates the empirical premise that memory improves answer quality. The experimental results remain independent evidence, and the paper's own Finding 4 shows the premise can fail at low similarity, so the general claim is bounded by that condition. The score reflects partial circularity in the theoretical claim, not in the experimental measurements.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The framework introduces no new physical or architectural entities. The free parameters are hand-chosen thresholds and hyperparameters that control the early-stopping behavior, and the axioms are standard domain assumptions plus an ad hoc DAG condition used in the proof.

free parameters (4)
  • Score threshold tau = 0.9 (value_thresh)
    Hand-chosen early-stopping threshold in Best-of-N and DFS; determines compute budget reduction. Higher tau increases cost, lower tau risks accuracy loss.
  • DFS pruning ratio = 0.4
    Hand-chosen tree-search pruning ratio; affects compute and search depth.
  • Max in-context examples = 3
    Limits the size of the in-context memory buffer; chosen without ablation.
  • SFT learning rate = 5e-4
    Hand-chosen for single-step LoRA update; memory quality depends on this value.
assumptions (3)
  • domain assumption Additional correct memory from relevant questions does not degrade model performance
    Theorem 2's premise (B0) that expected score is non-decreasing with memory; this is the empirical claim the experiments test, not a derived result.
  • ad hoc to paper Candidate answers are independent or satisfy a monotone DAG dependence structure
    Appendix A.2 introduces a DAG over candidate indices to generalize the proof; no justification that real test-time scaling methods satisfy this structure.
  • domain assumption The score function reliably estimates answer quality so that early stopping at threshold tau preserves accuracy
    The framework assumes LLM judge or PRM scores align with correctness; the paper notes the score-correctness gap in Section 5.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Past Experience Accelerate LLM Reasoning?." pith.science (2026). https://pith.science/paper/2CWSQ43T

@misc{pith2026250520643,
  author       = {Pith},
  title        = {Pith review of: Can Past Experience Accelerate LLM Reasoning?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2CWSQ43T}},
  note         = {Machine review of arXiv:2505.20643}
}
read the original abstract

Allocating more compute to large language models (LLMs) reasoning has generally been demonstrated to improve their effectiveness, but also results in increased inference time. In contrast, humans can perform tasks faster and better with increased experience and exposure. Hence, this paper aims to investigate the question: Can LLMs also become faster at reasoning through recurrent exposure on relevant tasks, and if so, how can it be achieved? To address these questions, we first formalize the problem setting of LLM reasoning speedup systematically in the dimensions of task relevancy and compute budget calculation. We then propose SpeedupLLM, a theoretically guaranteed framework to implement and benchmark such reasoning speedup behaviour based on adaptive compute allocation and memory mechanisms. We further conduct comprehensive experiments to benchmark such behaviour across different question similarity levels, memory methods, and reasoning methods. Results show that LLMs can generally reason faster with past experience, achieving up to a 56% reduction in compute cost when equipped with appropriate memory and reasoning methods.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 2 canonical work pages

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774,

    arXiv

  2. [7]

    Near-duplicate question detection

    Preetam Prabhu Srikar Dammu and Omar Alonso. Near-duplicate question detection. InCompanion Proceedings of the ACM Web Conference 2024, pages 493–496,

    work page 2024

  3. [8]

    Editing factual knowledge in language models.arXiv preprint arXiv:2104.08164,

    Nicola De Cao, Wilker Aziz, and Ivan Titov. Editing factual knowledge in language models.arXiv preprint arXiv:2104.08164,

    arXiv

  4. [10]

    Alphaedit: Null-space constrained knowledge editing for language models.arXiv preprint arXiv:2410.02355,

    Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Shi Jie, Xiang Wang, Xiangnan He, and Tat-Seng Chua. Alphaedit: Null-space constrained knowledge editing for language models.arXiv preprint arXiv:2410.02355,

    arXiv

  5. [11]

    Alphazero-like tree-search can guide large language model decoding and training.arXiv preprint arXiv:2309.17179,

    Xidong Feng, Ziyu Wan, Muning Wen, Stephen Marcus McAleer, Ying Wen, Weinan Zhang, and Jun Wang. Alphazero-like tree-search can guide large language model decoding and training.arXiv preprint arXiv:2309.17179,

    arXiv

  6. [12]

    Critic: Large language models can self-correct with tool-interactive critiquing.arXiv preprint arXiv:2305.11738,

    Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Nan Duan, and Weizhu Chen. Critic: Large language models can self-correct with tool-interactive critiquing.arXiv preprint arXiv:2305.11738,

    arXiv

  7. [13]

    rstar-math: Small llms can master math reasoning with self-evolved deep thinking.arXiv preprint arXiv:2501.04519,

    Xinyu Guan, Li Lyna Zhang, Yifei Liu, Ning Shang, Youran Sun, Yi Zhu, Fan Yang, and Mao Yang. rstar-math: Small llms can master math reasoning with self-evolved deep thinking.arXiv preprint arXiv:2501.04519,

    arXiv

  8. [14]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shi- rong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    arXiv

Show all 47 references
  1. [15]

    Token- budget-aware llm reasoning.arXiv preprint arXiv:2412.18547,

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

    arXiv

  2. [16]

    Training large language models to reason in a continuous latent space.arXiv preprint arXiv:2412.06769,

    Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. Training large language models to reason in a continuous latent space.arXiv preprint arXiv:2412.06769,

    arXiv

  3. [17]

    Chatdb: Augmenting llms with databases as their symbolic memory.arXiv preprint arXiv:2306.03901,

    12 Can Past Experience Accelerate LLM Reasoning? Chenxu Hu, Jie Fu, Chenzhuang Du, Simian Luo, Junbo Zhao, and Hang Zhao. Chatdb: Augmenting llms with databases as their symbolic memory.arXiv preprint arXiv:2306.03901,

    arXiv

  4. [19]

    Openai o1 system card.arXiv preprint arXiv:2412.16720,

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Hel- yar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720,

    arXiv

  5. [20]

    Test-time computing: from system-1 thinking to system-2 thinking.arXiv preprint arXiv:2501.02497,

    Yixin Ji, Juntao Li, Hai Ye, Kaixin Wu, Jia Xu, Linjian Mo, and Min Zhang. Test-time computing: from system-1 thinking to system-2 thinking.arXiv preprint arXiv:2501.02497,

    arXiv

  6. [21]

    How well do llms compress their own chain-of-thought? a token complexity approach.arXiv preprint arXiv:2503.01141,

    Ayeong Lee, Ethan Che, and Tianyi Peng. How well do llms compress their own chain-of-thought? a token complexity approach.arXiv preprint arXiv:2503.01141,

    arXiv

  7. [22]

    How long can context length of open-source llms truly promise? InNeurIPS 2023 Workshop on Instruction Tuning and Instruction Following,

    Dacheng Li, Rulin Shao, Anze Xie, Ying Sheng, Lianmin Zheng, Joseph Gonzalez, Ion Stoica, Xuezhe Ma, and Hao Zhang. How long can context length of open-source llms truly promise? InNeurIPS 2023 Workshop on Instruction Tuning and Instruction Following,

    2023

  8. [23]

    Graphreader: Building graph-based agent to enhance long- context abilities of large language models.arXiv preprint arXiv:2406.14550,

    Shilong Li, Yancheng He, Hangyu Guo, Xingyuan Bu, Ge Bai, Jie Liu, Jiaheng Liu, Xingwei Qu, Yangguang Li, Wanli Ouyang, et al. Graphreader: Building graph-based agent to enhance long- context abilities of large language models.arXiv preprint arXiv:2406.14550,

    arXiv

  9. [24]

    From system 1 to system 2: A survey of reasoning large language models.arXiv preprint arXiv:2502.17419,

    Zhong-Zhi Li, Duzhen Zhang, Ming-Liang Zhang, Jiaxin Zhang, Zengyan Liu, Yuxuan Yao, Haotian Xu, Junhao Zheng, Pei-Jie Wang, Xiuyi Chen, et al. From system 1 to system 2: A survey of reasoning large language models.arXiv preprint arXiv:2502.17419,

    arXiv

  10. [25]

    Lost in the middle: How language models use long contexts.arXiv preprint arXiv:2307.03172,

    Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts.arXiv preprint arXiv:2307.03172,

    arXiv

  11. [26]

    Can 1b llm surpass 405b llm? rethinking compute-optimal test-time scaling.arXiv preprint arXiv:2502.06703,

    Runze Liu, Junqi Gao, Jian Zhao, Kaiyan Zhang, Xiu Li, Biqing Qi, Wanli Ouyang, and Bowen Zhou. Can 1b llm surpass 405b llm? rethinking compute-optimal test-time scaling.arXiv preprint arXiv:2502.06703,

    arXiv

  12. [28]

    An empirical study of catastrophic forgetting in large language models during continual fine-tuning.arXiv preprint arXiv:2308.08747,

    Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. An empirical study of catastrophic forgetting in large language models during continual fine-tuning.arXiv preprint arXiv:2308.08747,

    arXiv

  13. [29]

    Fast model editing at scale.arXiv preprint arXiv:2110.11309,

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. Fast model editing at scale.arXiv preprint arXiv:2110.11309,

    arXiv

  14. [30]

    s1: Simple test-time scaling

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

    arXiv

  15. [31]

    Self- training elicits concise reasoning in large language models.arXiv preprint arXiv:2502.20122,

    Tergel Munkhbat, Namgyu Ho, Seo Hyun Kim, Yongjin Yang, Yujin Kim, and Se-Young Yun. Self- training elicits concise reasoning in large language models.arXiv preprint arXiv:2502.20122,

    arXiv

  16. [32]

    Inference-time computations for llm reasoning and planning: A benchmark and insights

    Shubham Parashar, Blake Olson, Sambhav Khurana, Eric Li, Hongyi Ling, James Caverlee, and Shuiwang Ji. Inference-time computations for llm reasoning and planning: A benchmark and insights. arXiv preprint arXiv:2502.12521,

    arXiv

  17. [33]

    Self-reflection in llm agents: Effects on problem-solving perfor- mance

    Matthew Renze and Erhan Guven. Self-reflection in llm agents: Effects on problem-solving perfor- mance. arXiv preprint arXiv:2405.06682,

    arXiv

  18. [34]

    Character-llm: A trainable agent for role-playing

    Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. Character-llm: A trainable agent for role-playing. arXiv preprint arXiv:2310.10158,

    arXiv

  19. [35]

    Stop overthinking: A survey on efficient reasoning for large language models.arXiv preprint arXiv:2503.16419,

    14 Can Past Experience Accelerate LLM Reasoning? Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Hanjie Chen, et al. Stop overthinking: A survey on efficient reasoning for large language models.arXiv prepri...

    arXiv

  20. [36]

    Fast best-of-n decoding via speculative rejection.arXiv preprint arXiv:2410.20290,

    Hanshi Sun, Momin Haider, Ruiqi Zhang, Huitao Yang, Jiahao Qiu, Ming Yin, Mengdi Wang, Peter Bartlett, and Andrea Zanette. Fast best-of-n decoding via speculative rejection.arXiv preprint arXiv:2410.20290,

    arXiv

  21. [37]

    Tinyr1-32b-preview: Boosting accuracy with branch-merge distillation

    Lin Sun, Guangxiang Zhao, Xiaoqi Jian, Yuhan Wu, Weihong Lin, Yongfu Zhu, Linglin Zhang, Jinzhu Wu, Junfeng Ran, Sai-er Hu, et al. Tinyr1-32b-preview: Boosting accuracy with branch-merge distillation. arXiv preprint arXiv:2503.04872,

    arXiv

  22. [38]

    Online adaptation of language models with a memory of amortized contexts.arXiv preprint arXiv:2403.04317,

    Jihoon Tack, Jaehyung Kim, Eric Mitchell, Jinwoo Shin, Yee Whye Teh, and Jonathan Richard Schwarz. Online adaptation of language models with a memory of amortized contexts.arXiv preprint arXiv:2403.04317,

    arXiv

  23. [39]

    Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, et al. Kimi k1. 5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599,

    arXiv

  24. [40]

    Enhancing large language model with self-controlled memory framework.arXiv preprint arXiv:2304.13343, 2023a

    Bing Wang, Xinnian Liang, Jian Yang, Hui Huang, Shuangzhi Wu, Peihao Wu, Lu Lu, Zejun Ma, and Zhoujun Li. Enhancing large language model with self-controlled memory framework.arXiv preprint arXiv:2304.13343, 2023a. Haochun Wang, Chi Liu, Nuwa Xi, Zewen Qiang, Sendong Zhao, Bin...

    arXiv

  25. [41]

    Sampling-efficient test-time scaling: Self-estimating the best-of-n sampling in early decoding.arXiv preprint arXiv:2503.01422,

    Yiming Wang, Pei Zhang, Siyuan Huang, Baosong Yang, Zhuosheng Zhang, Fei Huang, and Rui Wang. Sampling-efficient test-time scaling: Self-estimating the best-of-n sampling in early decoding.arXiv preprint arXiv:2503.01422,

  26. [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, 2025a. Yige Xu, Xu Guo, Zhiwei Zeng, and Chunyan Miao. Softcot: Soft chain-of-thought for efficient reasoning with llms.arXiv preprint arXiv...

    arXiv

  27. [43]

    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,

    arXiv

  28. [44]

    Deeper insights without updates: The power of in-context learning over fine-tuning

    Qingyu Yin, Xuzheng He, Luoao Deng, Chak Tou Leong, Fan Wang, Yanzhao Yan, Xiaoyu Shen, and Qiang Zhang. Deeper insights without updates: The power of in-context learning over fine-tuning. arXiv preprint arXiv:2410.04691,

    arXiv

  29. [45]

    Ruihong Zeng, Jinyuan Fang, Siwei Liu, and Zaiqiao Meng

    InThe First Workshop on System-2 Reasoning at Scale, NeurIPS’24. Ruihong Zeng, Jinyuan Fang, Siwei Liu, and Zaiqiao Meng. On the structural memory of llm agents. arXiv preprint arXiv:2412.15266,

    arXiv

  30. [46]

    When scaling meets llm finetuning: The effect of data, model and finetuning method.arXiv preprint arXiv:2402.17193, 2024a

    Biao Zhang, Zhongtao Liu, Colin Cherry, and Orhan Firat. When scaling meets llm finetuning: The effect of data, model and finetuning method.arXiv preprint arXiv:2402.17193, 2024a. MengqiZhang, XiaotianYe, QiangLiu, PengjieRen, ShuWu, andZhuminChen. Uncoveringoverfitting in lar...

    arXiv

  31. [47]

    Judging llm-as-a-judge with mt-bench and chatbot arena

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595–46623, 2023a. LongtaoZhe...

    arXiv

  32. [1982]

    Arigraph: Learning knowledge graph world models with episodic memory for llm agents.arXiv preprint arXiv:2407.04363,

    Petr Anokhin, Nikita Semenov, Artyom Sorokin, Dmitry Evseev, Mikhail Burtsev, and Evgeny Burnaev. Arigraph: Learning knowledge graph world models with episodic memory for llm agents.arXiv preprint arXiv:2407.04363,

    arXiv

  33. [1988]

    O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning.arXiv preprint arXiv:2501.12570,

    Haotian Luo, Li Shen, Haiying He, Yibo Wang, Shiwei Liu, Wei Li, Naiqiang Tan, Xiaochun Cao, and Dacheng Tao. O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning.arXiv preprint arXiv:2501.12570,

    arXiv

  34. [2020]

    Towards reasoning era: A survey of long chain-of-thought for reasoning large language models.arXiv preprint arXiv:2503.09567,

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

    arXiv

  35. [2021]

    Dynamic parallel tree search for efficient llm reasoning

    Yifu Ding, Wentao Jiang, Shunyu Liu, Yongcheng Jing, Jinyang Guo, Yingjie Wang, Jing Zhang, Zengmao Wang, Ziwei Liu, Bo Du, et al. Dynamic parallel tree search for efficient llm reasoning. arXiv preprint arXiv:2502.16235,

    arXiv

  36. [2022]

    Recommender ai agent: Inte- grating large language models for interactive recommendations.arXiv preprint arXiv:2308.16505, 2023a

    Xu Huang, Jianxun Lian, Yuxuan Lei, Jing Yao, Defu Lian, and Xing Xie. Recommender ai agent: Inte- grating large language models for interactive recommendations.arXiv preprint arXiv:2308.16505, 2023a. Ziheng Huang, Sebastian Gutierrez, Hemanth Kamana, and Stephen MacNeil. Memo...

    arXiv

  37. [2023]

    L1: Controlling how long a reasoning model thinks with reinforcement learning.arXiv preprint arXiv:2503.04697,

    11 Can Past Experience Accelerate LLM Reasoning? Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning.arXiv preprint arXiv:2503.04697,

    arXiv

  38. [2024]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901,

    1901

  39. [2025]

    Compressed chain of thought: Efficient reasoning through dense representations.arXiv preprint arXiv:2412.13171,

    Jeffrey Cheng and Benjamin Van Durme. Compressed chain of thought: Efficient reasoning through dense representations.arXiv preprint arXiv:2412.13171,

    arXiv

Pith tools

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