Pith. sign in

REVIEW 4 major objections 6 minor 66 references

EPiC: Towards Lossless Speedup for Reasoning Training through Edge-Preserving CoT Condensation

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 Training on only the head and tail of chain-of-thought traces—discarding the middle exploration—cuts reasoning-training time by over 34% while preserving 90.2% accuracy on MATH500.

desk verdict EPiC's head/tail pruning is a genuinely new and practically useful data-condensation trick, but the 'lossless' claim holds only for the primary MATH500/Qwen configuration, not across models and benchmarks. read the letter →

arxiv 2506.04205 v1 pith:R33YRYEB submitted 2025-06-04 cs.LG

Jinghan Jia , Hadi Reisizadeh , Chongyu Fan , Nathalie Baracaldo , Mingyi Hong , Sijia Liu This is my paper
classification cs.LG
keywords chain-of-thoughtcondensationreasoningdistillationsupervisedfine-tuningtrainingefficiencythought-levelpruningmutualinformationmathematicaloverthinking
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 long chain-of-thought reasoning trace is not uniformly informative for supervised reasoning training: the early stage (problem understanding) and the final stage (solution convergence) carry most of the signal, while the middle exploration stage is largely redundant or noisy. It proposes EPiC, an edge-preserving thought-level condensation method that keeps only the first and last $\lfloor \tau n/2\rfloor$ thoughts of an $n$-thought trace and discards the middle, and claims this is lossless. On the main experiment, training a 7B math instruction model on a 93k-example distillation dataset at $\tau=0.5$ reaches 90.2% on MATH500, identical to full-trace training, while cutting training time by over 34%. The paper further reports comparable generation length and reflection behavior, better out-of-domain accuracy on a graduate-level STEM benchmark, and transfer across a second dataset and two additional base models. If right, the result matters because reasoning distillation is currently expensive, and EPiC makes it cheaper without asking the student to abandon long, reflective reasoning.

What carries the argument

The load-bearing mechanism is the edge-preserving condensation operator paired with a three-stage model of reasoning traces. EPiC assumes each trace has a functional arc—understanding at the head, exploration in the middle, convergence at the tail—and therefore selects $\Omega$ as the union of the first and last $\lfloor \tau n/2\rfloor$ thoughts. The argument is carried by a nearest-neighbor mutual-information estimator between hidden-state embeddings of condensed and full traces, which ranks EPiC above random, head-only, middle-only, and tail-only condensation at every tested ratio, and by a perturbation analysis showing that corrupting the middle content degrades accuracy least. These tools tie the simple positional pruning rule to the claim of losslessness.

What would settle it

Annotate a random sample of teacher traces from the main dataset, marking the first thought that contains the essential equation or substitution from which the final answer follows, and record its relative position; if a substantial fraction of solved examples have that decisive thought inside the middle 50% that EPiC removes, the head-tail split is not lossless for those examples and accuracy should drop when they are pruned.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the information needed to teach a small model to reason is concentrated at the edges of the trace. The paper defines a condensed trace as the head and tail index set $\Omega = \{1,\dots,\lfloor \tau n/2\rfloor\} \cup \{n - \lfloor \tau n/2\rfloor + 1,\dots,n\}$, with $\tau\in[0,1]$ the fraction of thoughts retained, and trains the student on those shortened traces. The support for calling this lossless is threefold: mutual information between condensed and full trace embeddings, where EPiC at $\tau=0.5$ scores 8.70 against 8.77 for the full trace; perturbation experiments in which randomizing the middle content costs the least accuracy (89.4% vs. 90.2% MATH500); and end-to-end training, where EPiC matches full-trace accuracy while cutting wall-clock time from 51.9 to 34.0 hours on the main dataset and also improves out-of-domain GPQA-Diamond accuracy (41.9% vs. 38.4%).

Load-bearing premise

The load-bearing premise is that a reasoning trace's position in the sequence tells you its job—head means understanding, middle means discardable exploration, tail means convergence—so the middle can be pruned uniformly without ever removing the decisive step.

Editorial extensions

If this is right

  • Training on EPiC-condensed traces at a 50% condensation ratio reaches the same MATH500 accuracy as full-trace training while cutting wall-clock training time by roughly one third, so reasoning distillation becomes affordable on fewer GPUs.
  • Models trained on condensed traces still emit long, reflective outputs with similar counts of reflection tokens, so the pruning does not collapse the student's visible reasoning behavior into terse answers.
  • The out-of-domain STEM accuracy gain over full-trace training suggests that removing middle exploration can act as a denoiser, not just a cost saver.
  • The same head-tail recipe transfers to a second reasoning dataset and to two additional base-model initializations, indicating the effect is tied to the structure of verbose reasoning traces rather than to one model or benchmark.

Reading between the lines

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

  • Beyond the paper's global condensation ratio, one could make the split per-example and difficulty-aware, retaining more of the trace where the decisive step sits late; the paper leaves this explicitly open.
  • The mutual-information score used for analysis could be turned into a cheap pre-training screening tool: estimate MI for candidate pruning policies and pick the highest scorer before spending GPU time.
  • If overthinking is concentrated in the middle of teacher traces, EPiC-style condensation could serve as a data-cleaning step in self-distillation loops, preventing verbose habits from being amplified into student models.
  • The positional head/middle/tail assumption could be tested on other trace formats, such as code-generation or tool-use reasoning, where exploration often contains the decisive API call rather than noise.
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 proposes EPiC, a thought-level CoT condensation method that retains the first and last floor(tau*n/2) thoughts of each trace and discards the middle, motivated by a claimed three-stage structure of reasoning (problem understanding, exploration, solution convergence). The authors fine-tune Qwen2.5-Math-7B-Instruct, Qwen2.5-7B-Instruct, and LLaMA3.1-8B-Instruct on OpenR1Math and GeneralThought-195k, and evaluate on MATH500, AIME24, and GPQA-Diamond. The headline empirical result is that at tau=0.5 on OpenR1Math, EPiC matches full-trace training on MATH500 (90.2 vs. 90.2) while reducing training time from 51.9 to 34.0 hours; the authors also report comparable MATH500 on GeneralThought (86.0 vs. 87.0) with time savings. Supportive analyses include a mutual-information comparison, a trajectory-landscape visualization, and a perturbation study. The paper concludes that middle thoughts are less informative and that EPiC achieves lossless reasoning supervision.

Significance. If the MATH500/Qwen2.5-Math/OpenR1Math result is taken as the claim, it is practically significant: a 34% training-time reduction with identical MATH500 accuracy on an open 7B model, with code released and no parameters tuned to the test benchmarks. The 'lossless' framing, however, is stronger than the evidence: AIME24 drops by 6.7 points on both OpenR1Math and GeneralThought, and LLaMA3.1-8B loses 3.2 points on MATH500. These are not resolved by the MI analysis, which is self-referential, and the paper's own Appendix E acknowledges that the positional segmentation may not align with semantic structure. The central contribution is therefore a benchmark-specific, empirically interesting finding, not a validated stage structure of reasoning.

major comments (4)
  1. [§5.2, Table 2] The 'lossless' claim is not supported by the AIME24 columns: EPiC drops from 46.7 to 40.0 on OpenR1Math and from 26.7 to 20.0 on GeneralThought, and Table 2 provides no repeated-seed or confidence-interval information that would show whether the MATH500 matches are stable. Please either restrict the lossless claim to MATH500 in the primary configuration, or provide an explanation such as per-problem analysis or variance estimates for the AIME24 degradation.
  2. [Table 3] The claim that EPiC is 'resilient to non-reasoning base model choice' is contradicted by the LLaMA3.1-8B result, where MATH500 accuracy falls from 78.2 to 75.0 while the full-trace model already saturates the 9000-token generation limit; this 3.2-point drop is larger than the MATH500 differences reported elsewhere. Please report multiple seeds and qualify the robustness conclusion, or identify a configuration-dependent condition under which losslessness holds.
  3. [§4, Table 1, Appendix B] The mutual-information analysis cannot establish that the pruned middle is least informative. Since E_Omega is constructed from a subset of the tokens used to form E_full, I(E_Omega; E_full) is expected to be high whenever the retained subset is large or contiguous, regardless of whether the discarded middle contains the decisive reasoning; it does not measure the training signal contributed by the middle thoughts. Please add a control such as equal-sized random token subsets or compute MI against the final answer or ground truth, or reframe the MI results as a descriptive redundancy statistic rather than evidence for the pruning decision.
  4. [Appendix E, §4 Eq. (EPiC)] The paper's own limitation statement says the head/middle/tail segmentation 'may not align with the true semantic structure of reasoning,' which is precisely the assumption underlying Eq. (EPiC) and the three-stage rationale in Section 4. This admission should be reflected in the abstract and conclusion; the current text presents the stage structure as validated by the MI and perturbation analyses, but those analyses use the same positional segmentation they are meant to justify. Please either provide an independent validation of the stage structure or explicitly frame EPiC as an empirical, position-based heuristic with benchmark-specific support.
minor comments (6)
  1. [Abstract and Conclusion] The abstract is precise in saying 'lossless reasoning accuracy on MATH500,' but the title and conclusion use 'lossless' without qualification; recommend aligning them with the benchmark-specific result.
  2. [§5.2] There is a typo 'Table,2', and the dataset name 'GeneralThought195k' is inconsistent with 'GeneralThoughts' used elsewhere in the paper.
  3. [Eq. (EPiC)] Equation (EPiC) does not specify behavior when floor(tau*n/2) is zero or when the head and tail intervals overlap for small n; this can matter for short reasoning traces.
  4. [Figure 1 and Figure 7] The numeric labels in Figure 1 (right) and Figure 7(d) are not explained clearly; please label whether the numbers are accuracy values, training hours, or speedup factors.
  5. [Appendix C.2] AIME24 pass@1 is computed with 32 samples per query, but Table 2 does not state this in the main text; please add this information for reproducibility.
  6. [§3] The sentence 'we first reviews CoT-based reasoning training' should be 'we first review', and the model name 'QWEN' should be written consistently as 'Qwen' throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EPiC's central lossless claim is evaluated on held-out benchmarks, and its heuristic head/middle/tail segmentation is an acknowledged assumption rather than a fitted prediction or self-citation chain.

full rationale

EPiC's core operation is a fixed positional pruning rule (Eq. EPiC), not a parameter fitted to the test benchmarks. The abstract's 'lossless' claim is explicitly scoped to MATH500, where Table 2 shows 90.2% for both full-data training and EPiC, and the comparison is externally evaluated on MATH500, AIME24, and GPQA-Diamond. No parameter is fitted to those benchmarks, and no fitted quantity is renamed as a prediction. The mutual information analysis (Table 1) and perturbation study (Figure 6) use the same head/middle/tail positional split that EPiC implements, so they are corroborative but not independent evidence for the segmentation; this is an evidentiary weakness, not a circular derivation. Appendix E explicitly concedes that the segmentation is heuristic and may not align with the true semantic structure of reasoning, and the failure to reproduce lossless accuracy on AIME24 (46.7 to 40.0) and for LLaMA3.1-8B on MATH500 (78.2 to 75.0) is an empirical limitation of the general 'lossless' wording, not a definitional reduction of the claim to its inputs. There is no self-citation chain, no uniqueness theorem imported from the authors, and no equation that reduces the reported accuracy to the method's definition by construction. Under the required standard of quoting a specific reduction or fitted-input-as-prediction, no circular step can be exhibited.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The method introduces no new entities or forces. It relies on a positional segmentation heuristic and a chosen condensation ratio. The MI analysis is auxiliary evidence, not a derivation.

free parameters (1)
  • condensation ratio tau = 0.5 default, swept from 0.01 to 1.0
    The fraction of thoughts retained is a hand-chosen hyperparameter central to the method. It is not fitted to the evaluation benchmarks, but the 'lossless' claim depends on the default value.
assumptions (3)
  • domain assumption CoT traces have a three-stage structure (problem understanding, exploration, solution convergence) aligned with head, middle, and tail positions.
    Section 4 and Figure 4 assert this mapping without independent validation that positional segments match functional stages.
  • domain assumption Thoughts can be reliably segmented by the '\n\n' delimiter, and each thought is an atomic unit for pruning.
    Section 3 adopts the delimiter from LightThinker [39]; if the delimiter splits mid-thought, pruning may break coherence.
  • domain assumption Mutual information between mean-pooled embeddings of a condensed trace and the full trace is a valid proxy for training informativeness.
    Section 4 uses the Kraskov estimator on a sample of 2500 examples; the validity of this proxy is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EPiC: Towards Lossless Speedup for Reasoning Training through Edge-Preserving CoT Condensation." pith.science (2026). https://pith.science/paper/R33YRYEB

@misc{pith2026250604205,
  author       = {Pith},
  title        = {Pith review of: EPiC: Towards Lossless Speedup for Reasoning Training through Edge-Preserving CoT Condensation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R33YRYEB}},
  note         = {Machine review of arXiv:2506.04205}
}
read the original abstract

Large language models (LLMs) have shown remarkable reasoning capabilities when trained with chain-of-thought (CoT) supervision. However, the long and verbose CoT traces, especially those distilled from large reasoning models (LRMs) such as DeepSeek-R1, significantly increase training costs during the distillation process, where a non-reasoning base model is taught to replicate the reasoning behavior of an LRM. In this work, we study the problem of CoT condensation for resource-efficient reasoning training, aimed at pruning intermediate reasoning steps (i.e., thoughts) in CoT traces, enabling supervised model training on length-reduced CoT data while preserving both answer accuracy and the model's ability to generate coherent reasoning. Our rationale is that CoT traces typically follow a three-stage structure: problem understanding, exploration, and solution convergence. Through empirical analysis, we find that retaining the structure of the reasoning trace, especially the early stage of problem understanding (rich in reflective cues) and the final stage of solution convergence, is sufficient to achieve lossless reasoning supervision. To this end, we propose an Edge-Preserving Condensation method, EPiC, which selectively retains only the initial and final segments of each CoT trace while discarding the middle portion. This design draws an analogy to preserving the "edge" of a reasoning trajectory, capturing both the initial problem framing and the final answer synthesis, to maintain logical continuity. Experiments across multiple model families (Qwen and LLaMA) and benchmarks show that EPiC reduces training time by over 34% while achieving lossless reasoning accuracy on MATH500, comparable to full CoT supervision. To the best of our knowledge, this is the first study to explore thought-level CoT condensation for efficient reasoning model distillation.

Figures

Figures reproduced from arXiv: 2506.04205 by the authors.

Figure 1
Figure 1. Overview of EPiC. Left: EPiC removes the middle portion of CoT while preserving the head (problem understanding) and tail (solution convergence). EPiC applies to training data in OpenR1Math. Right: Performance and training time comparison between EPiC and full CoT training based on QWEN2.5-MATH-7B-INSTRUCT. At 50% condensation ratio, EPiC achieves competitive accuracy with 1.5× faster training. supervision quality d… view at source ↗
Figure 2
Figure 2. Accuracy and training time for reasoning training on OpenR1Math (93k ex￾amples), LIMO (0.8k examples), and S1 (1k examples), using QWEN2.5-MATH-7B￾INSTRUCT as the base non-reasoning LLM. Accuracy is evaluated on the MATH500 bench￾mark. In addition to standard CoT datasets, we also include a thought-level condensed version of OpenR1Math, where 50% of the intermedi￾ate thoughts in each CoT trace are randomly retained … view at source ↗
Figure 3
Figure 3. Performance of TokenSkip-based token-level con￾densation for reasoning training. (Left) Visualization of a CoT trace pruned by TokenSkip [38] with a 50% pruning ratio. To￾kens highlighted in red are retained, while the rest are removed. (Right) Final answer accuracy of models trained on three datasets: TokenSkip-pruned (50%), random thought-level condensation (50%), and the original full dataset, evaluated on AIME a… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Illustration of three-stage struc￾ture of long CoT reasoning: problem under￾standing (head), exploration (middle), and solution convergence (tail). Edge-preserving condensation (EPiC). First, we segment each CoT trace into three distinct stages based on their functions…
Figure 5
Figure 5. Figure 5: Visualization of a reasoning trajectory generated by DEEPSEEK-R1-DISTILL-QWEN-7B on the AQuA [57] dataset. The plot is produced using the trajectory landscape tool from [58], where each node represents the model’s rea￾soning state in a latent space after k thought step…
Figure 6
Figure 6. Figure 6: Final answer accuracy comparison for reasoning training using QWEN2.5-MATH-7B￾INSTRUCT on various perturbed CoT training sets, evaluated on AIME and MATH500 at test time. Per￾turbations are applied to specific regions of the CoT trace–head, tail, middle, entire trace-o…
Figure 7
Figure 7. Figure 7: Reasoning accuracy of CoT training at different condensation ratios using EPiC, HoC, and ToC, on three benchmarks: (a) AIME24, (b) MATH500, and (c) GPQA-Diamond. All models are fine-tuned using Qwen2.5-Math-7B-Instruct on OpenR1Math. The dashed line indicates performan…
Figure 8
Figure 8. Figure 8: Accuracy and generation length across difficulty levels on the Math500 benchmark. (Top) Accuracy comparison of condensation methods (50% condensation ratio) and full-data baseline across five difficulty levels. (Bottom) Average number of generated tokens per method and…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 17 canonical work pages

  1. [1]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  2. [2]

    Let’s verify step by step

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. In The Twelfth International Conference on Learning Representations, 2023

    2023

  3. [5]

    Gpqa: A graduate-level google-proof q&a benchmark

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling, 2024

    2024

  4. [6]

    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, 2025

    arXiv 2025

  5. [7]

    Open r1: A fully open reproduction of deepseek-r1, January 2025

    Hugging Face. Open r1: A fully open reproduction of deepseek-r1, January 2025. URL https://github.com/ huggingface/open-r1

    2025

  6. [8]

    Open Thoughts

    OpenThoughts Team. Open Thoughts. https://open-thoughts.ai, January 2025

    2025

  7. [9]

    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, 2025

    arXiv 2025

  8. [10]

    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

    arXiv 2025

Show all 66 references
  1. [11]

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

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

    2024 arXiv

  2. [12]

    Thoughts are all over the place: On the underthinking of o1-like llms

    Yue Wang, Qiuzhi Liu, Jiahao Xu, Tian Liang, Xingyu Chen, Zhiwei He, Linfeng Song, Dian Yu, Juntao Li, Zhuosheng Zhang, et al. Thoughts are all over the place: On the underthinking of o1-like llms. arXiv preprint arXiv:2501.18585, 2025

    2025 arXiv

  3. [13]

    Teaching algorithmic reasoning via in-context learning

    Hattie Zhou, Azade Nova, Hugo Larochelle, Aaron Courville, Behnam Neyshabur, and Hanie Sedghi. Teaching algorithmic reasoning via in-context learning. arXiv preprint arXiv:2211.09066, 2022

    2022 arXiv

  4. [14]

    Distilling reasoning capabilities into smaller language models

    Kumar Shridhar, Alessandro Stolfo, and Mrinmaya Sachan. Distilling reasoning capabilities into smaller language models. Findings of the Association for Computational Linguistics: ACL 2023, pages 7059–7073, 2023

    2023

  5. [15]

    Specializing smaller language models towards multi-step reasoning

    Yao Fu, Hao Peng, Litu Ou, Ashish Sabharwal, and Tushar Khot. Specializing smaller language models towards multi-step reasoning. In International Conference on Machine Learning, pages 10421–10430. PMLR, 2023

    2023

  6. [16]

    Generalthought-195k

    GeneralReasoning. Generalthought-195k. https://huggingface.co/datasets/GeneralReasoning/ GeneralThought-195K, 2024. Accessed: 2025-05-06

    2024

  7. [17]

    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,

  8. [18]

    Bespoke-stratos: The unreasonable effectiveness of reasoning distillation

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

    2025

  9. [19]

    Re-distilling smaller deepseek r1 models for better performance, January 2025

    Appu Shaji Hicham Badri. Re-distilling smaller deepseek r1 models for better performance, January 2025. URL https://mobiusml.github.io/r1_redistill_blogpost/

    2025

  10. [20]

    Small models struggle to learn from strong reasoners

    Yuetai Li, Xiang Yue, Zhangchen Xu, Fengqing Jiang, Luyao Niu, Bill Yuchen Lin, Bhaskar Ramasubramanian, and Radha Poovendran. Small models struggle to learn from strong reasoners. arXiv preprint arXiv:2502.12143, 2025

    2025

  11. [21]

    Redstar: Does scaling long-cot data unlock better slow-reasoning systems? arXiv preprint arXiv:2501.11284, 2025

    Haotian Xu, Xing Wu, Weinong Wang, Zhongzhi Li, Da Zheng, Boyuan Chen, Yi Hu, Shijia Kang, Jiaming Ji, Yingying Zhang, et al. Redstar: Does scaling long-cot data unlock better slow-reasoning systems? arXiv preprint arXiv:2501.11284, 2025

    2025 arXiv

  12. [22]

    The first few tokens are all you need: An efficient and effective unsupervised prefix fine-tuning method for reasoning models

    Ke Ji, Jiahao Xu, Tian Liang, Qiuzhi Liu, Zhiwei He, Xingyu Chen, Xiaoyuan Liu, Zhijie Wang, Junying Chen, Benyou Wang, et al. The first few tokens are all you need: An efficient and effective unsupervised prefix fine-tuning method for reasoning models. arXiv preprint arXiv:25...

    2025 arXiv

  13. [23]

    Chain-of-thought prompting elicits reasoning in large language models, 2023

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. URL https: //arxiv.org/abs/2201.11903

    2023 arXiv

  14. [24]

    Inference scaling laws: An empirical analysis of compute-optimal inference for problem-solving with language models, 2024

    Yangzhen Wu, Zhiqing Sun, Shanda Li, Sean Welleck, and Yiming Yang. Inference scaling laws: An empirical analysis of compute-optimal inference for problem-solving with language models, 2024. URL https://arxiv. org/abs/2408.00724

    2024 arXiv

  15. [25]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei F...

    2025 arXiv

  16. [26]

    Scaling llm test-time compute optimally can be more effective than scaling model parameters, 2024

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute optimally can be more effective than scaling model parameters, 2024. URL https://arxiv.org/abs/2408.03314

    2024 arXiv

  17. [27]

    Self-consistency improves chain of thought reasoning in language models, 2023

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models, 2023. URL https: //arxiv.org/abs/2203.11171

    2023 arXiv

  18. [28]

    Let’s sample step by step: Adaptive-consistency for efficient reasoning and coding with llms

    Pranjal Aggarwal, Aman Madaan, Yiming Yang, et al. Let’s sample step by step: Adaptive-consistency for efficient reasoning and coding with llms. arXiv preprint arXiv:2305.11860, 2023

    2023 arXiv

  19. [29]

    Large language monkeys: Scaling inference compute with repeated sampling

    Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V Le, Christopher Ré, and Azalia Mirhoseini. Large language monkeys: Scaling inference compute with repeated sampling. arXiv preprint arXiv:2407.21787, 2024

    2024 arXiv

  20. [30]

    Griffiths, Yuan Cao, and Karthik Narasimhan

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models, 2023. URL https://arxiv.org/ abs/2305.10601

    2023 arXiv

  21. [31]

    Huajian Xin, Z. Z. Ren, Junxiao Song, Zhihong Shao, Wanjia Zhao, Haocheng Wang, Bo Liu, Liyue Zhang, Xuan Lu, Qiushi Du, Wenjun Gao, Qihao Zhu, Dejian Yang, Zhibin Gou, Z. F. Wu, Fuli Luo, and Chong Ruan. Deepseek-prover-v1.5: Harnessing proof assistant feedback for reinforcem...

    2024 arXiv

  22. [32]

    Generat- ing sequences by learning to self-correct

    Sean Welleck, Ximing Lu, Peter West, Faeze Brahman, Tianxiao Shen, Daniel Khashabi, and Yejin Choi. Generat- ing sequences by learning to self-correct. In The Eleventh International Conference on Learning Representations,

  23. [33]

    Self-refine: Iterative refinement with self-feedback, 2023

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. Self-refine: Iterativ...

    2023 arXiv

  24. [34]

    From decoding to meta-generation: Inference-time algorithms for large language models,

    Sean Welleck, Amanda Bertsch, Matthew Finlayson, Hailey Schoelkopf, Alex Xie, Graham Neubig, Ilia Kulikov, and Zaid Harchaoui. From decoding to meta-generation: Inference-time algorithms for large language models,

  25. [35]

    Can 1b llm surpass 405b llm? rethinking compute-optimal test-time scaling

    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, 2025

    2025 arXiv

  26. [36]

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

    Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning. arXiv preprint arXiv:2503.04697, 2025

    2025 arXiv

  27. [37]

    O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning

    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, 2025

    2025 arXiv

  28. [38]

    Tokenskip: Controllable chain-of-thought compression in llms

    Heming Xia, Yongqi Li, Chak Tou Leong, Wenjie Wang, and Wenjie Li. Tokenskip: Controllable chain-of-thought compression in llms. arXiv preprint arXiv:2502.12067, 2025

    2025

  29. [39]

    Lightthinker: Thinking step-by-step compression

    Jintian Zhang, Yuqi Zhu, Mengshu Sun, Yujie Luo, Shuofei Qiao, Lun Du, Da Zheng, Huajun Chen, and Ningyu Zhang. Lightthinker: Thinking step-by-step compression. arXiv preprint arXiv:2502.15589, 2025

    2025

  30. [40]

    Similar: Submodular information measures based active learning in realistic scenarios

    Suraj Kothawade, Nathan Beck, Krishnateja Killamsetty, and Rishabh Iyer. Similar: Submodular information measures based active learning in realistic scenarios. Advances in Neural Information Processing Systems, 34, 2021

    2021

  31. [41]

    Grad-match: A gradient matching based data subset selection for efficient learning

    Krishnateja Killamsetty, Durga Sivasubramanian, Baharan Mirzasoleiman, Ganesh Ramakrishnan, Abir De, and Rishabh Iyer. Grad-match: A gradient matching based data subset selection for efficient learning. arXiv preprint arXiv:2103.00123, 2021

    2021 arXiv

  32. [42]

    Deduplicating training data makes language models better

    Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison-Burch, and Nicholas Carlini. Deduplicating training data makes language models better. arXiv preprint arXiv:2107.06499, 2021

    2021 arXiv

  33. [43]

    Dataset pruning for resource-constrained spoofed audio detection

    Abdul Hameed Azeemi, Ihsan Ayyub Qazi, and Agha Ali Raza. Dataset pruning for resource-constrained spoofed audio detection. Proc. Interspeech 2022, pages 416–420, 2022

    2022

  34. [44]

    Contextual diversity for active learning

    Sharat Agarwal, Himanshu Arora, Saket Anand, and Chetan Arora. Contextual diversity for active learning. In ECCV, pages 137–153. Springer, 2020

    2020

  35. [45]

    Selection via proxy: Efficient data selection for deep learning

    Cody Coleman, Christopher Yeh, Stephen Mussmann, Baharan Mirzasoleiman, Peter Bailis, Percy Liang, Jure Leskovec, and Matei Zaharia. Selection via proxy: Efficient data selection for deep learning. arXiv preprint arXiv:1906.11829, 2019

    1906 arXiv

  36. [46]

    Active learning is a strong baseline for data subset selection

    Dongmin Park, Dimitris Papailiopoulos, and Kangwook Lee. Active learning is a strong baseline for data subset selection. In Has it Trained Yet? NeurIPS 2022 Workshop, 2022

    2022

  37. [47]

    Coresets for data-efficient training of machine learning models

    Baharan Mirzasoleiman, Jeff Bilmes, and Jure Leskovec. Coresets for data-efficient training of machine learning models. In ICML. PMLR, 2020

    2020

  38. [48]

    An empirical study of example forgetting during deep neural network learning

    Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geof- frey J Gordon. An empirical study of example forgetting during deep neural network learning. arXiv preprint arXiv:1812.05159, 2018

    2018 arXiv

  39. [49]

    Deep learning on a data diet: Finding important examples early in training

    Mansheej Paul, Surya Ganguli, and Gintare Karolina Dziugaite. Deep learning on a data diet: Finding important examples early in training. Advances in Neural Information Processing Systems, 34:20596–20607, 2021

    2021

  40. [50]

    Fewer is more: Boosting llm reasoning with reinforced context pruning

    Xijie Huang, Li Lyna Zhang, Kwang-Ting Cheng, Fan Yang, and Mao Yang. Fewer is more: Boosting llm reasoning with reinforced context pruning. arXiv preprint arXiv:2312.08901, 2023

    2023 arXiv

  41. [51]

    Staff: Specu- lative coreset selection for task-specific fine-tuning

    Xiaoyu Zhang, Juan Zhai, Shiqing Ma, Chao Shen, Tianlin Li, Weipeng Jiang, and Yang Liu. Staff: Specu- lative coreset selection for task-specific fine-tuning. In The Thirteenth International Conference on Learning Representations

  42. [52]

    Less: Selecting influential data for targeted instruction tuning

    Mengzhou Xia, Sadhika Malladi, Suchin Gururangan, Sanjeev Arora, and Danqi Chen. Less: Selecting influential data for targeted instruction tuning. arXiv preprint arXiv:2402.04333, 2024

    2024 arXiv

  43. [53]

    Lima: Less is more for alignment

    Chunting Zhou, Pengfei Liu, Puxin Xu, Srinivasan Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. Lima: Less is more for alignment. Advances in Neural Information Processing Systems, 36: 55006–55021, 2023

    2023

  44. [54]

    Openai o1 system card

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024

    2024 arXiv

  45. [55]

    Imitate, explore, and self-improve: A reproduction report on slow-thinking reasoning systems

    Yingqian Min, Zhipeng Chen, Jinhao Jiang, Jie Chen, Jia Deng, Yiwen Hu, Yiru Tang, Jiapeng Wang, Xiaoxue Cheng, Huatong Song, et al. Imitate, explore, and self-improve: A reproduction report on slow-thinking reasoning systems. arXiv preprint arXiv:2412.09413, 2024

    2024 arXiv

  46. [56]

    Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression

    Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Menglin Xia, Xufang Luo, Jue Zhang, Qingwei Lin, Victor Rühle, Yuqing Yang, Chin-Yew Lin, et al. Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression. arXiv preprint arXiv:2403.12968, 2024

    2024 arXiv

  47. [57]

    Program induction by rationale generation: Learning to solve and explain algebraic word problems

    Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blunsom. Program induction by rationale generation: Learning to solve and explain algebraic word problems. arXiv preprint arXiv:1705.04146, 2017

    2017 arXiv

  48. [58]

    Land- scape of thoughts: Visualizing the reasoning process of large language models

    Zhanke Zhou, Zhaocheng Zhu, Xuan Li, Mikhail Galkin, Xiao Feng, Sanmi Koyejo, Jian Tang, and Bo Han. Land- scape of thoughts: Visualizing the reasoning process of large language models. arXiv preprint arXiv:2503.22165, 2025

    2025

  49. [59]

    Estimating mutual information

    Alexander Kraskov, Harald Stögbauer, and Peter Grassberger. Estimating mutual information. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 69(6):066138, 2004

    2004

  50. [60]

    Pointer sentinel mixture models, 2016

    Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models, 2016

    2016

  51. [61]

    Math-verify: Math verification library, 2024

    Hynek Kydlíˇcek. Math-verify: Math verification library, 2024. URL https://github.com/huggingface/ math-verify. Apache-2.0 License

    2024

  52. [62]

    An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, et al. Qwen2. 5-math technical report: Toward mathematical expert model via self-improvement. arXiv preprint arXiv:2409.12122, 2024

    2024 arXiv

  53. [63]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024

    2024 arXiv

  54. [64]

    The llama 3 herd of models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  55. [65]

    Aime problems and solutions

    MAA Committees . Aime problems and solutions. https://artofproblemsolving.com/wiki/index.php/ AIME_Problems_and_Solutions. Appendix A Visualization of Condensed Reasoning Examples To provide qualitative insight into how EPiC condense long reasoning traces, Table A1 presents an...

  56. [2023]

    URL https://openreview.net/forum?id=hH36JeQZDaO

  57. [2024]

    URL https://arxiv.org/abs/2406.16838

    arXiv

  58. [2025]

    Accessed: 2025-01-09

    2025

Pith tools

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