Pith. sign in

REVIEW 6 major objections 5 minor 25 references

SCA: Segment-Wise CoT Compression with Answer Alignment

T0 review · 6 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A post-training method, DSS-GRPO, can shorten a reasoning model's thinking traces by roughly 40-60% while keeping final answers at the same length and accuracy as the untrained model, by routing compression rewards only to thinking tokens a

desk verdict The body of the paper describes a coherent, novel RL recipe for CoT compression, but the abstract describes a different method, and the 'same answers' claim rests on a length proxy that is never validated against actual answer content. read the letter →

arxiv 2603.07598 v2 pith:ZAAGN6NL submitted 2026-03-08 cs.AI cs.LG

Ye Tian , Hongyu Lin This is my paper
classification cs.AIcs.LG
keywords chain-of-thoughtcompressionsegment-wisereinforcementlearningGRPOanswer-lengthalignmentdifficulty-awarescalingmathreasoningtokenefficiencypost-trainingoptimization
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 is trying to establish that a reasoning model's hidden thinking trace can be compressed substantially after training without changing what the model ultimately says to the user. The proposed method, DSS-GRPO, treats the think segment and the answer segment of each completion as separate optimization targets, sends length-reduction pressure only to the think tokens, and anchors answer length to the original model. On several math benchmarks with two open-weight reasoning models, this yields roughly 40-60% shorter thinking traces while Pass@1 and answer length stay close to the base model. The contrast case is completion-level reinforcement learning, where a single length-reduction signal leaks into the answer and produces systematically shorter, less informative answers. The authors also report that the right amount of compression depends on difficulty and model capacity, so fixed length targets are brittle.

What carries the argument

The central object is DSS-GRPO (Difficulty-Scaled Segment-Wise GRPO), an RL update for post-training compression. It parses each completion into think and answer segments at boundary markers, computes separate group-relative advantages for each segment, and routes them with hard binary masks so compression rewards touch only think tokens and answer-alignment rewards only answer tokens. A quality gate limits structural rewards to correct, well-formed completions; a difficulty weight amplifies positive think advantages on hard prompts.

What would settle it

Sample a DSS-GRPO-compressed model and the base model on the same math problems, then have human raters or a strong evaluator judge whether same-length answers are equally complete, faithful, and helpful. If compressed-model answers systematically omit explanation steps, caveats, or parts of the solution while matching the base model's token count, the 'same answers' claim is falsified despite the tables in Section 4. A simpler automated check: compare answer content overlap (e.g., n-gram or entailment) between the two models on a held-out set.

Watch

Extended reading notes

Core claim

SCA/DSS-GRPO defines, for each sampled completion, binary masks over think and answer tokens, computes group-relative advantages separately for each segment, and multiplies the token-level policy-gradient objective by the routed advantages. The think advantage uses a gated, within-group min-max efficiency reward modulated by a difficulty weight; the answer advantage uses a length-alignment reward with a tolerance band around the frozen base model's answer length. The paper reports that this preserves base-model Pass@1 on five math benchmarks, shortens average think length by 40-60%, and keeps answer length near the reference, whereas a completion-level baseline with the same gating and think

Load-bearing premise

The load-bearing premise is that keeping the answer length aligned with the frozen base model's reference length is a sufficient proxy for preserving answer behavior and content; the paper measures length and Pass@1 but not semantic completeness or helpfulness.

Editorial extensions

If this is right

  • Post-trained models can emit roughly 40-60% shorter thinking traces on math benchmarks while keeping Pass@1 and answer length near the base model, cutting inference cost per query.
  • Completion-level GRPO under length pressure produces answer drift (shorter, terser answers); isolating the gradient by segment prevents this.
  • Compression pressure should scale with prompt difficulty and model competence; a fixed target is brittle because harder problems retain longer necessary reasoning.
  • LoRA-only post-training on one dataset does not reliably transfer compression to harder out-of-domain benchmarks; full-parameter training does.
  • The answer-length alignment reward permits slight over-length answers but mainly prevents under-length ones, so user-facing answers do not degrade into terse responses.

Reading between the lines

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

  • Because answer preservation is checked only via token length and Pass@1, the strong reading of 'same answers' — identical useful content — is untested. A content-level evaluation (e.g., human judged completeness or semantic similarity) would either validate or bound the claim.
  • The segment-routing trick is generic: any structured output with separable roles (e.g., tool calls vs. user-facing text, or executable code vs. comments) could use the same mask-based advantage routing to compress one part without distorting the other.
  • The difficulty-scaling mechanism is a testable recipe for adaptive inference: models could expose a competence estimate per prompt and decide how aggressively to compress at serving time, not just during training.
  • Reported gains are on math reasoning; the method's value outside math (e.g., code, legal, or multilingual reasoning) remains an open empirical question.
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

6 major / 5 minor

Summary. The paper proposes DSS-GRPO, a segment-wise GRPO variant for post-training reasoning LLMs to compress the 'think' segment of structured chain-of-thought outputs while preserving the 'answer' segment. It partitions each completion into think and answer tokens via hard masks, computes separate group-relative advantages for each segment, applies difficulty scaling only to the think advantage, and adds an answer-length alignment reward anchored to the frozen base model's reference length L_ref(x). Experiments on Qwen3-4B and Qwen3-8B over five math benchmarks report Pass@1 and average think/answer lengths, claiming roughly 40-60% think compression with stable answer lengths, together with a GSM8K LoRA case study.

Significance. If fully established, DSS-GRPO would be a practically valuable method for token-efficient reasoning: the reported think-length reductions are large and consistent, and the answer-length stability relative to a completion-level baseline is evident. The idea of segment-routed advantages for think/answer credit assignment is novel and worth pursuing. However, the paper does not currently substantiate the 'same answers' part of the central claim, and it omits any comparison to existing CoT-compression methods, so the stated significance as a state-of-the-art compression method is not yet supported.

major comments (6)
  1. [§3.5, Eq. (11), Table 2] The central claim 'preserving answer behavior' is not established. The only answer-side objective is a length-based proxy anchored to L_ref(x), and the evaluation reports only average answer length and Pass@1. Pass@1 is a binary correctness metric that cannot detect answers that are correct but less complete, less helpful, or substantively different; the reward plateau even permits over-length answers. To support 'same answers', the paper must either add a semantic/content-preservation evaluation (e.g., human or LLM judgments of answer informativeness and equivalence) or explicitly limit the claim to 'answer-length preservation.' This is load-bearing for the conclusion.
  2. [§4.2, Tables 1–2] No comparison is made to any prior CoT-compression method, although several are discussed in §2 (CoT-Valve, DeepCompress, TokenSkip, step entropy, etc.). The abstract claims 'state-of-the-art-level chain-of-thought compression,' but without these baselines the SOTA claim is unsupported. At minimum, report results for the most relevant prior methods on the same benchmarks and metrics, with matched training budgets and decoding settings.
  3. [Abstract vs. full text] The arXiv abstract describes a method named 'SCA' that 'protects answer tokens through length and distribution alignment,' while the full text presents a different title and a method (DSS-GRPO) that uses length-only alignment (Eq. 11) and explicitly rejects distribution matching. This is a serious inconsistency that misrepresents the contribution and must be resolved (whether by aligning the abstract to the full text or by clarifying which method is being submitted).
  4. [Table 1] Table 1 reports no error bars, confidence intervals, or significance tests. Many accuracy differences between Base and DSS-GRPO are small (e.g., Qwen3-4B MATH-500: 97.2 vs. 97.2; Qwen3-8B AMC23: 95.0 vs. 97.5). Without variance estimates, the claim that DSS-GRPO 'preserves' accuracy is not statistically supported. Provide standard errors or paired tests across problems or repeated sampling seeds.
  5. [§4.2, ablations] The only baseline, 'Naive GRPO,' removes both segment routing and difficulty scaling simultaneously, so it cannot isolate the contribution of the difficulty-scaling mechanism (Eqs. 7–8, 14) or of the answer-length alignment reward. The paper presents difficulty scaling as a key contribution but gives no ablation that varies Wdiff while holding routed advantages fixed, nor an ablation that removes only the answer-alignment term. A proper ablation matrix is needed to support the component claims.
  6. [§4.1, evaluation] The Pass@1 metric is ambiguous. It is unclear whether it is the mean over problems of the fraction of correct samples among N=8, or the fraction of problems with at least one correct sample. These can differ substantially, especially at high accuracies. Please define the estimator precisely and cite the convention used.
minor comments (5)
  1. [§4.3, Figure 1 caption] The caption reads 'DDS-GRPO' but should be 'DSS-GRPO'.
  2. [§3.5] L_ref(x) is not defined operationally: is it the answer length from a single base-model sample, an average over multiple samples, or the gold answer length? Since the reward depends on L_ref(x), this must be specified for reproducibility.
  3. [Table 3] The training configuration says 'batch=1 prompt/step (per-device), grad accum=1 (global=1)' — clarify whether the global batch size is 1 or the product with the number of devices.
  4. [Table 1] 'Avg Acc' is reported without defining the weighting across benchmarks. State whether it is a simple average or weighted by dataset size.
  5. [§A.1] The text says 'guaranteeing at least a 2× amplification when successes are extremely scarce' but with the global scale s=1.5 the effective multiplier is 3× on positive advantages. The wording is imprecise.

Circularity Check

2 steps flagged · score 6.0 of 10

Answer-length preservation is optimized into the loss and then reported as evidence of 'same answers'; semantic equivalence is never tested.

  1. fitted input called prediction [Section 3.5 (Eq. 11); Algorithm 1 step 19; Section 4.2 (Table 2); Section 5]
    "we add an explicit length-alignment reward that anchors answer length to a reference behavior. Although one could match answer-token distributions to a reference model with a KL term, a length-based proxy is cheaper and effective ... it directly targets the dominant drift (answers getting too short). ... DSS-GRPO mitigates this collapse by explicitly isolating segment updates and anchoring answer length, maintaining answer-length behavior while achieving comparable think compression."

    Eq. (11) defines R_len so that R_len is maximized when L_ans lies in [L_ref, L_ref+f]; Algorithm 1 computes R_len for every completion; Eqs. (13)-(15) convert it into the answer-token advantage; Eq. (16) is the training loss. Table 2 then reports E[L_ans] close to L_ref as evidence that DSS-GRPO 'maintain[s] answer-length behavior,' and the conclusion claims it 'preserv[es] answer behavior.' The reported length match is the optimized objective itself, not an independent test of answer preservation. No semantic comparison of answer content is provided.

  2. self definitional [Section 3.5; Section 5 Conclusion]
    "The plateau [Lref, Lref + f] allows slightly longer, more user-friendly answers, while mainly counteracting under-length answers ... DSS-GRPO shortens think while preserving answer behavior, avoiding the answer-shortening drift of completion-level GRPO."

    The paper's only operationalization of 'answer behavior' for optimization is the length band in Eq. (11). The conclusion elevates this to the claim that DSS-GRPO 'preserv[es] answer behavior,' but answer behavior is never measured except by token length and Pass@1. Since Pass@1 is a correctness metric, not a content-preservation metric, the claim 'same answers' is the training target under a new name; the paper explicitly chose a length proxy over distribution matching without validating that the proxy correlates with content preservation.

full rationale

The think-compression and accuracy results are not circular: Reff is a within-group length-shaping reward gated by format/correctness, and the reported Pass@1 and think-length reductions are evaluated on held-out benchmarks against an independent Naive GRPO baseline. No load-bearing self-citation appears; the cited GRPO/CoT works are external and do not define the method. The circularity is localized to the 'same answers' / 'answer behavior' claim. Eq. (11) makes L_ref(x) the target of an explicit answer-length reward; Algorithm 1 computes R_len; Eqs. (13)-(15) route the mean-centered R_len advantage to answer tokens; Eq. (16) is the policy-gradient loss. Thus, E[L_ans] close to L_ref in Table 2 is a check that the optimization worked, not an independent test that answer behavior is preserved. The paper explicitly chooses a length proxy over distribution matching for cost reasons but never validates the proxy against answer content or completeness; Pass@1 is a binary correctness metric and cannot detect answers that are correct but less complete or less helpful. Therefore, the conclusion 'preserving answer behavior' is, in the evidence presented, equivalent to the training objective. This is partial circularity: the think-length compression and accuracy results are genuinely independent, but the answer-preservation half of the central claim reduces to the fitted target. Score 6.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim depends on hand-set hyperparameters (m=256, f=32, s=1.5, K=16, temperature schedule), a clean think/answer parser, and the assumption that length alignment to L_ref preserves answer content. No sensitivity analysis or content-similarity metric is provided.

free parameters (6)
  • think margin m = 256
    Length plateau in Eq. 10; chosen by hand, no sensitivity analysis.
  • answer tolerance band f = 32
    Over-length tolerance in Eq. 11; chosen by hand.
  • difficulty scale s = 1.5
    Multiplies positive think advantages in Eq. 14; hand-set, no ablation.
  • group size K = 16
    Number of completions per prompt for GRPO; chosen by hand.
  • temperature schedule = cosine 1.3 -> 0.7
    Annealing schedule used in training; chosen by hand, no systematic comparison except a LoRA pilot.
  • reference answer length L_ref(x) = not reported; from frozen base model rollout(s)
    Target for answer alignment in Eq. 11; the number of samples used to estimate it and the decoding setting are unspecified.
assumptions (5)
  • domain assumption Each completion can be deterministically partitioned into think and answer via boundary markers (</think>, <|im_end|>).
    Section 3.1: masks depend on clean parsing; malformed outputs would route rewards incorrectly.
  • domain assumption A reliable correctness signal I_corr(y) is available for every prompt.
    Section 3.2: the quality gate g(k) multiplies all rewards; a noisy verifier would break the gating.
  • domain assumption Answer length alignment to a frozen base model preserves answer quality.
    Section 3.5: length is used as a proxy for content; the paper does not measure answer semantics.
  • ad hoc to paper Group success rate p_succ(x) is a valid competence proxy for difficulty scaling.
    Eqs. 7-8: W_diff = 2 - p_succ is introduced without derivation or validation.
  • standard math GRPO group-relative advantages and the policy-gradient loss in Eq. 16 are applicable to this setting.
    Borrowed from GRPO literature (ref [24]); assumed stable under segment masks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SCA: Segment-Wise CoT Compression with Answer Alignment." pith.science (2026). https://pith.science/paper/ZAAGN6NL

@misc{pith2026260307598,
  author       = {Pith},
  title        = {Pith review of: SCA: Segment-Wise CoT Compression with Answer Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZAAGN6NL}},
  note         = {Machine review of arXiv:2603.07598}
}
read the original abstract

Chain-of-thought (CoT) reasoning improves problem solving, but long think traces increase inference cost. Existing CoT compression methods usually optimize completion-level length. For structured thinking models, however, a completion contains both a think segment and an answer segment, so completion-level compression can save tokens by compressing not only the CoT but also the answer. We call this failure mode answer drift. We propose Segment-wise CoT Compression with Answer Alignment (SCA), an answer-preserving think-compression method. SCA parses completions into functional segments, routes compression rewards only to successful think tokens, and protects answer tokens through length and distribution alignment to a frozen base model. Experiments show that, across datasets from multiple domains, SCA achieves state-of-the-art-level chain-of-thought compression while preserving the base model's performance and answer alignment. Training data and code are included in the supplementary code and data package.

Figures

Figures reproduced from arXiv: 2603.07598 by the authors.

Figure 1
Figure 1. GSM8K-test length distributions at evaluation [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. GSM8K Example: Base vs DSS-GRPO. Our results also highlight that the shortest sufficient reasoning length depends on difficulty and capacity: harder benchmarks retain longer traces, and LoRA-only GSM8K training does not reliably transfer compression to harder out-of-domain tests, whereas full-parameter post-training is more effective. Future work includes extending routing to finer-grained structure and broader task… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 1 linked inside Pith

  1. [1]

    Skip-thinking: Chunk-wise chain-of-thought distillation enable smaller language models to reason better and faster

    Xiaoshu Chen, Sihang Zhou, Ke Liang, Xiaoyu Sun, and Xinwang Liu. Skip-thinking: Chunk-wise chain-of-thought distillation enable smaller language models to reason better and faster. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 12153–12168, 2025

    2025

  2. [2]

    Navigate through enigmatic labyrinth a survey of chain of thought reasoning: Advances, frontiers and future

    Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Tao He, Haotian Wang, Weihua Peng, Ming Liu, Bing Qin, and Ting Liu. Navigate through enigmatic labyrinth a survey of chain of thought reasoning: Advances, frontiers and future. InAnnual Meeting of the Association for Computational Linguistics, 2023

    2023

  3. [3]

    Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    arXiv 2021

  4. [4]

    Bryan L. M. de Oliveira, Felipe Vieira Frujeri, Marcos P. C. M. Queiroz, Luana G. B. Martins, Telma W. de L. Soares, and Luckeciano Carvalho Melo. Learning without critics? revisiting grpo in classical reinforcement learning environments.ArXiv, abs/2511.03527, 2025

    arXiv 2025

  5. [5]

    Sutherland, Xiaoxiao Li, and Christos Thrampoulidis

    Wenlong Deng, Yi Ren, Muchen Li, Danica J. Sutherland, Xiaoxiao Li, and Christos Thrampoulidis. On the effect of negative gradient in group relative deep reinforcement optimization. In2nd AI for Math Workshop @ ICML 2025, 2025

    2025

  6. [6]

    Cot-evo: Evolutionary distillation of chain-of-thought for scientific reasoning

    Kehua Feng, Keyan Ding, Zhihui Zhu, Lei Liang, Qiang Zhang, and Huajun Chen. Cot-evo: Evolutionary distillation of chain-of-thought for scientific reasoning. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  7. [7]

    Solving quantitative reasoning problems with language models

    Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, Yuhuai Wu, Behnam Neyshabur, Guy Gur-Ari, and Vedant Misra. Solving quantitative reasoning problems with language models. InAdvances in Neural Information Processing Systems, 2022

    2022

  8. [8]

    SSVPO: Effective step-level credit assignment for RL training of language models

    Yugu Li, Zehong Cao, Jianglin Qiao, and Siyi Hu. SSVPO: Effective step-level credit assignment for RL training of language models. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

Show all 25 references
  1. [9]

    Making slow thinking faster: Compressing LLM chain-of-thought via step entropy

    Zeju Li, Jianyuan Zhong, Ziyang Zheng, Xiangyu Wen, Zhijian Xu, Yingying Cheng, Fan Zhang, and Qiang Xu. Making slow thinking faster: Compressing LLM chain-of-thought via step entropy. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  2. [10]

    Chain of thought empowers transformers to solve inherently serial problems

    Zhiyuan Li, Hong Liu, Denny Zhou, and Tengyu Ma. Chain of thought empowers transformers to solve inherently serial problems. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  3. [11]

    Deepcompress: A dual reward strategy for dynamically exploring and compressing reasoning chains

    Tian Liang, Wenxiang Jiao, Zhiwei He, Jiahao Xu, Haitao Mi, and Dong Yu. Deepcompress: A dual reward strategy for dynamically exploring and compressing reasoning chains. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  4. [12]

    Wu, Ilia Sucholutsky, Tania Lombrozo, and Thomas L

    Ryan Liu, Jiayi Geng, Addison J. Wu, Ilia Sucholutsky, Tania Lombrozo, and Thomas L. Griffiths. Mind your step (by step): Chain-of-thought can reduce performance on tasks where thinking makes humans worse, 2025

    2025

  5. [13]

    Can language models learn to skip steps? InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    Tengxiao Liu, Qipeng Guo, Xiangkun Hu, Cheng Jiayang, Yue Zhang, Xipeng Qiu, and Zheng Zhang. Can language models learn to skip steps? InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  6. [14]

    Cot-valve: Length- compressible chain-of-thought tuning

    Xinyin Ma, Guangnian Wan, Runpeng Yu, Gongfan Fang, and Xinchao Wang. Cot-valve: Length- compressible chain-of-thought tuning. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6025–6035, 2025

    2025

  7. [15]

    American invitational mathematics examination (aime)

    MAA. American invitational mathematics examination (aime). Mathematics Competition Series. n.d.a. URL https://maa.org/math-competitions/aime

  8. [16]

    American mathematics competitions (amc 12)

    MAA. American mathematics competitions (amc 12). Mathematics Competition Series. n.d.b. URL https://maa.org/math-competitions/amc

  9. [17]

    Codi: Compressing chain-of-thought into continuous space via self-distillation

    Zhenyi Shen, Hanqi Yan, Linhai Zhang, Zhanghao Hu, Yali Du, and Yulan He. Codi: Compressing chain-of-thought into continuous space via self-distillation. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 677–693, 2025. 10

    2025

  10. [18]

    Think silently, think fast: Dynamic latent compression of llm reasoning chains.ArXiv, abs/2505.16552, 2025

    Wenhui Tan, Jiaze Li, Jianzhong Ju, Zhenbo Luo, Jian Luan, and Ruihua Song. Think silently, think fast: Dynamic latent compression of llm reasoning chains.ArXiv, abs/2505.16552, 2025

    2025

  11. [19]

    Understanding chain-of-thought in LLMs through information theory

    Jean-Francois Ton, Muhammad Faaiz Taufiq, and Yang Liu. Understanding chain-of-thought in LLMs through information theory. InForty-second International Conference on Machine Learning, 2025

    2025

  12. [20]

    Polymath: Evaluating mathematical reasoning in multilingual contexts.arXiv preprint arXiv:2504.18428, 2025

    Yiming Wang, Pei Zhang, Jialong Tang, Haoran Wei, Baosong Yang, Rui Wang, Chenshu Sun, Feitong Sun, Jiran Zhang, Junxuan Wu, Qiqian Cang, Yichang Zhang, Fei Huang, Junyang Lin, Fei Huang, and Jingren Zhou. Polymath: Evaluating mathematical reasoning in multilingual contexts.ar...

    2025

  13. [21]

    SIM-cot: Supervised implicit chain-of-thought

    Xilin Wei, Xiaoran Liu, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Jiaqi Wang, Xipeng Qiu, and Dahua Lin. SIM-cot: Supervised implicit chain-of-thought. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  14. [22]

    TokenSkip: Controllable chain- of-thought compression in LLMs

    Heming Xia, Chak Tou Leong, Wenjie Wang, Yongqi Li, and Wenjie Li. TokenSkip: Controllable chain- of-thought compression in LLMs. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng, editors,Proceedings of the 2025 Conference on Empirical Methods ...

    2025

  15. [23]

    Qwen3 technical report, 2025

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

    2025

  16. [24]

    Group-relative REINFORCE is secretly an off-policy algorithm: Demystifying some myths about GRPO and its friends

    Chaorui Yao, Yanxi Chen, Yuchang Sun, Yushuo Chen, Wenhao Zhang, Xuchen Pan, Yaliang Li, and Bolin Ding. Group-relative REINFORCE is secretly an off-policy algorithm: Demystifying some myths about GRPO and its friends. InThe Fourteenth International Conference on Learning Repr...

    2026

  17. [25]

    what works

    Xianwei Zhuang, Zhihong Zhu, Zhichang Wang, Xuxin Cheng, and Yuexian Zou. UnicoTT: A unified framework for structural chain-of-thought distillation. InThe Thirteenth International Conference on Learning Representations, 2025. A Technical Appendices and Supplementary Material A...

    2025

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.