Pith. sign in

REVIEW 3 major objections 5 minor 38 references

A lightweight correctness-filtered SFT bootstrap fixes the cold-start bottleneck in on-policy self-distillation, making shorter reasoning traces more accurate than longer ones.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 22:26 UTC pith:2DX7V5CQ

load-bearing objection A competent two-stage recipe for compressed reasoning, but the central prefix-support mechanism is asserted, not measured. the 3 major comments →

arxiv 2607.15736 v1 pith:2DX7V5CQ submitted 2026-07-17 cs.CL

Better Starts, Better Ends: Bootstrapped Iterative Self-Reasoning Distillation for Compressed Reasoning

Leichao Dong , Dongxu Zhang , Yiding Sun , Qirui Wang , Yuhan Wang , Lin Chen , Jihua Zhu This is my paper
classification cs.CL
keywords large language modelschain-of-thoughton-policy self-distillationreasoning compressionprefix supporttoken efficiencyrejection samplingknowledge distillation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that on-policy self-distillation for compressing reasoning fails from a cold start because its loss is prefix-local: the concise teacher is only queried on trajectories the student already generates, and a verbose base model generates noisy, redundant, or off-track prefixes where the teacher can only make local corrections. BIRD fixes the start rather than the teacher: it samples concise solutions from the base model under a brevity instruction, keeps answer-correct ones, and fine-tunes them under the original prompt so conciseness becomes default behavior. Then it runs the same on-policy reverse-KL distillation from the warmed model, so the loss lands on cleaner prefixes. If this is right, the practical upshot is large: on Qwen3-8B MATH-500 accuracy goes from 86.2% to 92.0% while average output length falls from 3,099 to 1,115 tokens, and the gains transfer across model families.

Core claim

BIRD's central claim is that a concise self-teacher's value is determined by where it is queried: in on-policy distillation the student only receives KL supervision on prefixes it actually visits, and a verbose base model visits noisy, redundant, or off-track prefixes. By first performing a lightweight SFT bootstrap on self-generated, correctness-filtered concise traces learned under the original prompt, BIRD shifts the student onto shorter, on-track rollouts; the same reverse-KL objective then produces sharper, more actionable targets. The paper reports that on Qwen3-8B this raises MATH-500 accuracy from 86.2% to 92.0% while cutting average response length from 3,099 to 1,115 tokens, and im

What carries the argument

The load-bearing mechanism is prefix-local reverse-KL self-distillation: the student samples a rollout, and at each prefix the loss compares the student's next-token distribution with that of a stop-gradient concise teacher conditioned on a brevity instruction. Because supervision is applied only on visited prefixes, the paper's contribution is to change which prefixes are visited before distillation begins. This is done by a prompt-switch SFT bootstrap—traces generated under a brevity instruction but learned under the original prompt—so conciseness becomes a default behavior rather than an inference-time artifact. The bootstrap is deliberately lightweight: one rollout per problem, a 2,048-t

Load-bearing premise

The load-bearing premise is that the correctness-filtered concise self-traces used to warm start actually place the student on more informative prefixes, and that filtering on final-answer correctness does not simply select easy problems the base model can already answer concisely.

What would settle it

Stratify test problems by whether the base model solves them in short, correct traces; if BIRD's accuracy-efficiency gains appear only on that easy stratum, the bootstrap is doing dataset selection rather than prefix-support repair.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Cold-start OPSD plateaus at a lower token-efficiency frontier; the paper shows that longer cold-start training does not close the gap, so initialization matters more than extra optimization budget.
  • A single lightweight SFT bootstrap on correctness-filtered concise traces is enough to make the same KL objective reach 11.92 TE within 25 steps, already above CRISP's peak.
  • Applying the bootstrap after OPSD hurts accuracy—AIME 2025 falls from 40.83% to 28.33%—so the warm-start must come before on-policy distillation.
  • The recipe transfers across model families: DeepSeek-R1-Distill-Llama-8B gains 8.0 MATH-500 points while reducing length by 33%, even though its baseline traces are already compact.
  • Gains widen with scale: at 1.7B BIRD mostly compresses, while at 8B it improves accuracy and length jointly on all three benchmarks.

Where Pith is reading between the lines

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

  • A testable consequence the paper does not pursue: if prefix support is the mechanism, then any cheap way to bias the student's rollouts toward short, correct traces—rejection sampling, a length-penalized RL warm-up, or a difficulty-ranked curriculum—should produce similar accuracy-efficiency gains; BIRD's SFT is one implementation, not the only one.
  • The prompt-switch trick suggests that inference-time instructions can be compiled into model weights; an extension would be to measure how much warm-start data is needed to make conciseness persist without the instruction, and whether this works for non-math domains like code generation or long-document reasoning.
  • Because the gains grow with model scale, a plausible but implicit consequence is that the bottleneck is more severe for larger models that have more verbose base rollouts; comparing BIRD against a larger cold-start baseline on the same compute budget would help quantify how much of the benefit is scale-dependent.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes BIRD, a two-stage method for compressing chain-of-thought traces. Stage 1 ("static bootstrap") samples one concise solution per training problem from the base model under a brevity instruction, keeps only traces with the correct answer, and runs one epoch of LoRA SFT under the original task prompt (prompt switch). Stage 2 ("dynamic refinement") runs on-policy reverse-KL distillation (Eq. 3) from this warm-started model, with a concise self-teacher that is a periodically refreshed stop-gradient copy. Experiments on Qwen3-1.7B/4B/8B and DeepSeek-R1-Distill-Llama-8B on MATH-500, AIME 2024, and AIME 2025 report that BIRD improves Token Efficiency (TE = accuracy/ln(length), Eq. 5) over prompting and cold-start OPSD in 8/9 model-benchmark pairs; the largest gain is Qwen3-8B on MATH-500: accuracy 86.2→92.0 with length 3,099→1,115. The paper interprets these gains as evidence that warm-starting shifts the student's prefix support to "cleaner" contexts, making the same prefix-local KL objective more effective.

Significance. If the empirical results are reproducible, BIRD is a valuable practical recipe: a lightweight SFT bootstrap plus standard OPSD consistently improves the accuracy-efficiency frontier over cold-start OPSD across model scales and families, and the code is provided. The stage-order and stage-necessity ablations are consistent with the proposed ordering. However, the paper's central mechanistic claim—that the gains are caused by a shift in the support of visited prefixes—is not directly tested; the reported evidence is compatible with generic initialization benefits or selection effects. The scientific contribution would be much stronger if the prefix-support hypothesis were measured, not just inferred.

major comments (3)
  1. [§4.4, Fig. 3, Table 3] The central mechanism is asserted, not measured. The paper never reports a direct diagnostic of prefix quality: no fraction of off-track prefixes, no entropy of the concise teacher's next-token distributions at visited prefixes, no per-prefix KL sharpness, and no comparison of cold-start vs warm-start rollouts. Figure 1 is therefore an analogy, and the §6 statement that "these results highlight prefix support as a central factor" is unsupported. Please add at least one such diagnostic (e.g., teacher target entropy or off-track fraction before/after SFT) to the paper.
  2. [§3.3, Eq. (4), Table 3] The SFT bootstrap filters on answer correctness, which may select easy problems that the base model can already solve concisely. Because SFT-only already reaches TE 11.56 on MATH-500, close to CRISP's 11.71, the later gains of BIRD could reflect dataset selection rather than improved prefix support. A matched-difficulty control (e.g., SFT on unfiltered concise traces, or on filtered traces matched with base-model solve difficulty) is needed to rule out selection bias and to support the support-placement explanation.
  3. [§4.1, Tables 1–2] Reported numbers are for the checkpoint with the highest TE on each test benchmark (p. 7), which is oracle selection on the test set. No validation split is used. This can inflate the apparent advantage and makes the comparison between methods less meaningful. Please report the final checkpoint or a validation-selected checkpoint, and show checkpoint/stability variability.
minor comments (5)
  1. [Algorithm 1, §4.1] Algorithm 1 does not list T_KL, T_SFT, M, or the LoRA configuration as inputs; provide these values (T_SFT=2048, T_KL=1024, etc.) and the teacher refresh interval M in the experimental setup.
  2. [§4.4, Fig. 3] Figure 3 is based on a single training run; no error bars or seeds are reported. The convergence claim in §4.4 needs at least two seeds.
  3. [§4.2, Table 1] At 1.7B on AIME 2024, BIRD has lower TE than CRISP (3.33 vs 3.60). The "8 out of 9" statement is accurate, but the abstract and §4.2 should acknowledge the one exception.
  4. [References] Several references are 2026 arXiv preprints (e.g., [5], [10], [21], [25]); please verify and update publication status where applicable.
  5. [§3.2, Eq. (3)] In Eq. (3), the teacher p_t is conditioned on c, but the conciseness instruction is introduced only in §3.3; define c before Eq. (3) to avoid confusion.

Circularity Check

0 steps flagged

No significant circularity: BIRD's reported gains are independently evaluated on test benchmarks and are not equivalent to the filtering or KL objective by construction.

full rationale

The central derivation is self-contained. Stage 1 builds a filtered set by a binary correctness check on self-generated traces: D_filtered = {(x,y) | answer(y)=GT(x)} (Eq. 4), and the paper states: "Ground truth is used only for this binary verification step and is never provided as part of the model input during training." Stage 2 uses the same prefix-local reverse-KL objective as conventional OPSD (Eq. 3); the paper explicitly says "The teacher is a stop-gradient copy of the model" and "The objective is the same as in conventional OPSD." Thus the student-teacher relationship is self-distillation by design, not a hidden circular reduction. The reported accuracies and lengths are measured on MATH-500 and AIME (hold-out style benchmarks), so the final numbers are not re-statements of the filtering condition or the KL loss. The SFT bootstrap is fit to correctness-filtered self-generated traces from DAPO-Math-17k-dedup, not to the evaluation labels, and no fitted parameter is renamed as a prediction. The checkpoint-by-TE selection is an evaluation protocol, not a derivation step. The paper's self-citations appear in related-work and method-background contexts (e.g., LoRA, SFT compression methods) and are not load-bearing for the claimed bottleneck or the two-stage algorithm. The skeptical concern that the correctness filter may preferentially select easy problems and that the prefix-support mechanism is not directly measured is an evidence and effect-identification concern, not a circularity of the kind where an output equals an input by construction. Therefore no circularity is found.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

No new entities are introduced. The paper's contribution rests on several domain assumptions about when concise correct traces are trustworthy and about KL at visited prefixes capturing compression signal. Free parameters are mostly engineering choices; the most consequential is the per-benchmark best-TE checkpoint selection, which is disclosed in §4.1 but still affects every headline number.

free parameters (6)
  • KL truncation length T_KL = 1024 tokens
    Chosen by hand in §3.3; changing it could affect compressibility.
  • SFT truncation length T_SFT = 2048 tokens
    Chosen by hand in §3.3; affects which traces constitute the bootstrap.
  • LoRA configuration = Not specified (rank/alpha/learning rate)
    Used for both SFT and OPSD; details omitted, so the effective capacity is a hidden choice.
  • Teacher refresh interval M = Not specified
    Periodic stop-gradient refresh affects distillation stability.
  • Number of samples per problem in Stage 1 = 1
    Single rollout per problem; recall depends on base model's accurate-concise rate.
  • Token Efficiency metric definition = TE = Acc / ln(Length)
    The logarithmic shape is a metric choice that determines which checkpoint is 'best' and which method wins.
axioms (4)
  • domain assumption Self-generated concise traces from the base model, when correct, contain the pattern the student should adopt.
    Central premise of Stage 1; if base-model concise traces are merely short because they skip needed reasoning, the SFT warm-up injects a bad bias.
  • domain assumption The KL divergence at student-sampled prefixes is the right proxy for compression quality.
    Assumed throughout; no independent check that lower KL on visited prefixes predicts a better frontier.
  • domain assumption Answer-correctness filtering is a sufficient quality filter for reasoning traces.
    Eq. (4) filters only on final answer; a correct answer can still come from a flawed or lucky trace.
  • standard math Standard autoregressive next-token factorization for LLMs (Eq. 1).
    Uncontroversial background.

pith-pipeline@v1.3.0-alltime-deepseek · 11871 in / 10868 out tokens · 60929 ms · 2026-08-01T22:26:08.755258+00:00 · methodology

0 comments
read the original abstract

Large reasoning models often solve problems through long chain-of-thought (CoT) traces, yet much of this computation is spent on redundant derivations, repeated self-verification, and detours that do not improve the final answer. Existing on-policy self-distillation methods reduce this cost by matching a student model to a concise copy of itself on prefixes sampled from the student's own rollouts. We show that this objective has an initialization bottleneck. Since supervision is applied only to visited prefixes, training from a verbose base model places the KL loss on contexts that are often noisy, redundant, or already off track. In such regions, a concise teacher can provide only local corrections, while the student continues to explore trajectories that an efficient reasoner should avoid. In this paper, we propose BIRD(Bootstrapped Iterative Self-Reasoning Distillation), a two-stage self-reasoning distillation method that improves the rollout distribution before on-policy training. BIRD first samples concise solutions from the base model under a brevity instruction, keeps only answer-correct traces, and performs a lightweight prompt-switch SFT step. The traces are generated with the brevity instruction but learned under the original task prompt, turning instruction-induced conciseness into a default reasoning behavior. Starting from this warm model, BIRD then applies on-policy reverse-KL distillation with a concise self-teacher, now on cleaner and more informative prefixes. Across Qwen3 series models, BIRD achieves a stronger accuracy-efficiency trade-off than prompting and cold-start on-policy distillation on MATH-500 and AIME benchmarks. On Qwen3-8B, it improves MATH-500 accuracy from 86.2% to 92.0% while reducing the average response length from 3,099 to 1,115 tokens. These results highlight prefix support as a central factor in efficient reasoning distillation.

Figures

Figures reproduced from arXiv: 2607.15736 by Dongxu Zhang, Jihua Zhu, Leichao Dong, Lin Chen, Qirui Wang, Yiding Sun, Yuhan Wang.

Figure 1
Figure 1. Figure 1: Prefix support determines where on-policy KL supervision is applied. Cold-start OPSD [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of BIRD. Stage 1 constructs a static target from one self-generated concise trace per problem, filters it with ground-truth verification, and applies prompt-switch SFT. The traces are generated with the conciseness instruction x + c but learned under the original prompt x. This offline bootstrap moves the student from cold prefix support, which may contain detours and self-checks, to warmer suppor… view at source ↗
Figure 3
Figure 3. Figure 3: Training dynamics of CRISP and BIRD on Qwen3-8B evaluated on MATH-500. Top: average response length. Bottom: Token Efficiency (TE). BIRD starts from the SFT-warmed checkpoint before OPSD updates and maintains lower response length and higher TE throughout training. 4.2 Main Results and Scaling [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative trace analysis on a representative MATH-500 counting problem using Qwen3- [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

38 extracted references · 19 linked inside Pith

  1. [1]

    On-policy distillation of language models: Learning from self-generated mistakes

    Rishabh Agarwal, Nino Vieillard, Yongchao Zhou, Piotr Stanczyk, Sabela Ramos Garea, Matthieu Geist, and Olivier Bachem. On-policy distillation of language models: Learning from self-generated mistakes. In International Conference on Learning Representations, volume 2024, pages 21246–21263, 2024

    2024

  2. [2]

    L1: Controlling how long a reasoning model thinks with reinforcement learning.arXiv preprint arXiv:2503.04697, 2025

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

    Pith/arXiv arXiv 2025

  3. [3]

    Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374, 2021

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374, 2021

    Pith/arXiv arXiv 2021

  4. [4]

    Over-reasoning and redundant calculation of large language models

    Cheng-Han Chiang and Hung-yi Lee. Over-reasoning and redundant calculation of large language models. InProceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers), pages 161–169, 2024

    2024

  5. [5]

    Ctrlcot: Dual-granularity chain-of-thought compression for controllable reasoning.arXiv preprint arXiv:2601.20467, 2026

    Zhenxuan Fan, Jie Cao, Yang Dai, Zheqi Lv, Wenqiao Zhang, Zhongle Xie, Peng LU, and Beng Chin Ooi. Ctrlcot: Dual-granularity chain-of-thought compression for controllable reasoning.arXiv preprint arXiv:2601.20467, 2026

    arXiv 2026

  6. [6]

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

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    Fade: Mitigating hallucinations by reducing language-prior dominance in large vision- language models, 2026

    Yichen Guo, Kai Tang, Fenglai Lin, Yiding Sun, Dongxu Zhang, Wenya Wang, Lin William Cong, and Shanghang Zhang. Fade: Mitigating hallucinations by reducing language-prior dominance in large vision- language models, 2026. URLhttps://arxiv.org/abs/2606.29431

    Pith/arXiv arXiv 2026

  8. [8]

    Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022

    2022

  9. [9]

    Cfms: A coarse-to-fine multimodal synthesis framework for enhanced tabular reasoning.arXiv preprint arXiv:2604.10973, 2026

    Qixian Huang, Hongqiang Lin, Tong Fu, Yingsen Wang, Zhenghui Fu, Qirui Wang, Yiding Sun, and Dongxu Zhang. Cfms: A coarse-to-fine multimodal synthesis framework for enhanced tabular reasoning.arXiv preprint arXiv:2604.10973, 2026

    Pith/arXiv arXiv 2026

  10. [10]

    Stepwise penalization for length-efficient chain-of-thought reasoning.arXiv preprint arXiv:2603.00296, 2026

    Xintong Li, Sha Li, Rongmei Lin, Hongye Jin, Linwei Li, Hejie Cui, Sarah Zhang, Chia-Yuan Chang, Kewei Cheng, Besnik Fetahu, et al. Stepwise penalization for length-efficient chain-of-thought reasoning.arXiv preprint arXiv:2603.00296, 2026

    arXiv 2026

  11. [11]

    Leash: Adaptive length penalty and reward shaping for efficient large reasoning model

    Yanhao Li, Lu Ma, Jiaran Zhang, Lexiang Tang, Wentao Zhang, and Guibo Luo. Leash: Adaptive length penalty and reward shaping for efficient large reasoning model. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2846–2856, 2026

    2026

  12. [12]

    Filter, then reweight: Rethinking optimization granularity in on-policy distillation.arXiv preprint arXiv:2606.02684, 2026

    Yuying Li, Leqi Zheng, Yongzi Yu, Wenrui Zhou, Xuchang Zhong, Xing Hu, Jing Jin, Huangjie Yuan, and Tao Feng. Filter, then reweight: Rethinking optimization granularity in on-policy distillation.arXiv preprint arXiv:2606.02684, 2026

    Pith/arXiv arXiv 2026

  13. [13]

    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

  14. [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 Computa- tional Linguistics (Volume 1: Long Papers), pages 6025–6035, 2025

    2025

  15. [15]

    Revisiting overthinking in long chain-of-thought from the perspective of self-doubt.arXiv preprint arXiv:2505.23480, 2025

    Keqin Peng, Liang Ding, Yuanxin Ouyang, Meng Fang, and Dacheng Tao. Revisiting overthinking in long chain-of-thought from the perspective of self-doubt.arXiv preprint arXiv:2505.23480, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    The benefits of a concise chain of thought on problem-solving in large language models.arXiv preprint arXiv:2401.05618, 2024

    Matthew Renze and Erhan Guven. The benefits of a concise chain of thought on problem-solving in large language models.arXiv preprint arXiv:2401.05618, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    Crisp: Compressed reasoning via iterative self-policy distillation.arXiv preprint arXiv:2603.05433, 2026

    Hejian Sang, Yuanda Xu, Zhengze Zhou, Ran He, Zhipeng Wang, and Jiachen Sun. Crisp: Compressed reasoning via iterative self-policy distillation.arXiv preprint arXiv:2603.05433, 2026. 12 Better Starts, Better Ends: Bootstrapped Iterative Self-Reasoning Distillation for Compressed Reasoning

    Pith/arXiv arXiv 2026

  18. [18]

    Stop overthinking: A survey on efficient reasoning for large language models.arXiv preprint arXiv:2503.16419, 2025

    Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Na Zou, et al. Stop overthinking: A survey on efficient reasoning for large language models.arXiv preprint arXiv:2503.16419, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    Mitigating hallucinations via inter-layer consistency aggregation in large vision-language models, 2026

    Kai Tang, Jinhao You, Yichen Guo, Yiding Sun, Dongxu Zhang, Wenya Wang, Hanze Li, Tao Luo, Renyuan Li, Xiande Huang, and Shanghang Zhang. Mitigating hallucinations via inter-layer consistency aggregation in large vision-language models, 2026. URLhttps://arxiv.org/abs/2505.12343

    Pith/arXiv arXiv 2026

  20. [20]

    Seeme: Mitigating hallucinations in large vision-language models through effective visual token engineering, 2026

    Kai Tang, Jinhao You, Bohua Zhang, Yichen Guo, Yiding Sun, Dongxu Zhang, Chenxi Li, Xiande Huang, and Shanghang Zhang. Seeme: Mitigating hallucinations in large vision-language models through effective visual token engineering, 2026. URLhttps://arxiv.org/abs/2607.04163

    Pith/arXiv arXiv 2026

  21. [21]

    Mitigating overthinking in large reasoning models via difficulty-aware reinforcement learning.arXiv preprint arXiv:2601.21418, 2026

    Qian Wan, Ziao Xu, Luona Wei, Xiaoxuan Shen, and Jianwen Sun. Mitigating overthinking in large reasoning models via difficulty-aware reinforcement learning.arXiv preprint arXiv:2601.21418, 2026

    arXiv 2026

  22. [22]

    Pointrft: Explicit reinforcement fine-tuning for point cloud few-shot learning.arXiv preprint arXiv:2603.23957, 2026

    Yankai Wang, Yiding Sun, Qirui Wang, Pengbo Li, Chaoyi Lu, and Dongxu Zhang. Pointrft: Explicit reinforcement fine-tuning for point cloud few-shot learning.arXiv preprint arXiv:2603.23957, 2026

    arXiv 2026

  23. [23]

    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

  24. [24]

    Stop spinning wheels: Mitigating llm overthinking via mining patterns for early reasoning exit.arXiv preprint arXiv:2508.17627, 2025

    Zihao Wei, Liang Pang, Jiahao Liu, Jingcheng Deng, Shicheng Xu, Zenghao Duan, Jingang Wang, Fei Sun, Xunliang Cai, Huawei Shen, et al. Stop spinning wheels: Mitigating llm overthinking via mining patterns for early reasoning exit.arXiv preprint arXiv:2508.17627, 2025

    arXiv 2025

  25. [25]

    Intern-atlas: A methodological evolution graph as research infrastructure for ai scientists.arXiv preprint arXiv:2604.28158, 2026

    Yujun Wu, Dongxu Zhang, Xinchen Li, Jinhang Xu, Yiling Duan, Yumou Liu, Jiabao Pan, Qiyuan Zhu, Xuanhe Zhou, Jingxuan Wei, et al. Intern-atlas: A methodological evolution graph as research infrastructure for ai scientists.arXiv preprint arXiv:2604.28158, 2026

    Pith/arXiv arXiv 2026

  26. [26]

    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. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 3351–3363, 2025

    2025

  27. [27]

    Tip: Token importance in on-policy distillation.arXiv preprint arXiv:2604.14084, 2026

    Yuanda Xu, Hejian Sang, Zhengze Zhou, Ran He, Zhipeng Wang, and Alborz Geramifard. Tip: Token importance in on-policy distillation.arXiv preprint arXiv:2604.14084, 2026

    Pith/arXiv arXiv 2026

  28. [28]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    Dynamic early exit in reasoning models.arXiv preprint arXiv:2504.15895, 2025

    Chenxu Yang, Qingyi Si, Yongjie Duan, Zheliang Zhu, Chenyu Zhu, Qiaowei Li, Minghui Chen, Zheng Lin, and Weiping Wang. Dynamic early exit in reasoning models.arXiv preprint arXiv:2504.15895, 2025

    arXiv 2025

  30. [30]

    Dapo: An open-source llm reinforcement learning system at scale.Advances in Neural Information Processing Systems, 38:113222–113244, 2026

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, et al. Dapo: An open-source llm reinforcement learning system at scale.Advances in Neural Information Processing Systems, 38:113222–113244, 2026

    2026

  31. [31]

    Not all errors are created equal: Ascot addresses late-stage fragility in efficient llm reasoning.arXiv Prepr

    Dongxu Zhang, Yujun Wu, Yiding Sun, Jihua Zhu, Jinnan Yang, Miao Xin, and Baoliang Tian. Not all errors are created equal: Ascot addresses late-stage fragility in efficient llm reasoning.arXiv Prepr. arXiv:2508.05282, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Not all queries need deep thought: Coficot for adaptive coarse-to-fine stateful refinement

    Dongxu Zhang, Hongqiang Lin, Yiding Sun, Pengyu Wang, Qirui Wang, Ning Yang, and Jihua Zhu. Not all queries need deep thought: Coficot for adaptive coarse-to-fine stateful refinement. InAnn. Conf. Uncertain. Artif. Intell., 2026

    2026

  33. [33]

    Pointcot: A multi-modal benchmark for explicit 3d geometric reasoning

    Dongxu Zhang, Yiding Sun, Pengcheng Li, Yumou Liu, Hongqiang Lin, Haoran Xu, Xiaoxuan Mu, Liang Lin, Wenbiao Yan, Ning Yang, et al. Pointcot: A multi-modal benchmark for explicit 3d geometric reasoning. arXiv Prepr. arXiv:2602.23945, 2026

    arXiv 2026

  34. [34]

    Chain-of- thought compression should not be blind: V-skip for efficient multimodal reasoning via dual-path anchoring

    Dongxu Zhang, Yiding Sun, Cheng Tan, Wenbiao Yan, Ning Yang, Jihua Zhu, and Haijun Zhang. Chain-of- thought compression should not be blind: V-skip for efficient multimodal reasoning via dual-path anchoring. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 19402–19415, 2026. 13 Better...

    2026

  35. [35]

    find the time 1000 days later

    Xinliang Frederick Zhang, Anhad Mohananey, Alexandra Chronopoulou, Pinelopi Papalampidi, Somit Gupta, Tsendsuren Munkhdalai, Lu Wang, and Shyam Upadhyay. Do llms really need 10+ thoughts for “find the time 1000 days later”? towards structural understanding of llm overthinking. InProceedings of the 64th Annual Meeting of the Association for Computational L...

    2026

  36. [36]

    Opsdl: On-policy self-distillation for long-context language models.arXiv preprint arXiv:2604.17535, 2026

    Xinsen Zhang, Zhenkai Ding, Tianjun Pan, Run Yang, Chun Kang, Xue Xiong, and Jingnan Gu. Opsdl: On-policy self-distillation for long-context language models.arXiv preprint arXiv:2604.17535, 2026

    Pith/arXiv arXiv 2026

  37. [37]

    Self- distilled reasoner: On-policy self-distillation for large language models.arXiv preprint arXiv:2601.18734, 2026

    Siyan Zhao, Zhihui Xie, Mengchen Liu, Jing Huang, Guan Pang, Feiyu Chen, and Aditya Grover. Self- distilled reasoner: On-policy self-distillation for large language models.arXiv preprint arXiv:2601.18734, 2026

    Pith/arXiv arXiv 2026

  38. [38]

    Rosd: Reflective on-policy self-distillation for language model reasoning across domains.arXiv preprint arXiv:2605.28014, 2026

    Ziqi Zhao, Xinyu Ma, Liu Yang, Yujie Feng, Daiting Shi, Jingzhou He, Xin Xin, Zhaochun Ren, and Xiao-Ming Wu. Rosd: Reflective on-policy self-distillation for language model reasoning across domains.arXiv preprint arXiv:2605.28014, 2026. 14

    Pith/arXiv arXiv 2026