Pith. sign in

REVIEW 3 major objections 5 minor 17 references

Long-context reasoning models copy the prompt too much; adding a grounding reward and a distractor penalty to reinforcement learning measurably improves accuracy and cuts wasted tokens.

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 ↗

GEAR adds a grounding reward for overlap with key evidence and a distractor penalty to the RL accuracy signal, reducing repetitive copying and improving long-context reasoning accuracy by up to 4.6 points over accuracy-only RL.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Solid diagnosis, plausible reward fix, but headline gains are tuned on the test set and lack variance — deserve referees but not at face value. the 3 major comments →

arxiv 2607.19345 v2 pith:EBBXJUN2 submitted 2026-07-21 cs.CL cs.AI

Copy Less, Ground More: Overcoming Repetitive Copying in Long-Context Reasoning via Evidence-Aware Reinforcement Learning

Lizhe Fang , Weizhou Shen , Tianyi Tang , Yisen Wang This is my paper
classification cs.CL cs.AI
keywords repetitive copyinglong-context reasoningreinforcement learningreward shapingevidence groundingn-gram overlapdistractor penaltythinking length
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 reading

The paper identifies a failure mode in long-context reasoning models: they copy large chunks of the prompt into their thinking traces, more so as context grows, and this "repetitive copying" correlates with wrong answers and inflated token use. The paper argues the root cause is insufficient grounding—models copy indiscriminately instead of focusing on the evidence that matters. To fix it, the paper proposes GEAR, a reward-shaping scheme that adds to the accuracy reward a bonus for n-gram overlap with key evidence and a penalty for overlap with distractor context. Training with this reward across three model scales and five benchmarks improves accuracy by up to 4.6 average points over accuracy-only reinforcement learning, with larger gains at 128k than at 32k contexts, while also reducing repetitive copying and thinking length. A sympathetic reader would care because it suggests a cheap, annotation-light way to make long-context reasoning models both more accurate and more efficient.

Core claim

On its own terms, the paper's central discovery is that long-context reasoning models fail not because they copy from the prompt, but because they copy without grounding: correctly answered samples show higher overlap with key evidence and lower overlap with distractor context. The paper converts this diagnostic into a training signal—the GEAR reward—and shows that combining a grounding reward with a distractor penalty yields consistent accuracy gains over accuracy-only RL across three model scales and five held-out benchmarks, with the biggest gains at 128k contexts, while also reducing repetitive copying and thinking length. Both reward components are necessary; the grounding bonus alone m

What carries the argument

The central object is the GEAR reward: R(x,y) = R_acc + α·Overlap_n(y∥x_key) − β·Overlap_n(y∥x_dist), where Overlap_n is the fraction of n-gram windows in the reasoning trace that appear in the designated evidence (or distractor) part of the prompt. It works by giving the policy a dense, string-matching signal to engage with task-relevant evidence and avoid transcribing irrelevant context. The n-gram overlap measure (Eq. 1) is the load-bearing mechanism: it requires only support annotations and no external verifier or retrieval module.

Load-bearing premise

The automated pipeline that samples 1–3 random chunks as the designated evidence assumes those chunks are exactly what a correct answer needs; if a question can be answered from elsewhere in the document, the grounding reward rewards text-matching and the distractor penalty punishes legitimate reasoning.

What would settle it

Train GEAR on a long-context QA set constructed so that the designated evidence chunks are arbitrary and the question is answerable from other parts of the document; if the GEAR-trained model does not beat accuracy-only RL, the reward is riding on annotation artifacts rather than grounding.

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

If this is right

  • Long-context RL can be improved without extra verifiers, using only n-gram overlap statistics computed from existing evidence annotations.
  • The accuracy gains transfer to contexts four times longer than the training distribution (128k vs. 16–32k), suggesting that evidence-grounding behavior generalizes to unseen context lengths.
  • Both reward components are essential: removing the distractor penalty or the grounding reward hurts accuracy, so reward design for long-context reasoning should pair positive and negative shaping.
  • GEAR simultaneously reduces repetitive copying and thinking length, implying efficiency gains that compound with accuracy gains.
  • The method works on both synthetic and natural-language documents, so it may extend to arbitrary corpora through the proposed automated annotation pipeline.

Where Pith is reading between the lines

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

  • The grounding ratio diagnostic (key-overlap divided by distractor-overlap) could serve as a cheap, interpretable evaluation-time probe for long-context comprehension, complementing end-task accuracy.
  • The result implies that suppressing copying outright is the wrong target; selective copying from relevant evidence may be a productive strategy, so methods that penalize all prompt reuse could backfire.
  • The same reward shape could be applied to general retrieval-augmented settings where evidence is selected by a retriever rather than by construction, potentially improving grounding in RAG pipelines.
  • The reported gains are modest in absolute terms and may be benchmark-specific; whether they compound with scale or saturate at even longer contexts is untested.
Share X Bluesky LinkedIn Reddit HN

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. This paper identifies a failure mode in long-context reasoning models: repetitive copying of input text into reasoning traces. Using GSM-Infinite, the authors show that this behavior increases with context length, correlates with lower accuracy, and that correct answers exhibit more selective overlap with key evidence versus distractor context. They propose GEAR, a reward that augments an accuracy signal with a grounding reward α·Overlap_n(y∥x_key) and a distractor penalty β·Overlap_n(y∥x_dist) (Eq. 3), and they build an automated three-stage pipeline to generate evidence-annotated QA pairs from arbitrary documents. Experiments across three Qwen3.5 model scales and five benchmarks report consistent average accuracy gains over accuracy-only GSPO of up to +4.6 points, with reduced repetitive copying and shorter thinking traces.

Significance. The diagnosis of repetitive copying in long-context reasoning is timely and well-motivated, and GEAR is a simple, cheap, and potentially widely applicable reward-shaping idea. The automated evidence-annotation pipeline is a useful contribution that could extend the method beyond synthetic benchmarks. If the empirical claims are robust, this paper would provide a practical way to improve grounding in long-context RL and would sharpen understanding of a real failure mode. However, the current support for the headline numbers is weakened by (i) hyperparameter selection on the same test benchmarks, and (ii) the absence of any variance or multi-seed reporting for the RL runs. These issues are fixable but must be addressed before the central claim can be taken at face value.

major comments (3)
  1. [§5.3, Tables 3 and 4] The default hyperparameters (α=0.1, β=0.3, n=3) are selected by maximizing average accuracy on the same four 32k benchmarks that later appear in Table 1. The 32k gains in Table 1 are therefore selected results, not unbiased estimates. The 128k results use hyperparameters chosen on the 32k subsets of the same benchmark families, so selection effects may propagate. Please use a held-out validation set for hyperparameter selection (or pre-specify hyperparameters) and report test scores only after fixing the configuration. Without this, the 'up to +4.6 average points' claim is not properly supported.
  2. [§5.1, Table 1] Every row of Table 1 is a single RL run; no seeds, variances, or significance tests are reported. GSPO/GRPO-style training is typically noisy, and several reported gains are small (e.g., +1.5 for Qwen3.5-27B at 32k, +2.1 at 128k). These differences may fall within run-to-run variation. Please report means and standard deviations over at least 3 random seeds per configuration, and ideally paired comparisons across benchmarks. This is necessary to establish the central claim of consistent improvement.
  3. [§4.2, Stage 2 and Stage 3] The natural-language pipeline samples 1–3 random chunks as the 'constraint context' and only verifies that the generated question is answerable from that region; it does not test whether the question is also answerable from the distractor context. Thus a generated question may have multiple valid evidence regions. In such cases the distractor penalty can penalize legitimate reasoning that uses other parts of the document, and the grounding reward can reward coincidental n-gram overlap. The authors should quantify how often the designated evidence is unique (e.g., by re-answering from distractor chunks or a human sample) and filter or down-weight ambiguous pairs. This is important for the construct validity of the training reward.
minor comments (5)
  1. [§5.2, Table 2] The reported reduction in Overlap_n(y∥x_dist) is partly the direct objective of the distractor penalty, so the causal claim that accuracy gains 'stem from' reduced indiscriminate copying is not fully established. Consider an additional analysis that isolates the effect of the penalty (e.g., ablating the reward term while keeping the same data) or soften the causal wording.
  2. [§4.3] The sentence 'our ablation study confirms this is the optimal setting' refers to Tables 3–4, which are performed on the test benchmarks (see Major Comment 1). This wording should be revised once hyperparameters are selected on a proper validation split.
  3. [Eq. (1)] The typesetting of the fraction in Overlap_n has a line-break artifact ('m−n+ 1 m−n+1X'). Please clean up the equation presentation.
  4. [§3.2, Figure 1] The diagnostic trends in Figure 1 are monotonic and convincing, but adding error bars or per-model variance would strengthen the claim that repetitive copying intensifies with context length, especially because the paper emphasizes single-run results elsewhere.
  5. [Table 1] The base-model rows at 32k have an inconsistent placeholder for AA-LCR (the dash is missing in the 'Base' rows for 32k). Use '—' consistently for benchmarks not evaluated at that context length.

Circularity Check

2 steps flagged

Gains are benchmark-tested and not definitionally forced, but the headline hyperparameters are selected on the very evaluation benchmarks used to report the +4.6 gain, and the mechanism check partly restates the reward objective.

specific steps
  1. fitted input called prediction [§4.3 Training setup; §5.3 Ablation study (Tables 3–4); Table 1 (Qwen3.5-9B, 32k)]
    "We set n= 3... as shorter n-grams provide denser reward signal that is more effective for RL optimization, and our ablation study confirms this is the optimal setting. ... The default setting (α= 0.1, β= 0.3) achieves the best average of 81.3."

    The default hyperparameters (α=0.1, β=0.3, n=3) are chosen because they maximize average accuracy on the same four benchmarks (Ruler, LongBench-v2, BrowseComp-LC, GraphWalks at 32k) that are reported in Table 1. Table 3's best average (81.3) is literally the GEAR row's average for Qwen3.5-9B at 32k in Table 1, and the same configuration underlies the 128k +4.6 headline. Thus the reported 'consistent improvements' are in-sample grid maxima on the evaluation set, not predictions under a pre-fixed configuration. The gains are not definitionally forced (the RL runs could have failed), so this is a protocol-level selection rather than a full reduction, but the headline magnitude is selected, not predicted.

  2. other [§4.1 Eq. 3; §5.2 Effect on Repetitive Copying and Grounding]
    "R(x, y) =R_acc +α·Overlap_n(y∥x_key)−β·Overlap_n(y∥x_dist) (3) ... Only when the distractor penalty R_dist is added does the overlap drop below all other configurations: 27.0% on Ruler and 22.6% on LongBench-v2."

    The distractor penalty is defined as −β·Overlap_n(y∥x_dist) with the stated purpose of 'discouraging' indiscriminate copying (§4.1), so demonstrating that adding this penalty reduces copying is partly a restatement of the objective rather than an independent confirmation of the mechanism. The non-trivial part survives because Table 2 measures overlap against the whole input (no key/distractor split) on unseen benchmarks, so the observed drop does have transfer content; only the causal attribution 'the gains stem from reduced indiscriminate copying' (contribution 4) is inferred, not separately identified.

full rationale

The paper's central derivation is not circular in the strong sense: the diagnosis in §3.4 is an observational correlation on GSM-Infinite (correct samples have higher key-overlap and lower distractor-overlap), and §4.1's reward operationalizes that correlation, but the central accuracy claim is tested against five benchmarks (Ruler, LongBench-v2, BrowseComp-LC, GraphWalks, AA-LCR) that are disjoint from the 3.2k training mixture. The accuracy signal R_acc is external to the grounding terms, the ablation shows the grounding-only variant hurts, and the full GEAR reward still requires a genuinely successful RL run to improve accuracy. The diagnosis→reward loop is therefore standard observation→intervention science, not definitional equivalence. Self-citations (Fang et al. 2025, Wu et al. 2025, QwenLong) are related-work context and are not load-bearing; there is no imported uniqueness theorem or ansatz smuggled via citation. The circularity burden is real but contained: (1) the default (α,β,n) are tuned on the same four 32k benchmarks that appear in Table 1, with the best grid average (81.3) reported verbatim as GEAR's headline number, so the +4.6 claim is an in-sample selected maximum rather than an out-of-sample prediction — a statistical-inference flaw rather than a forced equivalence; and (2) the §5.2 'mechanism' check that the distractor penalty reduces copying partially restates the penalty's definition, although the whole-prompt overlap measurements on held-out benchmarks give it some independent content. Accordingly, the score is 3: partial circularity at the protocol level, while the central empirical claim retains independent content and would need pre-specified validation splits and seed variance to be taken at face value.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

No new entities are postulated. The grounding ratio (Eq. 2) and overlap-rate metrics (Eq. 1) are formal definitions, not invented entities with independent falsifiable handles; the GEAR reward is a training objective whose value is tested solely via downstream benchmark accuracy.

free parameters (4)
  • α (grounding reward weight) = 0.1
    Chosen after sensitivity analysis on the same four benchmarks (Ruler, LB-v2, BrowseComp-LC, GraphWalks) used for the main results; Table 3 shows the default setting achieves the best average.
  • β (distractor penalty weight) = 0.3
    Same ablation: β must not be too small relative to α; β=0.1 drops 2.9 points, and β=0.3 is best.
  • n (n-gram size) = 3 (training), 10 (diagnostics)
    Selected by ablation (Table 4): n=3 gives the densest reward signal for RL; n=10 is used in the diagnosis for precision. A hand choice that shapes the reward's behavior.
  • Training data composition = 3,200 samples (1k GSM-Infinite, 1k PhantomWiki, 1.2k RedPajama-v2)
    Manual mixture choice; the central claim's generality across synthetic and natural-language data depends on it.
axioms (5)
  • domain assumption n-gram overlap (Eq. 1) is a valid proxy for copying and grounding: high overlap with the prompt = copying; overlap with support spans = productive engagement.
    Invoked throughout §3 for the diagnosis and in Eq. 3 for the reward itself. A false proxy would invalidate both diagnosis and method, since incidental lexical overlap is counted as copying (partly mitigated by using n=10 in diagnostics, but the training reward uses n=3).
  • domain assumption GSM-Infinite's procedural support indices correctly partition prompts into key evidence vs. distractor context.
    §3.4 relies on this to define the grounding ratio (Eq. 2) and to establish the correlation between selective grounding and correctness.
  • ad hoc to paper Sampling 1–3 random chunks as the constraint context yields a faithful label of task-relevant evidence for natural-language QA.
    §4.2 Stage 2: the auto-pipeline defines key evidence by construction (the chunks the question was generated from), but nothing guarantees these chunks are the unique or true evidence, so the reward can penalize legitimate alternative reasoning paths.
  • domain assumption Single-valued answers verified by exact string match are a sufficient accuracy signal for RLVR on natural-language data.
    §4.2 Stage 3 restricts answers to numbers or entity phrases to automate reward computation; format-matching rather than semantic correctness may reward or punish the wrong thing.
  • domain assumption The correlation between grounding ratio and correctness (§3.4) is causally exploitable: optimizing grounding via reward shaping transfers to held-out benchmarks.
    The entire GEAR design premise. The held-out accuracy gains support it, but the mechanism is only verified with the metric the reward directly optimizes (§5.2), leaving the causal chain partially open.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Copy Less, Ground More: Overcoming Repetitive Copying in Long-Context Reasoning via Evidence-Aware Reinforcement Learning." pith.science (2026). https://pith.science/paper/EBBXJUN2

@misc{pith2026260719345,
  author       = {Pith},
  title        = {Pith review of: Copy Less, Ground More: Overcoming Repetitive Copying in Long-Context Reasoning via Evidence-Aware Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EBBXJUN2}},
  note         = {Machine review of arXiv:2607.19345}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large language models that generate step-by-step reasoning traces have achieved strong performance on complex tasks, and extending them to long-context settings has emerged as an important frontier. However, we identify a critical failure mode in this regime: \emph{repetitive copying}, where models extensively copy text from the input into their reasoning traces rather than productively solving the problem. We show that this behavior is pervasive across frontier long-context LLMs and intensifies with context length. By separating each prompt into task-relevant key evidence and irrelevant distractor context, we further show that the root cause is insufficient grounding: models copy from the prompt indiscriminately, and those that fail to focus on key evidence are far more likely to answer incorrectly. Motivated by this diagnosis, we propose GEAR (Grounding Evidence-Aware Reward), a reward shaping method that augments the accuracy signal with a grounding reward for overlap with key evidence and a distractor penalty for overlap with irrelevant context. To enable GEAR on natural-language data, we develop an automated pipeline that constructs evidence-annotated training data from arbitrary documents. We validate GEAR across multiple model scales and benchmarks, showing consistent improvements of up to +4.6 average points over standard RL with accuracy-based rewards, with larger gains at longer contexts, while also reducing repetitive copying and thinking length. Our findings suggest that, even as long-context evaluation shifts from simple retrieval toward complex reasoning, accurate grounding in relevant evidence remains an indispensable capability with substantial room for improvement.

Figures

Figures reproduced from arXiv: 2607.19345 by Lizhe Fang, Tianyi Tang, Weizhou Shen, Yisen Wang.

Figure 1
Figure 1. Figure 1: n-gram overlap rate on GSM-Infinite as a function of context length across seven models. All models copy more as context grows, and the high 10-gram overlap confirms that copied spans are long contiguous passages. are known exactly, enabling precise measurement of what the model copies, i.e., whether it targets task-relevant evidence or merely reproduces filler text. We construct datasets at context scales… view at source ↗
Figure 2
Figure 2. Figure 2: Repetitive copying harms both accuracy and reasoning efficiency (DeepSeek-V3.2 on [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Copying behavior of correctly vs. incorrectly answered samples (DeepSeek-V3.2 on [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Accuracy and thinking length by 10-gram overlap rate bin across six models on GSM [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Probability density of grounding ratio ( [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Base model: the thinking trace is dominated by repetitive copying of the input, consuming [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: GSPO model: avoids repetitive input copying but enters a degenerate repetition loop, [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: GEAR model: describes a procedural strategy, reasons about word frequencies without [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Base model on LongBench-v2 Academic/32k: the model identifies the correct answer early [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: GSPO model on the same sample: reaches the correct answer but still exhibits moderate [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: GEAR model on the same sample: systematic analysis of each choice, identifies C as [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗

Review history (2 revisions) →

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

17 extracted references · 13 linked inside Pith

  1. [3]

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

    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,

    Pith/arXiv arXiv

  2. [4]

    Ruler: What’s the real context size of your long-context language models?arXiv preprint arXiv:2404.06654,

    Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg. Ruler: What’s the real context size of your long-context language models?arXiv preprint arXiv:2404.06654,

    Pith/arXiv arXiv

  3. [6]

    Golongrl: Capability-oriented long context reinforcement learning with multitask alignment.arXiv preprint arXiv:2605.19577,

    Minxuan Lv, Tiehua Mei, Tanlong Du, Junmin Chen, Zhenpeng Su, Ziyang Chen, Ziqi Wang, Zhennan Wu, Ruotong Pan, Ruiming Tang, et al. Golongrl: Capability-oriented long context reinforcement learning with multitask alignment.arXiv preprint arXiv:2605.19577,

    Pith/arXiv arXiv

  4. [7]

    Repetitions are not all alike: distinct mechanisms sustain repetition in language models.arXiv preprint arXiv:2504.01100,

    Matéo Mahaut and Francesca Franzon. Repetitions are not all alike: distinct mechanisms sustain repetition in language models.arXiv preprint arXiv:2504.01100,

    arXiv

  5. [8]

    Longr: Unleashing long-context reasoning via reinforcement learning with dense utility rewards.arXiv preprint arXiv:2602.05758,

    Bowen Ping, Zijun Chen, Yiyao Yu, Tingfeng Hui, Junchi Yan, and Baobao Chang. Longr: Unleashing long-context reasoning via reinforcement learning with dense utility rewards.arXiv preprint arXiv:2602.05758,

    arXiv

  6. [9]

    Deepseekmath: Pushing the limits of mathemat- ical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathemat- ical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    Pith/arXiv arXiv

  7. [11]

    Fanqi Wan, Weizhou Shen, Shengyi Liao, Yingcheng Shi, Chenliang Li, Ziyi Yang, Ji Zhang, Fei Huang, Jingren Zhou, and Ming Yan

    URL https://qwenlm.github.io/blog/qwq-32b/. Fanqi Wan, Weizhou Shen, Shengyi Liao, Yingcheng Shi, Chenliang Li, Ziyi Yang, Ji Zhang, Fei Huang, Jingren Zhou, and Ming Yan. Qwenlong-l1: Towards long-context large reasoning models with reinforcement learning.arXiv preprint arXiv:2505.17667,

    Pith/arXiv arXiv

  8. [12]

    Loongrl: Reinforcement learning for advanced reasoning over long contexts.arXiv preprint arXiv:2510.19363,

    Siyuan Wang, Gaokai Zhang, Li Lyna Zhang, Ning Shang, Fan Yang, Dongyao Chen, and Mao Yang. Loongrl: Reinforcement learning for advanced reasoning over long contexts.arXiv preprint arXiv:2510.19363,

    arXiv

  9. [13]

    Browsecomp: A simple yet challenging benchmark for browsing agents.arXiv preprint arXiv:2504.12516,

    Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Amelia Glaese. Browsecomp: A simple yet challenging benchmark for browsing agents.arXiv preprint arXiv:2504.12516,

    Pith/arXiv arXiv

  10. [14]

    When more is less: Understanding chain-of-thought length in llms.arXiv preprint arXiv:2502.07266,

    Yuyang Wu, Yifei Wang, Ziyu Ye, Tianqi Du, Stefanie Jegelka, and Yisen Wang. When more is less: Understanding chain-of-thought length in llms.arXiv preprint arXiv:2502.07266,

    Pith/arXiv arXiv

  11. [15]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    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,

    Pith/arXiv arXiv

  12. [16]

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

    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. InNeurIPS, 2025a. Tianyu Yu, Bo Ji, Shouli Wang, Shu Yao, Zefan Wang, Ganqu Cui, Lifan Yuan, Ning Ding, Yuan Yao, Zhiyuan Liu, et al. Rlpr: Extrapolating rlvr to ...

    Pith/arXiv arXiv

  13. [17]

    Gsm-infinite: How do your llms behave over infinitely increasing context length and reasoning complexity?arXiv preprint arXiv:2502.05252,

    Yang Zhou, Hongyi Liu, Zhuoming Chen, Yuandong Tian, and Beidi Chen. Gsm-infinite: How do your llms behave over infinitely increasing context length and reasoning complexity?arXiv preprint arXiv:2502.05252,

    Pith/arXiv arXiv

  14. [2023]

    Deepseek-v3

    Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, et al. Deepseek-v3. 2: Pushing the frontier of open large language models.arXiv preprint arXiv:2512.02556,

    Pith/arXiv arXiv

  15. [2024]

    Qwenlong-l1

    Weizhou Shen, Ziyi Yang, Chenliang Li, Zhiyuan Lu, Miao Peng, Huashan Sun, Yingcheng Shi, Shengyi Liao, Shaopeng Lai, Bo Zhang, et al. Qwenlong-l1. 5: Post-training recipe for long-context reasoning and memory management.arXiv preprint arXiv:2512.12967,

    arXiv

  16. [2025]

    Chatglm: A family of large language models from glm-130b to glm-4 all tools.arXiv preprint arXiv:2406.12793,

    Team Glm, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools.arXiv preprint arXiv:2406.12793,

    Pith/arXiv arXiv

  17. [2026]

    Do not think that much for 2+ 3=? on the overthinking of o1-like llms.arXiv preprint arXiv:2412.21187,

    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,

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.