Pith. sign in

REVIEW 4 major objections 4 minor 295 references

QLPO shows that resampling a GRPO group—not touching the reward—cuts reasoning length by 30–70% while preserving accuracy.

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 06:39 UTC pith:I6TKDVFR

load-bearing objection Practical idea, broad empirical sweep, but the method as written is underspecified and no code exists — the reported gains may hinge on an unstated fallback. the 4 major comments →

arxiv 2607.21793 v1 pith:I6TKDVFR submitted 2026-07-23 cs.AI

QLPO: Quadrant-weighted Sampling for Length-aware Policy Optimization

Siwei Chen , Siqi Chen , Xupeng Miao , Bin Cui This is my paper
classification cs.AI
keywords reinforcement learningpolicy optimizationresponse length controlchain-of-thought compressionGRPOquadrant-weighted samplingLLM reasoninginference efficiency
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.

QLPO claims that the length explosion in reasoning models can be curbed without any reward modification: it over-generates candidate responses per prompt, then resamples the training group to favor short correct and long incorrect trajectories while keeping the empirical correct/incorrect ratio. Because group-relative advantages still reflect correctness, the policy learns concise successful reasoning and suppresses verbose failures. Across 1.5B to 32B models, the method cuts average response length by 30–70% while matching or slightly improving accuracy on hard mathematical and scientific benchmarks. The authors' Section F limits the method to verbose chain-of-thought regimes and notes over-generation can add training cost; those boundaries constrain deployment, not the accuracy-preservation claim. If the claim holds, QLPO is a drop-in resampling change that buys large inference-cost savings with little tuning.

Core claim

On its own terms, the paper's discovery is that the empirical composition of the rollout group—not the reward function—carries the length signal. QLPO draws K=16 responses, labels each by correctness and by whether it is shorter or longer than the class median, and selects M=8 responses with target counts that keep the correct/incorrect ratio of the over-generated pool while oversampling (Correct, Short) and (Incorrect, Long) according to a coefficient α=1/3. The authors argue through a group–trajectory gradient decomposition that this changes which trajectories contribute token-level gradient mass without perturbing the group-level correctness signal: length penalties can flip the sign of a

What carries the argument

Quadrant-weighted sampling: partition each prompt's over-generated candidate pool into four quadrants by correctness (correct/incorrect) and relative length (short/long, split at the class median), then compute target counts N+ and N− from the empirical ratio and allocate them to quadrants with a length-preference coefficient α (1/3 in experiments), using ceilings to protect preferred quadrants. The resampled group of size M then feeds the standard GRPO objective unchanged. The paper's supporting analytic device is a two-level gradient decomposition separating group-level advantage allocation (A(y) = (r−r̄)/σr) from trajectory-level token accumulation, which explains why selection can bias t

Load-bearing premise

The load-bearing premise is that after resampling by length, the group-relative advantages still encode the dominant correctness signal—that preserving the empirical correct/incorrect ratio keeps the baseline and scale undistorted so that correct long trajectories are not pushed to negative advantage; if length-based selection shifts the baseline, accuracy preservation fails even though the reward was never modified.

What would settle it

Train with QLPO on a hard benchmark while logging, for each selected group, the sign of the group-relative advantage of correct-long trajectories. If a nontrivial share (say, >10% averaged over training) of correct-long trajectories receive negative advantages, or if a rollout-budget-matched run (K=16,M=8 vs K=16,M=16) shows an accuracy drop on OlympiadBench or AIME beyond seed-level noise (~±1–2%), the accuracy-preservation claim would be refuted.

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

If this is right

  • Deployed reasoning models can serve answers with far fewer tokens per query, directly lowering inference latency and cost without a tuned length-penalty coefficient.
  • Because accuracy is preserved on hard multi-step benchmarks, the compressed responses still contain the reasoning needed for correctness, so surface verbosity is not a necessary component of performance.
  • Any GRPO-style group-based RL pipeline can adopt QLPO by adding candidate over-generation and a resampling step, leaving reward, advantage, and optimization code intact.
  • Preliminary results on multimodal geometry and code generation suggest the resampling mechanism transfers beyond text-only chain-of-thought.
  • The rollout-budget-matched comparison indicates the efficiency gain does not come from extra samples alone, and training wall-clock time rises at most about 16% (and can be lower).

Where Pith is reading between the lines

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

  • The same quadrant logic could be extended to non-binary trajectory attributes—step-level quality, confidence, or difficulty—potentially refining the accuracy-length frontier beyond what binary correctness permits; this is a natural next step the authors mention but do not take.
  • The method's asymmetry predicts a specific training signature: the length distribution of incorrect trajectories should compress faster than that of correct ones. The reported 38% vs 23% shifts are consistent with that prediction, making it checkable on other models.
  • Because QLPO acts on data selection rather than rewards, it is likely composable with explicit length penalties or adaptive difficulty controllers; a testable extension would combine both and measure whether compression compounds without accuracy loss.
  • If accuracy ever degrades under QLPO on a new benchmark, the two-level gradient view localizes the likely cause—baseline shift in group-relative advantages—before blaming the resampling idea itself, giving practitioners a diagnostic.

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

4 major / 4 minor

Summary. The paper proposes QLPO, a resampling layer on top of GRPO. For each prompt it samples K=16 candidate responses, splits them into correct/incorrect and short/long quadrants by per-class median length, and selects M=8 responses while approximately preserving the pool's correct/incorrect ratio and biasing toward short-correct and long-incorrect responses. The selected group is then used in the standard GRPO advantage computation and policy update. The authors report accuracy--length improvements across 1.5B--32B models, claim 30--70% length reductions, and introduce a group--trajectory gradient decomposition to explain why the method preserves correctness while shortening outputs. A rollout-budget-matched comparison (Table 2) is used to separate the effect of selection from the effect of additional sampling.

Significance. The core idea is practically significant: if the reported gains reproduce, QLPO offers a simple, reward-function-agnostic way to control CoT length, and the K=16 rollout-matched comparison in Table 2 is the right experiment to address the 'more rollouts' confounder. The evaluation breadth over dense/MoE models, base and reasoning models, and extensions to multimodal and code settings is a genuine strength. However, the method as written is not fully specified: the target quadrant counts are often infeasible under the natural no-replacement reading, and no code is released. The headline 30--70% range also overstates the per-benchmark reductions in Table 1. The central idea is defensible, but the paper needs revision to make the algorithm executable and to align the claims with the evidence.

major comments (4)
  1. [§3.1, Eq. (3)] As written, QLPO is not executable for many prompts. With α=1/3, the target count for the preferred quadrants is bN_long^- = ceil(3N^-/4) and bN_short^+ = ceil(3N^+/4), but after the median split the long-incorrect (respectively short-correct) quadrant has at most ceil(N^-/2) (respectively ceil(N^+/2)) trajectories. For N^-=5 the procedure requests 4 long-incorrect trajectories but only 3 exist; all-correct/all-incorrect pools create even larger shortages. The text does not state whether sampling is with replacement, whether counts are capped, or how shortages are resolved. This matters because r̄, σ_r, and the token-level normalization in Eq. (1) depend on the exact composition of the selected group. Without a precise specification or release of the implementation, the reported length/accuracy numbers cannot be attributed to the described algorithm. Please specify the fallback and confi
  2. [Abstract, §4.2/Table 1] The abstract and conclusion state that QLPO reduces response length by 30--70%, but Table 1 contains numerous settings below 30% and only one at 67.6%. Examples: DeepScaleR-1.5B GPQA reduces 16.4%, Qwen2.5-32B GSM8K 13.5% and MATH-500 21.7%, Qwen3-30B-A3B GPQA 12.4% and OlympiadBench 25.4%. Even the aggregate reductions quoted in §5.3 are 30--38%, not 30--70. The headline should be reworded to 'up to ~68%' or should report the actual per-pair range over the evaluated matrix.
  3. [Table 1, Appendix C] The accuracy-preservation claim is supported by seed-level variation for only two of the five model settings. Table 1 reports a single run for DeepScaleR-1.5B, Qwen2.5-32B, and Qwen3-30B-A3B; §4.2's assertion that small drops are 'well within ordinary training stochasticity' is not verified for those models. Since the central claim is 'preserving reasoning performance,' please report seed or evaluation variance for all main settings, or temper the claim to the settings with repeated runs.
  4. [§3.2, Eq. (2)] The group--trajectory decomposition in §3.2 is used to argue that QLPO, unlike length penalties, preserves the correctness-oriented signal of the original advantage. However, the decomposition ignores the clipping in Eq. (1), and the paper does not show in a quantifiable sense that the advantages after resampling remain close to the original advantages. Preserving the correct/incorrect count ratio is not equivalent to preserving the advantage distribution, especially when N+ or N- is small or when the labels used for quadrant assignment contain errors. At minimum, provide a diagnostic (e.g., fraction of selected responses whose advantage sign/magnitude changes relative to the K=16 pool) and report the α=1 control numerically to isolate the length preference from random ratio-preserving resampling.
minor comments (4)
  1. [Figures 5--7, §5.2] The rendered text in Figures 5--7 and in §5.2 contains literal '/uni000...' path strings. These need to be replaced with actual figure images or proper figure text; in the current form the ablation and distribution-shift results are not readable.
  2. [§3.1] The 'bN' notation in the count formulas should be defined explicitly and typeset consistently. Also clarify that the median split is performed separately for each prompt and each correctness class, and what happens when a class has zero or one trajectory.
  3. [Appendix B] The worked example is anecdotal. The section title promises an explanation of 'higher accuracy,' but the example only illustrates compression of a simple equation-solving response. Either report quantitative evidence tying the mechanism to accuracy or retitle the section.
  4. [§4.1] The evaluation generates multiple responses per prompt (3 or 16) but reports pass@1 accuracy. Please clarify whether this is per-response accuracy averaged over samples or accuracy of a single designated sample, since the definition affects how the length/accuracy numbers should be interpreted.

Circularity Check

1 steps flagged

Minor definitional framing in the correctness-preservation narrative; the empirical accuracy/length claims are not circular.

specific steps
  1. self definitional [Sec. 3.1 (target-count equations) and Sec. 3.2 ('A Group–Trajectory Gradient View')]
    "We then set the target positive and negative counts in the final group as N+(x) = round(M · Ñ+(x)/K) and N−(x) = M − N+(x), so that the final group contains exactly M samples while approximately preserving the empirical correct/incorrect ratio of the over-generated pool. ... By contrast, QLPO preserves the dominant correctness-oriented signal carried by the original group-relative advantage."

    The claim that QLPO 'preserves the dominant correctness-oriented signal' is a restatement of the Sec. 3.1 construction, which defines the resampled group to approximately keep the empirical correct/incorrect ratio. It is not derived from the gradient analysis: the selected group's r̄ and σ_r are recomputed from the resampled M responses in Eq. (1), so preserving the binary ratio does not by itself guarantee that group-relative advantages keep their sign or magnitude. The paper uses this by-construction property as the explanation for accuracy preservation. This is explanatory framing rather than a fitted prediction, and the central accuracy/length results come from external benchmark comparisons, so the circularity is minor.

full rationale

QLPO's central claims are empirical: the trained policies are evaluated on five external benchmarks against GRPO, GFPO, rollout-matched GRPO, and compression baselines, and the rollout-budget-matched control (K=16, M=16 vs K=16, M=8) isolates the effect of selection from extra sampling. The dependence on prior work is normal: GRPO, GFPO, DAPO, and LSPO are external baselines; the only author-overlapping reference (DARTS) appears in the reference list and is not load-bearing in the body. The one definitional element is the Sec. 3.2 narrative that resampling 'preserves the dominant correctness-oriented signal,' which restates the count-preserving construction rather than proving the group-relative advantage signal is unchanged. Because the reported accuracy-length trade-offs are not derived from this narrative, this is a minor rhetorical circularity, not a forced result.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

No invented entities. Two hyperparameters (alpha, K) are tuned and carry the method's behavior. The analysis depends on unproved domain assumptions about binary verifier quality, median-length quadrant splits, and a clipping-free gradient decomposition.

free parameters (2)
  • length preference coefficient alpha = 1/3
    Chosen by ablation on DeepScaleR-1.5B-Preview/DAPO-MATH; it controls the within-class retention ratio and therefore the strength of the length bias in every reported run.
  • candidate pool size K = 16 = 2M
    Selected by ablation on DeepScaleR-1.5B-Preview; it must exceed the update group size M=8 for quadrant resampling to have effect.
axioms (4)
  • standard math Group-relative advantage estimates in GRPO are a valid baseline for policy-gradient updates in this setting.
    Invoked via Eq. (1) and Sec. 2; standard but unproved in this paper.
  • domain assumption The binary correctness labels (rule-based matching plus GPT-4.1 fallback) accurately partition candidates into correct and incorrect on all five benchmarks.
    Sec. 4.1 describes the verifier; if labels are noisy, the quadrant split and preserved ratio are built on noise.
  • ad hoc to paper Splitting each correctness class at the median length and resampling with alpha=1/3 preserves the dominant correctness signal while inducing the desired length bias.
    Secs. 3.1-3.2 assert this; there is no theoretical guarantee that the resampled group's baseline r-bar and sigma-r do not suppress long correct reasoning.
  • ad hoc to paper The group-trajectory gradient decomposition remains a faithful description when clipping and token-level loss normalization are active.
    Sec. 3.2 explicitly sets clipping aside ('For simplicity, we ignore the clipping-induced case distinction'); later qualitative claims about GFPO failure rely on this simplification.

pith-pipeline@v1.3.0-alltime-deepseek · 18429 in / 15080 out tokens · 146124 ms · 2026-08-01T06:39:45.868810+00:00 · methodology

0 comments
read the original abstract

Recent large reasoning models often develop long chain-of-thought responses during reinforcement learning (RL), resulting in high inference latency and deployment cost. Existing methods for response length control typically rely on explicit length penalties or additional control modules, which require careful tuning and may compromise reasoning quality. We propose Quadrant-weighted Sampling for Length-aware Policy Optimization (QLPO), a simple resampling-based variant of GRPO that introduces implicit length control without modifying the reward function. QLPO first over-generates candidate responses and then resamples the training group by preserving the empirical correct/incorrect ratio while favoring short correct responses and long incorrect responses. This reshapes the training distribution and implicitly encourages shorter model outputs. Across models ranging from 1.5B to 32B parameters, including both base models and strong reasoning models, QLPO consistently improves the accuracy-length trade-off. It reduces response length by 30% to 70% while preserving reasoning performance. These results suggest that structured resampling provides an effective and robust approach to efficient reasoning.

Figures

Figures reproduced from arXiv: 2607.21793 by Bin Cui, Siqi Chen, Siwei Chen, Xupeng Miao.

Figure 1
Figure 1. Figure 1: Conceptual illustration of the optimiza [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Training dynamics of different methods for Qwen2.5-32B on DAPO-MATH. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the QLPO training pipeline. QLPO first generates [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison with representative compression baselines. Each point shows a [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Response Length Distribution Shift Across Trajectory Correctness for DeepScaleR [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation of hyperparameters for DeepScaleR-1.5B-Preview on DAPO-MATH. The effect of QLPO depends on both the candidate pool size K and the length pref￾erence coefficient α. We conduct an abla￾tion study on DeepScaleR-1.5B-Preview. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Response Length Distribution Shift Across Question Complexity for DeepScaleR [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Phi-3.5-mini-instruct on Eurus-2-RL-Data. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qwen2.5-3B-VL on Geo3K. B Why Can QLPO Produce Shorter Responses with Higher Accuracy? To illustrate the qualitative difference induced by QLPO, we show a representative example from real model outputs below. The two responses are produced by Qwen2.5-32B models trained with GRPO and QLPO, respectively, on the prompt: “Solve for x: 4(x + 2) − 3 = 2x + 9. Show your reasoning step by step and verify your fina… 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

295 extracted references · 9 canonical work pages

  1. [1]

    arXiv preprint arXiv:2605.30859 , year=

    DARTS: Distribution-Aware Active Rollout Trajectory Shaping for Accelerating LLM Reinforcement Learning , author=. arXiv preprint arXiv:2605.30859 , year=

    Pith/arXiv arXiv

  2. [2]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

  3. [3]

    2023 , eprint=

    DeepSpeed-Chat: Easy, Fast and Affordable RLHF Training of ChatGPT-like Models at All Scales , author=. 2023 , eprint=

    2023

  4. [4]

    2025 , eprint=

    OpenRLHF: An Easy-to-use, Scalable and High-performance RLHF Framework , author=. 2025 , eprint=

    2025

  5. [5]

    arXiv preprint arXiv:2509.18521 , year=

    April: Active partial rollouts in reinforcement learning to tame long-tail generation , author=. arXiv preprint arXiv:2509.18521 , year=

    arXiv

  6. [6]

    Proceedings of the Twentieth European Conference on Computer Systems , pages=

    Hybridflow: A flexible and efficient rlhf framework , author=. Proceedings of the Twentieth European Conference on Computer Systems , pages=

  7. [7]

    arXiv preprint arXiv:2506.06122 , year=

    Reinforcement Learning Optimization for Large-Scale Learning: An Efficient and User-Friendly Scaling Library , author=. arXiv preprint arXiv:2506.06122 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2505.24298 , year=

    AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning , author=. arXiv preprint arXiv:2505.24298 , year=

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2509.21009 , year=

    RollPacker: Mitigating Long-Tail Rollouts for Fast, Synchronous RL Post-Training , author=. arXiv preprint arXiv:2509.21009 , year=

    arXiv

  10. [10]

    arXiv preprint arXiv:2510.11345 , year=

    Part II: ROLL Flash--Accelerating RLVR and Agentic Training with Asynchrony , author=. arXiv preprint arXiv:2510.11345 , year=

    arXiv

  11. [11]

    arXiv preprint arXiv:2504.15930 , year=

    StreamRL: Scalable, Heterogeneous, and Elastic RL for LLMs with Disaggregated Stream Generation , author=. arXiv preprint arXiv:2504.15930 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2507.01663 , year=

    AsyncFlow: An Asynchronous Streaming RL Framework for Efficient LLM Post-Training , author=. arXiv preprint arXiv:2507.01663 , year=

    Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2508.18588 , year=

    History rhymes: Accelerating llm reinforcement learning with rhymerl , author=. arXiv preprint arXiv:2508.18588 , year=

    Pith/arXiv arXiv

  14. [14]

    arXiv preprint arXiv:2510.12633 , year=

    Laminar: A Scalable Asynchronous RL Post-Training Framework , author=. arXiv preprint arXiv:2510.12633 , year=

    arXiv

  15. [15]

    arXiv preprint arXiv:2501.12948 , year=

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

    Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2507.20534 , year=

    Kimi k2: Open agentic intelligence , author=. arXiv preprint arXiv:2507.20534 , year=

    Pith/arXiv arXiv

  17. [17]

    5: Scaling reinforcement learning with llms , author=

    Kimi k1. 5: Scaling reinforcement learning with llms , author=. arXiv preprint arXiv:2501.12599 , year=

    Pith/arXiv arXiv

  18. [18]

    2025 , author =

    Grok 4 , howpublished =. 2025 , author =

    2025

  19. [19]

    5-thinking: Advancing superb reasoning models with reinforcement learning , author=

    Seed1. 5-thinking: Advancing superb reasoning models with reinforcement learning , author=. arXiv preprint arXiv:2504.13914 , year=

    arXiv

  20. [20]

    arXiv preprint arXiv:2412.16720 , year=

    Openai o1 system card , author=. arXiv preprint arXiv:2412.16720 , year=

    Pith/arXiv arXiv

  21. [21]

    2025 , author =

    OpenAI o3 , howpublished =. 2025 , author =

    2025

  22. [22]

    2025 , author =

    Claude 3.7 Sonnet , howpublished =. 2025 , author =

    2025

  23. [23]

    2025 , author =

    Claude 4.5 Opus , howpublished =. 2025 , author =

    2025

  24. [24]

    2025 , author =

    Claude 4.5 Sonnet , howpublished =. 2025 , author =

    2025

  25. [25]

    2025 , author =

    Gemini 3 Pro , howpublished =. 2025 , author =

    2025

  26. [26]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  27. [27]

    5-math technical report: Toward mathematical expert model via self-improvement , author=

    Qwen2. 5-math technical report: Toward mathematical expert model via self-improvement , author=. arXiv preprint arXiv:2409.12122 , year=

    Pith/arXiv arXiv

  28. [28]

    , howpublished =

    QwQ-32B: Embracing the Power of Reinforcement Learning. , howpublished =. 2025 , author =

    2025

  29. [29]

    arXiv preprint arXiv:1707.06347 , year=

    Proximal policy optimization algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    Pith/arXiv arXiv

  31. [31]

    arXiv preprint arXiv:2503.14476 , year=

    Dapo: An open-source llm reinforcement learning system at scale , author=. arXiv preprint arXiv:2503.14476 , year=

    Pith/arXiv arXiv

  32. [32]

    Notion Blog , year=

    Deepcoder: A fully open-source 14b coder at o3-mini level , author=. Notion Blog , year=

  33. [33]

    arXiv preprint arXiv:2504.11536 , year=

    Retool: Reinforcement learning for strategic tool use in llms , author=. arXiv preprint arXiv:2504.11536 , year=

    Pith/arXiv arXiv

  34. [34]

    Yunfan Shao and Zhichao Geng and Yitao Liu and Junqi Dai and Hang Yan and Fei Yang and Li Zhe and Hujun Bao and Xipeng Qiu , title =. Sci. China Inf. Sci. , volume =. 2024 , url =. doi:10.1007/S11432-021-3536-5 , timestamp =

    work page doi:10.1007/s11432-021-3536-5 2024

  35. [35]

    Deep learning for code generation: a survey , author=. Sci. China Inf. Sci. , volume=. 2024 , publisher=

    2024

  36. [36]

    The rise and potential of large language model based agents: A survey , author=. Sci. China Inf. Sci. , volume=. 2025 , publisher=

    2025

  37. [37]

    Data Sci

    Xuanhe Zhou and Zhaoyan Sun and Guoliang Li , title =. Data Sci. Eng. , volume =. 2024 , url =. doi:10.1007/S41019-023-00235-6 , timestamp =

    doi:10.1007/s41019-023-00235-6 2024

  38. [38]

    2024 , issn =

    AI computing systems for LLMs training: a review , journal =. 2024 , issn =. doi:10.1007/s11390-024-4178-1 , url =

    work page doi:10.1007/s11390-024-4178-1 2024

  39. [39]

    Advances of Pipeline Model Parallelism for Deep Learning Training: An Overview , journal =

    Lei Guan and Dong. Advances of Pipeline Model Parallelism for Deep Learning Training: An Overview , journal =. 2024 , url =. doi:10.1007/S11390-024-3872-3 , timestamp =

    work page doi:10.1007/s11390-024-3872-3 2024

  40. [40]

    Nagrecha, Kabir and Kumar, Arun , title =. Proc. VLDB Endow. , month =. 2024 , issue_date =. doi:10.14778/3636218.3636227 , abstract =

    arXiv 2024

  41. [41]

    Big Bird: Transformers for Longer Sequences , booktitle =

    Manzil Zaheer and Guru Guruganesh and Kumar Avinava Dubey and Joshua Ainslie and Chris Alberti and Santiago Onta. Big Bird: Transformers for Longer Sequences , booktitle =

  42. [42]

    arXiv preprint arXiv:2309.10305 , year=

    Baichuan 2: Open large-scale language models , author=. arXiv preprint arXiv:2309.10305 , year=

    Pith/arXiv arXiv

  43. [43]

    arXiv preprint arXiv:2407.21783 , year=

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    Pith/arXiv arXiv

  44. [45]

    2023 , eprint=

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. 2023 , eprint=

    2023

  45. [47]

    Proceedings of the 2022 International Conference on Management of Data , pages =

    Nakandala, Supun and Kumar, Arun , title =. Proceedings of the 2022 International Conference on Management of Data , pages =. 2022 , isbn =. doi:10.1145/3514221.3517846 , abstract =

    arXiv 2022

  46. [48]

    Proceedings of the 2021 International Conference on Management of Data , pages =

    Nagrecha, Kabir , title =. Proceedings of the 2021 International Conference on Management of Data , pages =. 2021 , isbn =. doi:10.1145/3448016.3450571 , abstract =

    arXiv 2021

  47. [49]

    Forty-first International Conference on Machine Learning,

    Hantian Ding and Zijian Wang and Giovanni Paolini and Varun Kumar and Anoop Deoras and Dan Roth and Stefano Soatto , title =. Forty-first International Conference on Machine Learning,. 2024 , url =

    2024

  48. [50]

    Ma, Kaihao and Yan, Xiao and Cai, Zhenkun and Huang, Yuzhen and Wu, Yidi and Cheng, James , title =. Proc. ACM Manag. Data , month =. 2023 , issue_date =. doi:10.1145/3589310 , abstract =

    work page doi:10.1145/3589310 2023

  49. [51]

    Fang, Jingzhi and Shen, Yanyan and Wang, Yue and Chen, Lei , title =. Proc. ACM Manag. Data , month =. 2024 , issue_date =. doi:10.1145/3639323 , abstract =

    work page doi:10.1145/3639323 2024

  50. [52]

    Proceedings of the 2022 International Conference on Management of Data , pages =

    Wang, Qiange and Zhang, Yanfeng and Wang, Hao and Chen, Chaoyi and Zhang, Xiaodong and Yu, Ge , title =. Proceedings of the 2022 International Conference on Management of Data , pages =. 2022 , isbn =. doi:10.1145/3514221.3526134 , abstract =

    arXiv 2022

  51. [53]

    Wang, Qiange and Chen, Yao and Wong, Weng-Fai and He, Bingsheng , title =. Proc. ACM Manag. Data , month =. 2023 , issue_date =. doi:10.1145/3626733 , abstract =

    work page doi:10.1145/3626733 2023

  52. [54]

    Wan, Xinchen and Xu, Kaiqiang and Liao, Xudong and Jin, Yilun and Chen, Kai and Jin, Xin , title =. Proc. ACM Manag. Data , month =. 2023 , issue_date =. doi:10.1145/3589288 , abstract =

    doi:10.1145/3589288 2023

  53. [55]

    2009 , isbn =

    Drozdowski, Maciej , title =. 2009 , isbn =

    2009

  54. [56]

    Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 , pages=

    Heet: Accelerating Elastic Training in Heterogeneous Deep Learning Clusters , author=. Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 , pages=

  55. [57]

    Proceedings of the Eighteenth European Conference on Computer Systems , pages=

    Lyra: Elastic scheduling for deep learning clusters , author=. Proceedings of the Eighteenth European Conference on Computer Systems , pages=

  56. [58]

    21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24) , pages=

    \ DISTMM \ : Accelerating Distributed Multimodal Model Training , author=. 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24) , pages=

  57. [59]

    CoRR , volume =

    Jinze Bai and Shuai Bai and Yunfei Chu and Zeyu Cui and Kai Dang and Xiaodong Deng and Yang Fan and Wenbin Ge and Yu Han and Fei Huang and Binyuan Hui and Luo Ji and Mei Li and Junyang Lin and Runji Lin and Dayiheng Liu and Gao Liu and Chengqiang Lu and Keming Lu and Jianxin Ma and Rui Men and Xingzhang Ren and Xuancheng Ren and Chuanqi Tan and Sinan Tan ...

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.16609 2023

  58. [60]

    Gehringer and Daniel P

    Edward F. Gehringer and Daniel P. Siewiorek and Zary Segall. Parallel Processing: The Cm* Experience. Parallel Processing: The Cm* Experience. 1987

    1987

  59. [61]

    1992 , isbn =

    Hwang, Kai , title =. 1992 , isbn =

    1992

  60. [62]

    Kovalyov and Maciej Machowiak

    Jacek Blazewicz and Mikhail Y. Kovalyov and Maciej Machowiak. and Denis Trystram and Jan Weglarz , journal=. Preemptable malleable task scheduling problem , year=

  61. [63]

    Annals of Operations Research , year=

    Scheduling Malleable Tasks on Parallel Processors to Minimize the Makespan , author=. Annals of Operations Research , year=

  62. [64]

    Optimization and Control of Dynamic Operational Research Models , pages=

    Modelling and control of dynamic resource allocation project scheduling systems , author=. Optimization and Control of Dynamic Operational Research Models , pages=. 1982 , publisher=

    1982

  63. [65]

    Weglarz, Jan , title =. Manage. Sci. , month =. 1981 , issue_date =. doi:10.1287/mnsc.27.9.1040 , abstract =

    work page doi:10.1287/mnsc.27.9.1040 1981

  64. [66]

    Gomez and Lukasz Kaiser and Illia Polosukhin , title =

    Ashish Vaswani and Noam Shazeer and Niki Parmar and Jakob Uszkoreit and Llion Jones and Aidan N. Gomez and Lukasz Kaiser and Illia Polosukhin , title =. NeurIPS , pages =. 2017 , timestamp =

    2017

  65. [67]

    wav2vec 2.0:

    Alexei Baevski and Yuhao Zhou and Abdelrahman Mohamed and Michael Auli , editor =. wav2vec 2.0:. Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual , year =

    2020

  66. [68]

    Anurag Arnab and Mostafa Dehghani and Georg Heigold and Chen Sun and Mario Lucic and Cordelia Schmid , title =. 2021. 2021 , url =. doi:10.1109/ICCV48922.2021.00676 , timestamp =

    arXiv 2021

  67. [69]

    VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training , booktitle =

    Zhan Tong and Yibing Song and Jue Wang and Limin Wang , editor =. VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training , booktitle =. 2022 , url =

    2022

  68. [70]

    Large-Scale Self- and Semi-Supervised Learning for Speech Translation , booktitle =

    Changhan Wang and Anne Wu and Juan Pino and Alexei Baevski and Michael Auli and Alexis Conneau , editor =. Large-Scale Self- and Semi-Supervised Learning for Speech Translation , booktitle =. 2021 , url =. doi:10.21437/INTERSPEECH.2021-1912 , timestamp =

    work page doi:10.21437/interspeech.2021-1912 2021

  69. [71]

    Robust Speech Recognition via Large-Scale Weak Supervision , booktitle =

    Alec Radford and Jong Wook Kim and Tao Xu and Greg Brockman and Christine McLeavey and Ilya Sutskever , editor =. Robust Speech Recognition via Large-Scale Weak Supervision , booktitle =. 2023 , url =

    2023

  70. [72]

    The Tenth International Conference on Learning Representations,

    Hangbo Bao and Li Dong and Songhao Piao and Furu Wei , title =. The Tenth International Conference on Learning Representations,. 2022 , url =

    2022

  71. [73]

    Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis , pages=

    Efficient large-scale language model training on gpu clusters using megatron-lm , author=. Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis , pages=

  72. [74]

    Optimization Methods for Large-Scale Machine Learning , journal =

    L. Optimization Methods for Large-Scale Machine Learning , journal =. 2018 , timestamp =

    2018

  73. [75]

    ICML , volume =

    Xiangru Lian and Wei Zhang and Ce Zhang and Ji Liu , title =. ICML , volume =. 2018 , timestamp =

    2018

  74. [76]

    ICCC , pages=

    An Ensemble Learning Algorithm for Indoor Localization , author=. ICCC , pages=

  75. [77]

    Mechanical Systems and Signal Processing , volume=

    Rolling bearing fault feature learning using improved convolutional deep belief network with compressed sensing , author=. Mechanical Systems and Signal Processing , volume=

  76. [78]

    NeurIPS , pages=

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results , author=. NeurIPS , pages=

  77. [79]

    ICML , pages=

    Follow the moving leader in deep learning , author=. ICML , pages=

  78. [80]

    Andersen and Jun Woo Park and Alexander J

    Mu Li and David G. Andersen and Jun Woo Park and Alexander J. Smola and Amr Ahmed and Vanja Josifovski and James Long and Eugene J. Shekita and Bor. Scaling Distributed Machine Learning with the Parameter Server , booktitle =. 2014 , timestamp =

    2014

  79. [81]

    Taming unbalanced training workloads in deep learning with partial collective operations , booktitle =

    Shigang Li and Tal Ben. Taming unbalanced training workloads in deep learning with partial collective operations , booktitle =. 2020 , timestamp =

    2020

  80. [82]

    ICLR , year =

    Karen Simonyan and Andrew Zisserman , title =. ICLR , year =

Showing first 80 references.