Pith. sign in

REVIEW 3 major objections 5 minor 38 references

REFLEX reallocates expert budgets in diffusion language models by token refinement state, cutting allocated expert computation by 15% while keeping or improving quality on most math and code benchmarks.

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 ↗

A training-free, refinement-aware expert budget allocation scheme (REFLEX) reduces selected expert-token pairs by about 15% on MoE diffusion language models without hurting benchmark quality.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Useful heuristic with solid ablations; the sign of the frontier-progress signal is under-tested and the numbers lack error bars, but it deserves a real referee. the 3 major comments →

arxiv 2608.01784 v1 pith:NE75VE5J submitted 2026-08-03 cs.AI cs.CL

REFLEX: Rethinking MoE Inference as Refinement-Aware Compute Allocation in Diffusion Language Models

Xiang Xia , Cheng Yan , Yiming Zhang , Jiazheng Liu , Hongyu Zhang , Wuyang Zhang This is my paper
classification cs.AI cs.CL
keywords mixture-of-expertsdiffusion language modelsexpert allocationtoken-choice routinginference efficiencyblock-wise decodingconfidence progresstraining-free acceleration
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

REFLEX claims that Mixture-of-Experts inference in diffusion language models is best treated as a compute-allocation problem: each denoising forward revisits tokens in very different refinement states, yet the default router gives every token the same expert budget. The paper proposes a training-free allocator that leaves the router's expert ranking untouched but varies how many experts each token gets, guided first by how far a token's block is from the active block and then by a Frontier-Progress Score built from confidence history. On LLaDA-MoE and LLaDA2.0-mini across math and code benchmarks, this reduces allocated expert computation by about 15% on average while preserving or improving generation quality on most tasks, and gives a more consistent quality-computation trade-off than autoregressive-style variable-expert methods. If correct, it shows MoE efficiency in diffusion LMs does not require retraining or router changes—only smarter per-token expert budgeting.

Core claim

The paper's central claim is that expert-count allocation in MoE-based diffusion language models should be decoupled from expert ranking and conditioned on two observable refinement-state dimensions: a token's block-relative structural role in the iterative denoising process, and its dynamic progress toward commitment. REFLEX instantiates this via a coarse-to-fine hierarchy: Refinement-Role Budgeting assigns higher expert counts to active and adjacent blocks and lower counts to distant blocks using a three-level profile, and Frontier-Guided Expert Reallocation then redistributes the active block's total budget among priority groups defined by the Frontier-Progress Score. FPS combines commitm

What carries the argument

The load-bearing object is the Frontier-Progress Score S_t_i = F_t_i * P_t_i, where F_t_i is a sigmoid gate measuring whether a token's extrapolated confidence lies in the commitment-frontier interval [θ−w_cf, θ] and P_t_i measures whether its recent confidence change is below a velocity reference (limited progress). It is used to rank unresolved active-block tokens into equal-sized high-, middle-, and low-priority groups that receive K_cur+δ_cur, K_cur, and K_cur−δ_cur experts, so the block's total expert budget is preserved. The structural prior comes from Refinement-Role Budgeting's profile H(d), which sets expert counts by block-relative distance d to the active block (K_base+δ_role for

Load-bearing premise

The load-bearing premise is that the Frontier-Progress Score, estimated from a one-step extrapolation of past confidence, marks the tokens whose next denoising pass will actually benefit from more experts, and that taking experts away from low-scoring tokens will not hurt them—something validated only on the benchmarks tested, not derived from the model.

What would settle it

Apply REFLEX's active-block reallocation with the FPS ranking reversed—give the extra experts to the tokens the FPS calls lowest priority. If quality on the same math and code benchmarks does not drop below the default FPS ordering at the same AvgK, then the Frontier-Progress Score is not the causal signal and the reported gains would come from the block-relative budget profile alone.

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

If this is right

  • MoE-based DLMs can shed about 15% of allocated expert computation at inference time with no retraining and no change to the router.
  • Quality is preserved or improved on most math and code benchmarks; the notable exception is LLaDA2.0-mini on GSM8K, which loses 1.5 points while saving 10% of pairs.
  • Uniformly lowering the expert budget (top-4) hurts every task, so the savings come from where experts are reallocated, not from the lower average alone.
  • The active-block reallocation is budget-neutral, meaning the computation reduction is driven by the block-relative structural profile (RRB), while the FPS signal improves where within the active block the budget is spent.
  • Router-score-based and global-timestep baselines land off the empirical Pareto frontier, supporting refinement state rather than router preference or global denoising progress as the right allocation criterion.

Where Pith is reading between the lines

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

  • A natural test is whether the same coarse-to-fine allocation transfers to larger MoE-based DLMs and to non-block-wise DLM decoding schemes; the block-relative prior is specific to block-wise inference and would need reformulation for fully parallel decoding.
  • The confidence-history forecast is only first-order; a similar allocator could exploit richer uncertainty signals (e.g., prediction variance or attention entropy) to rank active-block tokens, potentially tightening the frontier-progress signal without changing the hierarchy.
  • Because REFLEX only changes expert counts, it can be stacked with expert offloading and unique-expert reduction methods; the combined gain in end-to-end latency could be larger than either alone, though the paper only reports allocated expert-token pairs.
  • The same refinement-aware logic may apply to masked image diffusion transformers with MoE layers, where block-relative roles could map to spatial regions and confidence to mask-scheduling priorities.
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. The paper proposes REFLEX, a training-free method for reallocating expert computation in MoE-based diffusion language models (DLMs). It decomposes MoE inference into router ranking (preserved) and expert-count allocation, viewing allocation as refinement-aware. REFLEX has two levels: Refinement-Role Budgeting (RRB), which assigns expert counts from the token's block-relative distance to the active block (Eq. 5), and Frontier-Guided Expert Reallocation (FGER), which reallocates a fixed active-block budget using the Frontier-Progress Score (FPS), a heuristic based on forecast confidence and confidence-velocity (Eqs. 6-9). The paper reports that on four benchmarks across LLaDA-MoE and LLaDA2.0-mini, REFLEX reduces selected expert-token pairs by about 15% on average while preserving or improving accuracy relative to Fixed TC top-8. It also includes component ablations, a role-shuffled RRB control, matched-computation signal comparisons, and hyperparameter sensitivity analyses.

Significance. If the empirical claim holds, REFLEX is a useful, training-free inference optimization for MoE DLMs. The paper has notable strengths: the role-shuffled RRB control (Table A) directly tests whether block-relative assignment matters beyond the count distribution, and the matched-computation comparison in Table B is a sound way to compare ranking signals while holding AvgK fixed. The paper is also candid that Eq. (4) is an allocation principle rather than an executable utility estimator. However, the load-bearing causal claim—that FPS identifies tokens for which extra experts are useful and that low-FPS tokens can safely lose experts—is not directly tested. The main results also lack error bars, and a model-specific hyperparameter is selected on the evaluation tasks. These issues are fixable with additional experiments and reporting, so the paper is a credible contribution pending revision.

major comments (3)
  1. [Appendix C, Table B; Methodology Eq. (9)] There is no directional (sign) control for the FGER allocation. Table B compares FPS against Confidence Only, Uncertainty Only, Frontier Only, and Progress Only, but every configuration allocates more experts to high-signal tokens and fewer to low-signal tokens (Eq. 9). This leaves open the possibility that any non-uniform reallocation of the active-block budget—not the specific FPS ordering—drives the quality gains. To support the causal claim that low-FPS tokens tolerate expert cuts, the authors should add a reversed or random-sign allocation at matched computation (e.g., giving +delta to low-FPS and -delta to high-FPS tokens). Without such a control, the +2.6/+1.9 point gains on Math500 and HumanEval could be explained by merely breaking the uniformity of top-8.
  2. [Table 1 and Table 2] The central quality-preservation claim is not accompanied by any variance estimate. Many reported differences are small: +2.6, +1.9, and -1.5 accuracy points, with a single deterministic run per configuration. On benchmarks known to have high prompt/seed variance, these differences are within plausible noise. The authors should report standard deviations or confidence intervals over multiple seeds/runs for the main results and component ablations, not only for the role-shuffled control. This is necessary to support the claim that REFLEX 'preserves or even improves' generation quality rather than merely not degrading it by a measurable amount.
  3. [Experimental Setting; Appendix D, Table C] The model-specific commitment-frontier width w_cf is selected on the evaluation tasks: w_cf=0.30 for LLaDA-MoE and w_cf=0.25 for LLaDA2.0-mini. Appendix D shows that on LLaDA-MoE, HumanEval accuracy varies from 51.83 to 57.32 across w_cf values, with the best value at 0.35 rather than the default 0.30. Because w_cf changes which tokens receive additional experts, this is not a cosmetic tuning choice. The authors need to clarify whether w_cf was chosen on a held-out validation split or by a pre-specified rule; otherwise part of the reported gains may reflect selection on the test benchmarks. Reporting the sensitivity tables is helpful, but it does not remove the selection concern.
minor comments (5)
  1. [Main Results, Section on Q1] The text reports that LLaDA2.0-mini GSM8K has NFE increasing by 11.6%, but NFE values are not tabulated. Since selected pairs depend on AvgK and NFE, reporting per-benchmark NFE (or pass counts) would help readers decompose the computation reduction.
  2. [Figure 1] The figure contains the annotation '3+3+3=4+3+2' in the RRB panel; the notation is confusing because the terms refer to expert counts for different roles, not an arithmetic identity. Please clarify with labeled K values for prompt/near/active/far blocks.
  3. [Metrics paragraph, Section Experimental Setup] There is a typo: 'Its normalized average is AvgK,K= ...' should be 'AvgK = ' or 'AvgK = pairs/(|L_moe| ...)'.
  4. [Appendix C, Table A] Role-Shuffled RRB reports mean and standard deviation over three seeds, but RRB Only is a single point. A confidence interval for RRB Only (or more seeds for both) would make the 3.7 and 8.8 point gaps more interpretable.
  5. [Discussion] The paper correctly notes that selected expert-token pairs are not end-to-end latency. It may be useful to state explicitly whether any of the compared methods affect attention or dispatch overhead differently, since the comparison is focused on one FLOP proxy.

Circularity Check

0 steps flagged

No significant circularity: REFLEX is an empirical allocation heuristic whose compute reduction is a transparent consequence of its own RRB profile, and whose quality claims rest on benchmark measurements and ablations rather than on a derivation from its inputs.

full rationale

The paper does not present a first-principles derivation of its quality gains. The formal allocation objective in Eq. (4) is explicitly acknowledged as an idealization: 'Since U_i^t is unobservable, Eq. (4) defines an allocation principle rather than an executable utility estimator.' The implemented rule in Eqs. (5)-(9) is a hand-designed proxy, and the paper consistently says FPS is used 'to rank active-block tokens from dynamic evidence, not to estimate marginal utility in Eq. (4).' The compute reduction is a direct arithmetic consequence of the RRB profile H(d) in Eq. (5): assigning Kbase-δ_role to d≥3 necessarily lowers AvgK relative to uniform Kbase, and Appendix A states transparently that 'RRB establishes the expert-count profile ... directly determines the reduction in AvgK.' This is method design, not a hidden equivalence between a prediction and an input. The quality-preservation claim is tested on standard benchmarks with component ablations, a role-shuffled intervention, and matched-computation signal comparisons (Tables A/B, 2); no equation is shown to reduce to its own inputs. Hyperparameter selection and the absence of a reversed-FPS directional control are real validity/robustness concerns about the heuristic's causal assumption, but they are not circularity under the required standard: no fitted value is renamed a prediction, no load-bearing self-citation chain exists, and no uniqueness theorem is imported. Therefore no circular step can be exhibited.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 0 invented entities

REFLEX introduces no new physical or structural entities. Its new constructs are algorithmic: a block-relative budgeting profile and a Frontier-Progress Score. The main contributions come from hand-chosen hyperparameters and domain assumptions about how confidence and block-relative roles relate to refinement demand.

free parameters (8)
  • Kbase = 8
    Base selected expert count; sets the nominal budget in Eq. (5). Chosen by authors, not derived.
  • delta_role = 4
    Role offset in Refinement-Role Budgeting; differentiates d<=1, d=2, d>=3 block-relative roles in Eq. (5).
  • delta_cur = 2
    Active-block reallocation offset in Eq. (9); controls how many experts are moved between high- and low-priority groups.
  • q = 0.25
    Outer-group fraction in FGER; sets the proportion of active-block tokens in the high- and low-priority groups.
  • v_ref = 0.01
    Velocity reference in Eq. (7); below this confidence progress is treated as slow.
  • tau = 0.05
    Gate temperature in Eqs. (6)-(7); controls smoothness of the frontier and progress factors.
  • w_cf = 0.30 for LLaDA-MoE; 0.25 for LLaDA2.0-mini
    Commitment-frontier width; differs per model, suggesting selection on the evaluation benchmarks. Sensitivity is reported in Appendix D.
  • Kmin, Kmax bounds = unspecified
    The per-token expert count bounds in Eq. (3) are defined but never given numerically, leaving the feasible allocation set under-specified.
axioms (5)
  • domain assumption Block-relative distance to the active block is a sufficient coarse signal for structural refinement role.
    Eq. (5) maps d(i) <= 1, d(i)=2, d(i)>=3 to distinct expert budgets. This is a modeling choice, not a derived property of DLMs.
  • domain assumption Confidence history can be extrapolated one forward to forecast commitment proximity.
    Eq. (6) uses ct-1 and ct-2 with first-order extrapolation and clipping to predict ct; this assumes the short-term confidence trend is informative.
  • domain assumption The default router's expert ranking remains valid when selected expert counts vary.
    REFLEX keeps router rankings and only changes counts; this assumes top-k preference order does not need retraining when k changes.
  • domain assumption Selected expert-token pairs are a valid proxy for allocated expert compute.
    Appendix A equates relative changes in selected pairs to routed-expert linear-layer FLOPs, explicitly excluding attention, router scoring, shared experts, and dispatch overhead.
  • domain assumption An additive state-wise utility surrogate with diminishing returns exists for expert allocation.
    Eq. (4) defines the ideal allocation as maximizing a sum of unobservable utilities U_i(k); the paper acknowledges this is a principle, not an executable estimator.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of REFLEX: Rethinking MoE Inference as Refinement-Aware Compute Allocation in Diffusion Language Models." pith.science (2026). https://pith.science/paper/NE75VE5J

@misc{pith2026260801784,
  author       = {Pith},
  title        = {Pith review of: REFLEX: Rethinking MoE Inference as Refinement-Aware Compute Allocation in Diffusion Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NE75VE5J}},
  note         = {Machine review of arXiv:2608.01784}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Mixture-of-experts (MoE) models increase parameter capacity by activating only a small subset of experts for each token. This conditional-computation paradigm has enabled autoregressive language models to scale model capacity without a proportional increase in per-token computation. In diffusion language models (DLMs), however, each denoising forward jointly revisits all token positions despite their sharply different refinement demands, while the default fixed token-choice routing assigns them a uniform expert budget, creating a mismatch between expert computation and refinement demand. We argue that MoE inference in DLMs should therefore be viewed as refinement-aware compute allocation across heterogeneous token refinement states. We propose REFLEX (\textbf{RE}finement-aware \textbf{FLEX}ible expert allocation), a training-free method that keeps the default router unchanged while reorganizing expert computation around the evolving refinement process. Specifically, REFLEX introduces a coarse-to-fine hierarchy for expert-budget allocation that aligns computation with block-relative refinement roles while using the Frontier-Progress Score to resolve active-block priorities. Across multiple widely used benchmarks on two representative MoE-based DLMs, LLaDA-MoE and LLaDA2.0-mini, REFLEX reduces allocated expert computation by 15\% on average while preserving or even improving generation quality on most benchmarks relative to default routing. Compared with autoregressive-style variable-expert routing methods, REFLEX also yields a more consistent quality--computation trade-off, further supporting the importance of allocating expert computation according to the heterogeneous refinement demands exposed within each denoising forward.

Figures

Figures reproduced from arXiv: 2608.01784 by Cheng Yan, Hongyu Zhang, Jiazheng Liu, Wuyang Zhang, Xiang Xia, Yiming Zhang.

Figure 1
Figure 1. Figure 1: Overview of REFLEX. (a) Fixed token-choice routing assigns the same expert count as token refinement states evolve [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Cross-task quality–computation trade-off on [PITH_FULL_IMAGE:figures/full_fig_p006_2.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 · 35 canonical work pages

  1. [1]

    Le and Geoffrey E

    Noam Shazeer and Azalia Mirhoseini and Krzysztof Maziarz and Andy Davis and Quoc V. Le and Geoffrey E. Hinton and Jeff Dean , title =. Proceedings of the 5th International Conference on Learning Representations , address =

    work page

  2. [2]

    Proceedings of the 9th International Conference on Learning Representations , address =

    Dmitry Lepikhin and HyoukJoong Lee and Yuanzhong Xu and Dehao Chen and Orhan Firat and Yanping Huang and Maxim Krikun and Noam Shazeer and Zhifeng Chen , title =. Proceedings of the 9th International Conference on Learning Representations , address =

    work page

  3. [3]

    Journal of Machine Learning Research , volume =

    William Fedus and Barret Zoph and Noam Shazeer , title =. Journal of Machine Learning Research , volume =

    work page

  4. [4]

    Albert Q. Jiang and Alexandre Sablayrolles and Antoine Roux and Arthur Mensch and Blanche Savary and Chris Bamford and Devendra Singh Chaplot and Diego de Las Casas and Emma Bou Hanna and Florian Bressand and Gianna Lengyel and Guillaume Bour and Guillaume Lample and L. CoRR , volume =

    work page

  5. [5]

    Damai Dai and Chengqi Deng and Chenggang Zhao and R. X. Xu and Huazuo Gao and Deli Chen and Jiashi Li and Wangding Zeng and Xingkai Yu and Y. Wu and Zhenda Xie and Y. K. Li and Panpan Huang and Fuli Luo and Chong Ruan and Zhifang Sui and Wenfeng Liang , title =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics , pages =

    work page

  6. [6]

    Johnson and Jonathan Ho and Daniel Tarlow and Rianne van den Berg , title =

    Jacob Austin and Daniel D. Johnson and Jonathan Ho and Daniel Tarlow and Rianne van den Berg , title =. Advances in Neural Information Processing Systems 34 , pages =

    work page

  7. [7]

    Hashimoto , title =

    Xiang Lisa Li and John Thickstun and Ishaan Gulrajani and Percy Liang and Tatsunori B. Hashimoto , title =. Advances in Neural Information Processing Systems 35 , pages =

    work page

  8. [8]

    Chiu and Zhihan Yang and Zhixuan Qi and Jiaqi Han and Subham Sekhar Sahoo and Volodymyr Kuleshov , title =

    Marianne Arriola and Aaron Gokaslan and Justin T. Chiu and Zhihan Yang and Zhixuan Qi and Jiaqi Han and Subham Sekhar Sahoo and Volodymyr Kuleshov , title =. Proceedings of the 13th International Conference on Learning Representations , address =

    work page

  9. [9]

    Freeman , title =

    Huiwen Chang and Han Zhang and Lu Jiang and Ce Liu and William T. Freeman , title =. Proceedings of the 35th

    work page

  10. [10]

    Large Language Diffusion Models , booktitle =

    Shen Nie and Fengqi Zhu and Zebin You and Xiaolu Zhang and Jingyang Ou and Jun Hu and Jun Zhou and Yankai Lin and Ji. Large Language Diffusion Models , booktitle =

    work page

  11. [11]

    CoRR , volume =

    Jiacheng Ye and Zhihui Xie and Lin Zheng and Jiahui Gao and Zirui Wu and Xin Jiang and Zhenguo Li and Lingpeng Kong , title =. CoRR , volume =

    work page

  12. [12]

    Findings of the 64th Annual Meeting of the Association for Computational Linguistics , pages =

    Shuang Cheng and Yihan Bian and Dawei Liu and Yuhua Jiang and Yihao Liu and Linfeng Zhang and Qian Yao and Zhongbo Tian and Wenhai Wang and Qipeng Guo and Kai Chen and Biqing Qi and Bowen Zhou , title =. Findings of the 64th Annual Meeting of the Association for Computational Linguistics , pages =

    work page

  13. [13]

    CoRR , volume =

    Fengqi Zhu and Zebin You and Yipeng Xing and Zenan Huang and Lin Liu and Yihong Zhuang and Guoshan Lu and Kangyu Wang and Xudong Wang and Lanning Wei and Hongrui Guo and Jiaqi Hu and Wentao Ye and Tieyuan Chen and Chenchen Li and Chengfu Tang and Haibo Feng and Jun Hu and Jun Zhou and Xiaolu Zhang and Zhenzhong Lan and Junbo Zhao and Da Zheng and Chongxua...

    work page

  14. [14]

    CoRR , volume =

    Tiwei Bie and Maosong Cao and Kun Chen and Lun Du and Mingliang Gong and Zhuochen Gong and Yanmei Gu and Jiaqi Hu and Zenan Huang and Zhenzhong Lan and Chengxi Li and Chongxuan Li and Jianguo Li and Zehuan Li and Huabin Liu and Lin Liu and Guoshan Lu and Xiaocheng Lu and Yuxin Ma and Jianfeng Tan and Lanning Wei and Ji. CoRR , volume =

    work page

  15. [15]

    CoRR , volume =

    Yuxuan Song and Zheng Zhang and Cheng Luo and Pengyang Gao and Fan Xia and Hao Luo and Zheng Li and Yuehang Yang and Hongli Yu and Xingwei Qu and Yuwei Fu and Jing Su and Ge Zhang and Wenhao Huang and Mingxuan Wang and Lin Yan and Xiaoying Jia and Jingjing Liu and Wei. CoRR , volume =

    work page

  16. [16]

    Proceedings of the 14th International Conference on Learning Representations , address =

    Chengyue Wu and Hao Zhang and Shuchen Xue and Zhijian Liu and Shizhe Diao and Ligeng Zhu and Ping Luo and Song Han and Enze Xie , title =. Proceedings of the 14th International Conference on Learning Representations , address =

    work page

  17. [17]

    CoRR , volume =

    Karl Cobbe and Vineet Kosaraju and Mohammad Bavarian and Mark Chen and Heewoo Jun and Lukasz Kaiser and Matthias Plappert and Jerry Tworek and Jacob Hilton and Reiichiro Nakano and Christopher Hesse and John Schulman , title =. CoRR , volume =

  18. [18]

    Proceedings of the 12th International Conference on Learning Representations , address =

    Hunter Lightman and Vineet Kosaraju and Yuri Burda and Harrison Edwards and Bowen Baker and Teddy Lee and Jan Leike and John Schulman and Ilya Sutskever and Karl Cobbe , title =. Proceedings of the 12th International Conference on Learning Representations , address =

    work page

  19. [19]

    Evaluating Large Language Models Trained on Code , journal =

    Mark Chen and Jerry Tworek and Heewoo Jun and Qiming Yuan and Henrique Pond. Evaluating Large Language Models Trained on Code , journal =

  20. [20]

    Nye and Maarten Bosma and Henryk Michalewski and David Dohan and Ellen Jiang and Carrie J

    Jacob Austin and Augustus Odena and Maxwell I. Nye and Maarten Bosma and Henryk Michalewski and David Dohan and Ellen Jiang and Carrie J. Cai and Michael Terry and Quoc V. Le and Charles Sutton , title =. CoRR , volume =

    work page

  21. [21]

    CoRR , volume =

    Alex Graves , title =. CoRR , volume =

    work page

  22. [22]

    Advances in Neural Information Processing Systems 35 , address =

    Tal Schuster and Adam Fisch and Jai Gupta and Mostafa Dehghani and Dara Bahri and Vinh Tran and Yi Tay and Donald Metzler , title =. Advances in Neural Information Processing Systems 35 , address =

    work page

  23. [23]

    CoRR , volume =

    Tiansheng Wen and Yifei Wang and Aosong Feng and Long Ma and Xinyang Liu and Yifan Wang and Lixuan Guo and Bo Chen and Stefanie Jegelka and Chenyu You , title =. CoRR , volume =

    work page

  24. [24]

    Metaxas , title =

    Can Jin and Hongwu Peng and Mingcan Xiang and Qixin Zhang and Xiangchi Yuan and Amit Hasan and Ohi Dibua and Yifan Gong and Yan Kang and Dimitris N. Metaxas , title =. Proceedings of the 43rd International Conference on Machine Learning , address =

    work page

  25. [25]

    CoRR , volume =

    Sicheng Feng and Zigeng Chen and Gongfan Fang and Xinyin Ma and Xinchao Wang , title =. CoRR , volume =

    work page

  26. [26]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics , pages =

    Quzhe Huang and Zhenwei An and Nan Zhuang and Mingxu Tao and Chen Zhang and Yang Jin and Kun Xu and Liwei Chen and Songfang Huang and Yansong Feng , title =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics , pages =

    work page

  27. [27]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics , pages =

    Baihui Liu and Kaiyuan Tian and Wei Wang and Zhaoning Zhang and Linbo Qiao and Dongsheng Li , title =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics , pages =

    work page

  28. [28]

    CoRR , volume =

    Hanchi Sun and Yixin Liu and Yonghui Wu and Lichao Sun , title =. CoRR , volume =

    work page

  29. [29]

    CoRR , volume =

    Chengcheng Wang and Tingzhang Luo and Wenhao Li and Jianyuan Guo and Chang Xu , title =. CoRR , volume =

    work page

  30. [30]

    Proceedings of the 43rd International Conference on Machine Learning , address =

    Linye Wei and Zixiang Luo and Pingzhi Tang and Meng Li , title =. Proceedings of the 43rd International Conference on Machine Learning , address =

    work page

  31. [31]

    CoRR , volume =

    Zhiben Chen and Youpeng Zhao and Yang Sui and Jun Wang and Yuzhang Shang , title =. CoRR , volume =

    work page

  32. [32]

    CoRR , volume =

    Shuibai Zhang and Caspian Zhuang and Chihan Cui and Zhihan Yang and Fred Zhangzhi Peng and Yanxin Zhang and Haoyue Bai and Zack Jia and Yang Zhou and Guanhua Chen and Ming Liu , title =. CoRR , volume =

    work page

  33. [33]

    Tom B. Brown and Benjamin Mann and Nick Ryder and Melanie Subbiah and Jared Kaplan and Prafulla Dhariwal and Arvind Neelakantan and Pranav Shyam and Girish Sastry and Amanda Askell and Sandhini Agarwal and Ariel Herbert. Language Models are Few-Shot Learners , booktitle =

  34. [34]

    Dai and Simon Tong and Dmitry Lepikhin and Yuanzhong Xu and Maxim Krikun and Yanqi Zhou and Adams Wei Yu and Orhan Firat and Barret Zoph and Liam Fedus and Maarten P

    Nan Du and Yanping Huang and Andrew M. Dai and Simon Tong and Dmitry Lepikhin and Yuanzhong Xu and Maxim Krikun and Yanqi Zhou and Adams Wei Yu and Orhan Firat and Barret Zoph and Liam Fedus and Maarten P. Bosma and Zongwei Zhou and Tao Wang and Yu Emma Wang and Kellie Webster and Marie Pellat and Kevin Robinson and Kathleen S. Meier. Proceedings of the 3...

    work page

  35. [35]

    Advances in Neural Information Processing Systems 33 , pages =

    Jonathan Ho and Ajay Jain and Pieter Abbeel , title =. Advances in Neural Information Processing Systems 33 , pages =

    work page

  36. [36]

    Proceedings of the 41st International Conference on Machine Learning , pages =

    Aaron Lou and Chenlin Meng and Stefano Ermon , title =. Proceedings of the 41st International Conference on Machine Learning , pages =

    work page

  37. [37]

    Sahoo and Marianne Arriola and Yair Schiff and Aaron Gokaslan and Edgar Marroquin and Justin T

    Subham S. Sahoo and Marianne Arriola and Yair Schiff and Aaron Gokaslan and Edgar Marroquin and Justin T. Chiu and Alexander Rush and Volodymyr Kuleshov , title =. Advances in Neural Information Processing Systems 37 , address =

    work page

  38. [38]

    Proceedings of the 13th International Conference on Learning Representations , address =

    Shen Nie and Fengqi Zhu and Chao Du and Tianyu Pang and Qian Liu and Guangtao Zeng and Min Lin and Chongxuan Li , title =. Proceedings of the 13th International Conference on Learning Representations , address =

    work page

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.