Pith. sign in

REVIEW 4 major objections 3 minor 30 references

A lightweight causal refiner, resolved in a few parallel Jacobi iterations, restores the missing token-to-token dependencies in one-pass diffusion drafters and raises acceptance length by about 30% and end-to-end decoding throughput by abou

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 ↗

xPress restores causality to diffusion-drafter drafts with a lightweight refiner solved by a few parallel Jacobi iterations, lifting acceptance length ~30% and throughput ~1.3x over dFlash on Qwen3-8B.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A genuinely new method for restoring causality in diffusion drafters, with internally consistent gains but an unverified top-k premise and missing artifacts. the 4 major comments →

arxiv 2608.02438 v1 pith:NAT62ADI submitted 2026-08-03 cs.AI

xPress: Parallel Refinement for Diffusion Drafters in Speculative Decoding

Zheng Wang , Davis Wertheimer , Yu Chin Fabian Lim , Mudhakar Srivatsa , Raghu K. Ganti , Minjia Zhang , Naigang Wang This is my paper
classification cs.AI
keywords speculative decodingdiffusion drafterscausal refinementJacobi decodingacceptance lengthlow-rank mixerlogit refinementparallel decoding
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

Speculative decoding accelerates language models by having a cheap draft model propose several tokens that the large model verifies at once. Diffusion-based drafters such as dFlash can propose an entire block in a single forward pass, but they sample each position independently, so the draft lacks the token-to-token dependence the verifying model expects, and early rejections limit the speedup. xPress adds a small causal refiner that re-ranks the drafter's already-plausible candidates using a low-rank triangular mixer, and it resolves the resulting causal dependency with a few parallel fixed-point (Jacobi) iterations instead of a serial left-to-right sweep. Training includes a total-variation loss that directly targets the acceptance rate. On Qwen3-8B across seven benchmarks, xPress raises acceptance length by about 30% (up to 56%) and end-to-end decoding throughput by about 1.3x (up to 1.7x) relative to the original diffusion drafter, with the improvement persisting in batched serving.

Core claim

The central claim is that the poor acceptance length of one-pass diffusion drafters is not due to weak per-position probabilities but to the absence of intra-block causal conditioning, and that a cheap refiner can repair this by re-ranking the drafter's own candidates. xPress is that refiner: it reads the drafter's hidden states and the current block guess, adds a per-position logit bias through a strictly lower-triangular low-rank mixer, and resolves the resulting autoregressive dependency with K parallel Jacobi iterations. Theorem 1 guarantees the iteration reaches the sequential greedy decode in at most B steps, and the paper finds K≈6 is enough in practice. The paper reports that on the

What carries the argument

The load-bearing object is the causal refiner: a low-rank (r=256) strictly lower-triangular linear mixer combined with a residual MLP and a shared low-rank readout, which maps the drafter's per-position hidden states, a block-global mean-pooled summary, and the previous token index to an additive logit bias. Its causal form (position k sees only the prefix x_{<k}) makes the refined block factorize autoregressively, and because fuse and mix are both linear, they fold into a single matmul at inference. The companion mechanism is Jacobi decoding: all B positions are updated in parallel from the previous iterate, with Theorem 1's guarantee of finite convergence to the sequential decode and empir

Load-bearing premise

The approach assumes the correct token is, at each position, already among the top-k candidates of the diffusion drafter's marginal distribution—a property the paper cites from prior work rather than re-measuring for dFlash on Qwen3-8B; if that recall is low, refining can only re-rank the wrong shortlist and the acceptance gains disappear.

What would settle it

On a held-out set of blocks from the dFlash drafter on Qwen3-8B, compute how often the token that the target model actually accepts at each position falls within the drafter's top-k (e.g., top-16) logits; if this recall is far below the method's implicit assumption, the reported acceptance-length gains should collapse when the test is rerun.

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

If this is right

  • Acceptance length rises roughly 30% on average relative to the plain dFlash drafter (18–56% greedy, 19–58% at sampling temperature 1) across math, code, and chat benchmarks.
  • End-to-end decoding throughput rises about 1.3x on average, up to 1.7x, under greedy decoding, and about 1.46x at T=1, with the gains surviving in batched serving (a 1.2–1.4x edge over the plain drafter at batch sizes 16–32).
  • xPress matches or beats a serial 15-step Markov correction head in acceptance length after only four Jacobi iterations, while the refiner then costs about 1.6x less than the serial head per block.
  • Because the refined block factorizes autoregressively, the method remains lossless under speculative sampling and can be composed with other drafters or with tree-based verification built on the refined logits.

Where Pith is reading between the lines

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

  • The method's ceiling is set by the drafter's top-k recall, so one likely extension is to train the drafter to improve top-k coverage jointly with the refiner, which should compound the acceptance gains.
  • An adaptive stopping rule for the Jacobi iterations (halt when the block stops changing) would cut the average K below the fixed K=6 and reduce refinement latency further; the paper's own convergence analysis implies this is safe, and the authors list it as a future direction.
  • The same parallel fixed-point idea could generalize to any one-shot multi-output model—parallel translation, block generation, or structured prediction—where per-position marginals are good but joint consistency is weak and a few Jacobi passes can inject mutual dependence without a serial decode.
  • Because the refiner is trained with a total-variation term against the frozen target, the method would likely transfer to other target/drafter pairs as long as the top-k premise holds, but the K needed may change with the drafter's convergence behavior.
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

4 major / 3 minor

Summary. The paper proposes xPress, a lightweight causal refiner attached to a block-diffusion drafter (dFlash). The refiner reads the drafter's hidden states and a block-global summary, down-projects to r=256, applies a per-channel strictly lower-triangular causal mixer, an MLP, and a low-rank readout, adding a logit bias to the drafter's base logits. The refined block is decoded by parallel Jacobi iterations (Eq. 11); Theorem 1 proves finite convergence to the sequential greedy decode in at most B iterations. Training combines teacher-forced cross-entropy, an ℓ1/TV acceptance loss (Eqs. 13-14), a drafter-anchor loss, and a one-step consistency loss (Eq. 15). On Qwen3-8B with a block-16 dFlash drafter, the paper reports acceptance-length gains of 18-56% (greedy) and 19-58% (T=1), throughput gains of about 1.29x (T=0) and 1.46x (T=1) over dFlash, and vLLM serving gains at batch sizes 16-32, with lower per-block latency than a sequential Markov-head baseline.

Significance. The contribution is real: the paper identifies a genuine mismatch in diffusion drafters (marginal vs path-conditioned proposals) and gives a cheap, parallel way to inject causality. The core identities are correct: Eq. (13), acceptance = 1 - TV, is a valid speculative-sampling identity, and the proof of Theorem 1 is sound, with the caveat that it concerns the refiner's own sequential decode, not target-model acceptance. The architecture is simple and the latency table is informative. If the empirical numbers are robust, xPress is a useful component for diffusion-based speculative decoding and compares favorably with the serial Markov head. However, the quantitative claims rest on an unverified design premise and on statistics that lack variance reporting, so the paper needs strengthening before publication.

major comments (4)
  1. [Section 1; Eq. (4)] The motivating claim that 'the correct token is frequently among the drafter's top-k candidates' is cited to the authors' own PRESTO [9] and not measured for dFlash on Qwen3-8B. Although the logit bias δ is full-vocabulary, the low-rank readout is biased toward re-ranking the drafter's favorites; if target-consistent tokens are far outside the base top-k, the refiner has limited capacity to promote them. This premise is load-bearing for the design. Please report recall@k of the dFlash base marginals on a held-out split of the training distribution (or on the eval benchmarks), and ideally an ablation showing how xPress performs as a function of base recall. A direct measurement would settle whether the mechanism is fixing ranking noise or doing something stronger.
  2. [Section 5.2, Fig. 3 / Table 1] Figure 3 is used to justify K≈6, but it is plotted on the same GSM8K/MATH-500/HumanEval/MBPP/LCB/MT-Bench curves that appear in the headline Table 1. If K is selected from these curves, the reported 28-30% acceptance gain and the latency advantage (Table 3) are test-set-tuned. State whether K was chosen on a separate validation set, and report results for a fixed K across a range (e.g., K=4,6,8) with confidence intervals so the reader can assess sensitivity. This is central because the throughput claims depend directly on K.
  3. [Tables 1-2; Sec. 5.1] No error bars or confidence intervals are given anywhere. T=1 τ is averaged over 5 seeds and throughput over 5 measured rounds, but no spread is reported, despite the text acknowledging several-percent GPU clock drift. Many comparisons in Table 1 are close (e.g., 6.6x vs 6.3x on HumanEval at T=0; 5.8x vs 5.4x on AIME), and the average ×Gain figures are the paper's main quantitative result. Please provide per-seed/per-round values, standard errors, or bootstrap CIs for the headline τ and throughput ratios.
  4. [Section 4.3, Eq. (15)] The consistency loss trains the refiner on only one self-conditioned forward pass (the drafter-seeded p̂r), while inference uses K≈6 Jacobi iterations. The manuscript does not apply any training signal to inputs arising from the second and later iterations, so the claimed alignment between training and the Jacobi trajectory is only partially realized. Either extend the consistency objective to multi-step unrolled iterations (with stop-gradient or similar), or add an ablation showing that one-step consistency is sufficient. This is directly relevant to the central claim that K≈6 parallel refinements are faithful to the sequential decode.
minor comments (3)
  1. [Throughout] Typos: 'accpeted' in Sec. 3.1; 'reply on massive drafters', 'itslef', 'diffuison' in Sec. 2.1; reference [14] listed as 'JetSPec' but should be 'JetSpec'; Table 3 header 'K=12' appears to be a typo for 'K=1 2'.
  2. [Eq. (14), Sec. 5.1] The hyperparameters α_ce, α_ℓ1, and γ are not specified. These are needed for reproducibility because they control the balance between cross-entropy and ℓ1/TV acceptance loss and the position-decay schedule.
  3. [Figure 3] The caption and y-axis use 'Acceptance length' but the values appear to be per-step τ; clarify the unit and whether this is averaged over all generated tokens or per verification step.

Circularity Check

1 steps flagged

Load-bearing top-k premise rests on the authors' own PRESTO citation and is not re-measured; the headline gains remain independent benchmark measurements.

specific steps
  1. self citation load bearing [Section 1, paragraph 6; Section 4.1, Eq. (3)-(4) and 'Lightweight' paragraph]
    "What makes this plausible is a property already established for diffusion drafters: the correct token is frequently among the drafter's top-k candidates at each position [9]. ... Consistent with the aforementioned top-k observation, the refiner does not score the vocabulary from scratch. It adds a small per-position logit bias δk that re-ranks the handful of candidates the drafter already favours."

    The method's recovery mechanism is restricted to re-ranking the drafter's pre-existing candidates ('does not score the vocabulary from scratch'), so the achievable acceptance length is bounded by the drafter's top-k recall. The paper justifies this restriction by citing [9] — the authors' own prior PRESTO paper — rather than measuring top-k coverage for the actual dFlash/Qwen3-8B setting used in the headline results. If the target-consistent token is not in the drafter's top-k, Eq. (4) cannot recover it and the claimed gains would vanish. This is load-bearing self-citation. However, Tables 1-2 report held-out target-verification measurements, not quantities derived from this premise, so the central empirical claim remains independent evidence.

full rationale

The central claims — ~30% higher acceptance length and ~1.3x end-to-end throughput over dFlash — are external measurements against the frozen Qwen3-8B target on held-out benchmarks, not quantities derived from the paper's own equations. Equation (13) is the standard acceptance/TV identity used as a training loss, and using it to train toward acceptance is a legitimate objective, not a fitted prediction. Theorem 1 is a self-contained Jacobi convergence proof for the refiner's own sequential greedy decode, with stated assumptions; it does not by itself guarantee target consistency, but it is not circular. The only circularity-adjacent step is the load-bearing top-k premise: the refiner only re-ranks drafter-favored candidates, and this restriction is justified by [9], a prior paper with overlapping authorship, without re-measuring top-k recall on the evaluated dFlash/Qwen3-8B pair. That self-citation supports the design's ceiling but does not generate the reported numbers, so the overall score is moderate rather than high.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The central claim (acceptance-length and throughput gains) rests on premises the paper does not itself establish: the standard Jacobi convergence fact (Theorem 1, from [18]/[19]); the acceptance = 1 − TV identity (Eq. 13, from [2]); the top-k containment of correct tokens in drafter marginals (from the authors' own PRESTO [9]); the fixed-hidden-state assumption that makes the Jacobi argument exact; and the training-to-inference transfer of the consistency loss. Free parameters (r=256, K≈6, λ=0.3, β 0.6→0.2, plus unstated α_ce, α_ℓ1, γ) are honest design choices, except that K is selected on the evaluation benchmarks' own τ-vs-K curves, so the headline operating point is partly tuned on the reported data. No invented entities: the refiner is a learned parameter module, not a new postulated thing.

free parameters (7)
  • r (refiner low-rank dimension) = 256
    Hand-chosen inner dimension for the refiner's r-space; not ablated.
  • K (number of Jacobi iterations at inference) = ≈6
    Picked from Figure 3's τ-vs-K plateaus on the same evaluation benchmarks; the paper notes K=4 already matches the Markov head and K=16 never beats the plateau.
  • λ (consistency loss weight) = 0.3
    Eq. (15); chosen, not ablated.
  • β (drafter-anchor loss weight) = annealed 0.6 → 0.2
    Eq. (15); annealed over training to shift capacity toward the refiner.
  • α_ce, α_ℓ1 (per-loss weights, Eq. 14) = not reported
    Weights of the cross-entropy and total-variation terms; values not given in the paper.
  • γ (prefix-position decay in w_k, Eq. 14) = not reported
    Exponential decay emphasizing early positions; inherited from dFlash, value not given.
  • learning rate = 6×10⁻⁴
    AdamW lr for all heads; standard hyperparameter rather than a fitted constant.
axioms (6)
  • standard math Jacobi fixed-point iteration on a strictly causal system converges to the sequential greedy decode in at most B iterations.
    Invoked in Theorem 1 (Section 4.2); the paper itself cites [18,19] for exactly this result, making the theorem a re-derivation.
  • standard math Per-token acceptance probability under speculative sampling equals Σ_x min(p(x), p_t(x)) = 1 − TV(p, p_t).
    Eq. (13), Section 4.3; from Leviathan et al. [2]; justifies the total-variation loss as an acceptance surrogate.
  • domain assumption The correct token is frequently among the drafter's top-k candidates at each position.
    Section 1 and Section 4.1 motivation; cited to the authors' own PRESTO [9], not re-measured on dFlash/Qwen3-8B in this paper.
  • domain assumption Drafter hidden states h_{1:B} and global summary g are computed once and held fixed across Jacobi iterations.
    Section 4.2, premise of Theorem 1; a modeling choice that makes the fixed-point argument exact.
  • domain assumption Refiner training under teacher forcing plus one drafter-seeded second pass transfers to the actual K-step Jacobi inference trajectory.
    Section 4.3; the consistency loss mitigates the train/inference mismatch but full Jacobi trajectories are not sampled during training.
  • domain assumption Lossless sampling at T=1 is preserved when the draft is produced by the refined distribution p_r.
    Section 3.1 asserts SD losslessness; the paper never specifies how T=1 drafts are drawn from the refiner (Algorithm 1 uses arg max) or verifies quality preservation.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of xPress: Parallel Refinement for Diffusion Drafters in Speculative Decoding." pith.science (2026). https://pith.science/paper/NAT62ADI

@misc{pith2026260802438,
  author       = {Pith},
  title        = {Pith review of: xPress: Parallel Refinement for Diffusion Drafters in Speculative Decoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NAT62ADI}},
  note         = {Machine review of arXiv:2608.02438}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Block-diffusion drafters like dFlash generate an entire block of draft tokens in a single forward pass, drastically reducing the overhead of multiple-token drafting in speculative decoding. The crucial final step of the single-pass discrete denoising process involves using the logit distribution at each position to sample conditionally independent tokens. The resulting draft is thus a set of per-position marginals, rather than a joint distribution: no draft token is guaranteed to depend on its predecessors. Such independently sampled marginals tend to produce sequences with tokens that are individually likely, but jointly improbable under the target model's distribution, which verifies each token conditionally. This can cause early rejection and limits acceptance length. To address this, we propose xPress as a means to restore the missing causality in diffusion drafters. xPress is a lightweight causal refiner that reconciles the whole diffusion block at once through parallel refinement, restoring and propagating causal dependencies across the draft without a token-by-token loop. On Qwen3-8B, across seven math, code, and chat benchmarks, xPress raises acceptance length by about 30% on average (up to +56%) and its end-to-end decoding throughput by about 1.3 on average (up to 1.7) compared to the original dFlash diffusion drafter.

Figures

Figures reproduced from arXiv: 2608.02438 by Davis Wertheimer, Minjia Zhang, Mudhakar Srivatsa, Naigang Wang, Raghu K. Ganti, Yu Chin Fabian Lim, Zheng Wang.

Figure 1
Figure 1. Figure 1: End-to-end decoding throughput speedup over the autoregressive baseline on Qwen3-8B with a block-16 dFlash drafter under greedy decoding (T=0, left) and lossless sampling (T=1, right). The throughput speedup values are measured on H200. xPress consistently outper￾forms the Markov head and dFlash diffusion drafter. acceptance length, which requires an unbroken chain of accepted verifications, does not. dFla… view at source ↗
Figure 2
Figure 2. Figure 2: (a) The full pipeline. The block-diffusion drafter produces hidden states, the target LM head reads out the base logits, and the refiner adds a learned logits bias to form the final logits. (b) Inside the refiner. The three inputs, the token id, the global hidden state, and the per-position hidden state, are down-projected into r-space, fused, mixed causally across block positions, passed through the r-spa… view at source ↗
Figure 3
Figure 3. Figure 3: Per-step τ versus the number of Jacobi iterations K, from K=1 to 16, with the Markov head as a horizontal baseline (dashed). τ climbs past the baseline within a few iterations, then flat￾tens; the small non-monotone wiggles past the plateau are dataset-dependent (e.g. LiveCodeBench). reaches a 6.2× average speedup at 8.02 accepted tokens per step, improving over the plain dFlash drafter (4.9×, 6.25) by 1.2… view at source ↗
Figure 4
Figure 4. Figure 4: Composition of the draft-side latency per block for the Markov head (GSM8K, block 16, [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

30 extracted references · 9 linked inside Pith

  1. [1]

    Accelerating large language model decoding with speculative sampling, 2023

    Charlie Chen, Sebastian Borgeaud, Geoffrey Irving, Jean-Baptiste Lespiau, Laurent Sifre, and John Jumper. Accelerating large language model decoding with speculative sampling, 2023

    2023

  2. [2]

    Fast inference from transformers via speculative decoding.ICML, 2023

    Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding.ICML, 2023. arXiv:2211.17192

    Pith/arXiv arXiv 2023

  3. [3]

    Eagle: Speculative sampling requires rethinking feature uncertainty.ICML, 2024

    Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. Eagle: Speculative sampling requires rethinking feature uncertainty.ICML, 2024. arXiv:2401.15077

    Pith/arXiv arXiv 2024

  4. [4]

    Eagle-2: Faster inference of language models with dynamic draft trees, 2024

    Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. Eagle-2: Faster inference of language models with dynamic draft trees, 2024

    2024

  5. [5]

    Eagle-3: Scaling up inference acceleration of large language models via training-time test, 2025

    Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. Eagle-3: Scaling up inference acceleration of large language models via training-time test, 2025

    2025

  6. [6]

    J. Chen, Y. Liang, and Z. Liu. Dflash: Block diffusion for flash speculative decoding. 2026. arXiv:2602.06036

    Pith/arXiv arXiv 2026

  7. [7]

    Fast-dllm v2: Efficient block-diffusion llm, 2025

    Chengyue Wu, Hao Zhang, Shuchen Xue, Shizhe Diao, Yonggan Fu, Zhijian Liu, Pavlo Molchanov, Ping Luo, Song Han, and Enze Xie. Fast-dllm v2: Efficient block-diffusion llm, 2025

    2025

  8. [8]

    Tidar: Think in diffusion, talk in autoregression, 2025

    Jingyu Liu, Xin Dong, Zhifan Ye, Rishabh Mehta, Yonggan Fu, Vartika Singh, Jan Kautz, Ce Zhang, and Pavlo Molchanov. Tidar: Think in diffusion, talk in autoregression, 2025

    2025

  9. [9]

    Z. Wang, Z. Ye, Y. Fu, Q. Cheng, Z. Wang, F. Zhu, H. Zhao, H. Shi, P. Molchanov, and M. Zhang. Presto: Prefix-aligned tree drafting for diffusion speculative decoding. InICML Workshop (AdaptFM), 2026. OpenReview:VKD70LJaDy

    2026

  10. [10]

    Ringel and Y

    L. Ringel and Y. Romano. Accelerating speculative decoding with block diffusion draft trees (ddtree). 2026. arXiv:2604.12989

    Pith/arXiv arXiv 2026

  11. [11]

    Diffuspec: Unlocking diffusion language models for speculative decoding, 2025

    Guanghao Li, Zhihui Fu, Min Fang, Qibin Zhao, Ming Tang, Chun Yuan, and Jun Wang. Diffuspec: Unlocking diffusion language models for speculative decoding, 2025

    2025

  12. [12]

    Christopher, Thomas Hartvigsen, and Ferdinando Fioretto

    Jameson Sandler, Jacob K. Christopher, Thomas Hartvigsen, and Ferdinando Fioretto. Specdiff-2: Scaling diffusion drafter alignment for faster speculative decoding, 2025

    2025

  13. [13]

    Nemotron-labs-diffusion: A tri-mode language model unifying autoregressive, diffusion, and self-speculation decoding, 2026

    Yonggan Fu, Lexington Whalen, Abhinav Garg, Chengyue Wu, Maksim Khadkevich, Nicolai Oswald, Enze Xie, Daniel Egert, Sharath Turuvekere Sreenivas, Shizhe Diao, Chenhan Yu, 14 Ye Yu, Weijia Chen, Sajad Norouzi, Jingyu Liu, Shiyi Lan, Ligeng Zhu, Jin Wang, Jindong Jiang, Morteza Mardani, Mehran Maghoumi, Song Han, Ante Jukić, Nima Tajbakhsh, Jan Kautz, and P...

    2026

  14. [14]

    Jetspec: Breaking the scaling ceiling of speculative decoding with parallel tree drafting, 2026

    Lanxiang Hu, Zhaoxiang Feng, Yulun Wu, Haoran Yuan, Yujie Zhao, Yu-Yang Qian, Bojun Wang, Peng Zhao, Daxin Jiang, Yibo Zhu, Tajana Rosing, and Hao Zhang. Jetspec: Breaking the scaling ceiling of speculative decoding with parallel tree drafting, 2026

    2026

  15. [15]

    Domino: Decoupling causal modeling from autoregressive drafting in speculative decoding

    Huang et al. Domino: Decoupling causal modeling from autoregressive drafting in speculative decoding. 2026. arXiv:2605.29707

    Pith/arXiv arXiv 2026

  16. [16]

    Cheng, X

    X. Cheng, X. Yu, C. Shao, J. Li, Y. Xiong, et al. Dspark: Confidence-scheduled speculative decoding with semi-autoregressive generation. 2026. DeepSeek-AI, arXiv:2607.05147

    Pith/arXiv arXiv 2026

  17. [17]

    S. L. Wang, P. Isola, and B. Cheung. The truth lies somewhere in the middle (of the generated tokens). 2026. arXiv:2605.09969

    Pith/arXiv arXiv 2026

  18. [18]

    Accelerating feedforward compu- tation via parallel nonlinear equation solving.ICML, 2021

    Yang Song, Chenlin Meng, Renjie Liao, and Stefano Ermon. Accelerating feedforward compu- tation via parallel nonlinear equation solving.ICML, 2021. arXiv:2002.03629

    Pith/arXiv arXiv 2021

  19. [19]

    Santilli et al

    A. Santilli et al. Accelerating transformer inference for translation via parallel decoding.ACL,

  20. [20]

    S. Kou, L. Hu, Z. He, Z. Deng, and H. Zhang. Cllms: Consistency large language models. ICML, 2024. arXiv:2403.00835

    Pith/arXiv arXiv 2024

  21. [21]

    Qwen3 technical report, 2025

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

    2025

  22. [22]

    The perfect blend: Redefining rlhf with mixture of judges, 2024

    Tengyu Xu, Eryk Helenowski, Karthik Abinav Sankararaman, Di Jin, Kaiyan Peng, Eric Han, Shaoliang Nie, Chen Zhu, Hejia Zhang, Wenxuan Zhou, Zhouhao Zeng, Yun He, Karishma Mandyam, Arya Talabzadeh, Madian Khabsa, Gabriel Cohen, Yuandong Tian, Hao Ma, Sinong Wang, and Han Fang. The perfect blend: Redefining rlhf with mixture of judges, 2024

    2024

  23. [23]

    Training verifiers to solve math word problems, 2021

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

    2021

  24. [24]

    Let’s verify step by step, 2023

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step, 2023

    2023

  25. [25]

    American invitational mathematics examination (aime).https://maa.org/ math-competitions/aime

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

  26. [26]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul 15 Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavar...

    2021

  27. [27]

    Programsynthesis with large language models, 2021

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, EllenJiang, CarrieCai, MichaelTerry, QuocLe, andCharlesSutton. Programsynthesis with large language models, 2021

    2021

  28. [28]

    Livecodebench: Holistic and contamination free evaluation of large language models for code, 2024

    Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Ar- mando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code, 2024

    2024

  29. [29]

    Xing, Hao Zhang, Joseph E

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023

    2023

  30. [30]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention, 2023. 16

    2023

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