Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

LANTERN++: Enhancing Relaxed Speculative Decoding with Static Tree Drafting for Visual Auto-regressive Models

T0 review · 5 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper argues that static draft trees plus a multiplicative relaxation bound can speed up visual autoregressive decoding by up to 2.56x without retraining the target model.

desk verdict A plausible incremental extension of LANTERN whose headline quality claim is undercut by its own FID numbers; worth a referee but needs a direct comparison against LANTERN and EAGLE-2. read the letter →

arxiv 2502.06352 v2 pith:JMNKKV5V submitted 2025-02-10 cs.CV

Sihwan Park , Doohyuk Jang , Sungyub Kim , Souvik Kundu , Eunho Yang This is my paper · ORCID
classification cs.CV
keywords speculativedecodingvisualautoregressivemodelstokenselectionambiguitystatictreedraftingdynamicrelaxedacceptancemultiplicativeboundimagegenerationacceleration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that the usual recipe for speeding up autoregressive image generation—dynamic tree drafting, where the draft tree grows only where the drafter is confident—is the wrong tool for visual autoregressive models, because image-token distributions are flat and ambiguous. It proposes LANTERN++, which instead uses a fixed, deep draft tree and relaxes the acceptance test so a draft token is accepted when the target model's probability mass near that token is high, scaled by a multiplicative bound. On Lumina-mGPT, Anole, and LlamaGen-XL, this raises step compression over EAGLE-1 and cuts latency by up to 2.56x relative to standard decoding. The price is a tunable quality shift: larger relaxation buys speed at the cost of higher FID. If correct, the framework shows that the bottleneck in visual speculative decoding is drafting structure, not drafter accuracy.

What carries the argument

The central machinery is static tree drafting with a fixed 58-node draft tree, combined with a neighborhood-based relaxed acceptance condition governed by a multiplicative bound. The static tree is generated regardless of drafter confidence, so low-confidence steps do not prune depth. The relaxed acceptance aggregates target probability over the k nearest codebook neighbors of a draft token and accepts when the aggregate is below a multiplicative factor times the draft token's own target probability, replacing LANTERN's additive TVD bound. This keeps the acceptance boost proportional to the target's own likelihood, avoiding over-amplification of low-probability tokens while allowing deeper draft sequences to survive.

What would settle it

Generate the same prompts with standard autoregressive decoding and with LANTERN++ at no relaxation versus a strong relaxation, then measure per-image perceptual distance between outputs; if images diverge sharply as the bound increases even when FID is similar, the relaxed acceptance is changing content rather than just accelerating. A sharper test is to replace each accepted draft token with a random codebook neighbor at the same latent distance and check whether image quality stays flat; if it drops, latent proximity alone does not guarantee visual interchangeability.

Watch

Extended reading notes

Core claim

LANTERN++ claims that replacing confidence-adaptive dynamic tree drafting with a fixed static tree, combined with a multiplicative relaxation bound, unlocks the acceleration that relaxed speculative decoding promised for visual autoregressive models. The acceptance rule sums target-model probability over the k codebook neighbors of a draft token and accepts the token if that sum stays below a multiplicative factor times the token's own target probability. The fixed tree lets deep draft sequences survive flat token distributions, and the multiplicative bound keeps the relaxation proportional across tokens with very different probabilities. The reported result is up to 2.56x latency speedup and 3.63x step compression over standard autoregressive decoding, beating EAGLE-1's 2.94x compression on Lumina-mGPT, with FID rising from 28.93 to 33.91 at lambda=3 on that model. The paper frames this FID shift as a tunable speed-quality trade-off rather than a failure of the method.

Load-bearing premise

The whole speedup rests on the assumption that tokens that are close in the latent codebook produce visually interchangeable image content, so summing the target's probability over a draft token's neighbors is a valid stand-in for accepting that token.

Editorial extensions

If this is right

  • On the three evaluated visual AR models, LANTERN++ outperforms EAGLE-1 in step compression and latency, indicating that static drafting is the stronger baseline under token selection ambiguity.
  • Larger values of the relaxation bound and larger neighborhoods monotonically raise step compression, giving users a practical dial between speed and quality.
  • Because the target model is unchanged, the acceleration applies without retraining the visual autoregressive model; only the single-layer drafter needs training.
  • The quality cost is measurable and grows with the bound: on Lumina-mGPT, FID rises from 28.93 to 33.91 at the strongest setting evaluated, so the operating point must be chosen deliberately.
  • The comparison implies that dynamic tree drafting should not be assumed superior for visual AR models, since the modality's flat distributions invert the usual LLM trade-off.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not explore it, but the same relaxed acceptance could apply to any autoregressive model with a continuous codebook—audio, video, or 3D—where token selection ambiguity is expected, not just images.
  • Because FID cost grows with the relaxation bound, an adaptive schedule that shrinks the bound when the drafter is confident and grows it when confidence is low could recover quality without losing speed; this is an extension, not a paper claim.
  • A direct human or perceptual study on whether latent codebook neighbors are truly interchangeable would decide how far the bound can be pushed before artifacts appear; the paper's FID numbers alone leave this open.
  • The paper's framing suggests the speed-quality trade-off is adjustable, which implies deployment could tune to a target latency; we infer this makes the method more practical than methods with a single operating point.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes LANTERN++, a method for accelerating visual autoregressive models by combining static tree drafting with a multiplicative relaxation bound λ in the acceptance condition of speculative decoding. It argues that dynamic tree drafting suffers from token selection ambiguity in visual AR models, leading to shallow draft trees and low acceptance rates, and that static drafting with a multiplicative bound enables deeper accepted sequences. Experiments on Lumina-mGPT, Anole, and LlamaGen-XL report step compression up to ×3.63 and latency speedup up to ×2.56 over standard AR decoding, with FID and CLIP scores reported as quality metrics.

Significance. The paper addresses a relevant and timely problem: making visual autoregressive generation faster on a single GPU without retraining the target model. The conceptual analysis of why dynamic tree drafting underperforms for visual AR models is plausible, and the proposed static-tree-plus-multiplicative-relaxation recipe is simple and reproducible, with code released. The method also introduces a concrete hyperparameter (λ) that trades speed for quality. However, the central quality-preservation claim is not established at the operating point where the headline speedup is reported, the comparison omits the two most relevant baselines (LANTERN and EAGLE-2), and no statistical uncertainty is provided. If the missing baselines and quality analysis are supplied, the contribution could be useful for practitioners.

major comments (5)
  1. [Abstract; §5, Table 2] The claim that LANTERN++ 'maintain[s] high image quality' while achieving ×2.56 latency reduction is not supported by the paper's own measurements. On Lumina-mGPT, FID increases from 28.93 (standard AR) to 33.91 at λ=3, and on Anole from 20.28 to 25.48 at λ=3; these are substantial degradations, not minimal. The ×2.56 latency and ×3.63 step compression figures are reported exactly at λ=3, while the smaller FID degradation at λ=2 (30.11 and 21.10) is accompanied by lower speedups (×2.28 and ×3.19 on Lumina-mGPT, ×1.85 and ×2.95 on Anole). Thus no single operating point is shown where the headline speedup and high image quality co-occur.
  2. [§5, Table 2; §3, Table 1] The paper motivates LANTERN++ as a refinement of LANTERN and argues against dynamic tree drafting (EAGLE-2), yet neither LANTERN nor EAGLE-2 appears in the main acceleration comparison. Table 1 reports EAGLE-1 versus EAGLE-2 on LlamaGen-3B and Vicuna-7B only, which are not the models evaluated in Table 2. Consequently, the central improvement-over-LANTERN claim and the claim that static drafting outperforms dynamic drafting for visual AR models are not directly tested. Please add LANTERN and EAGLE-2 results on the same models and settings.
  3. [§2.2, §4] The relaxed acceptance condition is valid only if tokens in the latent neighborhood Ak,λ(bx) are visually interchangeable with bx. The paper asserts this in Section 2.2 but provides no validation, and the FID degradation in Table 2 indicates the assumption has limits. Please provide direct evidence, for example a study of nearest-neighbor visual similarity, a comparison of FID at fixed speedup across k and λ, or an analysis of which tokens are accepted by relaxation versus exact speculative sampling.
  4. [§5, Appendix C.2, Appendix C.3] No confidence intervals, standard errors, or multi-seed runs are reported for any acceleration or quality metric. Given that several FID differences are small (e.g., LlamaGen-XL Stage I: 23.64 versus 23.89; Stage II: 40.52 versus 39.80), it is impossible to determine whether the reported improvements and degradations are statistically meaningful. Please report variance or multiple seeds for the central numbers.
  5. [§4] The paper introduces λ as a multiplicative bound but provides no theorem or measurement showing that the output distribution remains close to the target model's distribution. The claim of 'preserving distributional consistency' is supported only by FID and CLIP scores at two λ values. Please either provide a formal distortion bound or an empirical distribution-distance measurement.
minor comments (5)
  1. [§2, Notations] The roles of p and q are defined in one sentence but the drafter/target notation is used in several later subsections; please standardize by stating once that p always denotes the drafter distribution and q the target distribution.
  2. [Appendix C.2, Table 3] The baseline 'EAGLE-1: 2.81' is repeated in both the multiplicative and additive subtables; state explicitly whether this is the same run and whether the λ/δ results use the original or extended static tree.
  3. [Figure 3] The two tree diagrams are difficult to read because node labels overlap; please redraw with clearer spacing and annotate the depth levels.
  4. [Abstract, §5] The phrase 'up to ×2.56 speedup' should be qualified as 'at λ=3 with the reported FID degradation' to avoid overstatement, since the speedup and quality figures do not co-occur at a single setting.
  5. [Appendix B, Drafter Training] The drafter for Lumina-mGPT is trained on 30K generated images, whereas Anole uses 118K; given that drafter quality directly affects compression ratios, please discuss whether the reported differences are sensitive to drafter training set size.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: headline speedups are measured against standard AR decoding and EAGLE-1 with hand-set hyperparameters; self-citations to LANTERN motivate the relaxation but do not by construction force the reported results.

full rationale

The paper's derivation chain is empirical rather than definitional. The central claims (up to x2.56 latency reduction and x3.63 step compression) are measurements reported in Table 2 against standard AR decoding and EAGLE-1, and the comparison uses the same extended static tree for both EAGLE-1 and LANTERN++; no parameter is fitted to the test set and then renamed as a prediction. The relaxation bound lambda and neighborhood size k are hand-set hyperparameters, and the acceptance rule min(1, sum_{x in A_{k,lambda}} q(x|s) / p(bx|s)) is an algorithmic choice whose quality consequences are evaluated by FID and CLIP rather than asserted by construction. The paper does rely heavily on the authors' prior LANTERN work for the token-ambiguity framing and the latent-similarity premise, but that premise is not the source of the speedup numbers and is in fact tested, and partly challenged, by the paper's own Table 2, where lambda=3 raises FID from 28.93 to 33.91 on Lumina-mGPT and from 20.28 to 25.48 on Anole. That degradation is a correctness or robustness concern about the 'minimal degradation' claim, not a circularity: the reported acceleration does not reduce to the latent-similarity assumption, and no equation in the paper is equivalent to its own input by construction. The self-citations are therefore not load-bearing for the measured speedups, and the paper is self-contained against external baselines.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim depends on hand-chosen hyperparameters lambda, k, and the static tree shape, plus domain assumptions about latent-space semantics inherited from LANTERN. There are no invented physical entities.

free parameters (4)
  • lambda (multiplicative relaxation bound) = lambda = 2, 3, 5, 10, 20 tested; headline uses lambda = 3
    Chosen by hand to trade acceptance rate against distributional distortion; not derived. The abstract's 2.56x speedup uses lambda = 3.
  • k (latent neighborhood size) = k = 5, 10, 20, 50 tested; headline uses k = 10
    Number of nearest codebook neighbors aggregated in the relaxed acceptance rule; selected by hand over a grid.
  • static tree structure = 58-node left-heavy tree (N=58)
    Hand-designed extension of EAGLE-1's 26-node tree to match EAGLE-2's 59 dynamic nodes; affects both step compression and latency.
  • Drafter training set sizes = 100K (LlamaGen), 118K (Anole), 30K (Lumina-mGPT) images
    Self-generated or LAION-COCO samples chosen by the authors; drafter quality and all downstream results depend on these choices.
assumptions (4)
  • domain assumption Visual AR next-token distributions are dispersed, with many tokens sharing similarly low probabilities (token selection ambiguity).
    Section 2.2 presents this as a known characteristic and cites the authors' own LANTERN paper; it motivates the design but is not independently established here.
  • domain assumption Tokens close in the codebook latent space are visually similar and interchangeable.
    Section 2.2: the relaxed acceptance rule aggregates target probability over latent neighbors; if false, accepted drafts can be visually wrong.
  • ad hoc to paper A multiplicative bound lambda preserves the target distribution well enough that image quality is maintained.
    Section 4 claims the bound keeps relaxation proportional and stable, but Table 2 shows FID degrades substantially at lambda=3, so this assumption is only weakly supported.
  • standard math A single forward pass with a tree-aware attention mask verifies the whole draft tree equivalently to sequential verification.
    Section 2.1 takes this from EAGLE/Medusa prior work; it is standard but unproved in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LANTERN++: Enhancing Relaxed Speculative Decoding with Static Tree Drafting for Visual Auto-regressive Models." pith.science (2026). https://pith.science/paper/JMNKKV5V

@misc{pith2026250206352,
  author       = {Pith},
  title        = {Pith review of: LANTERN++: Enhancing Relaxed Speculative Decoding with Static Tree Drafting for Visual Auto-regressive Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JMNKKV5V}},
  note         = {Machine review of arXiv:2502.06352}
}
abstract

Speculative decoding has been widely used to accelerate auto-regressive (AR) text generation. However, its effectiveness for visual AR models remains limited due to token selection ambiguity, where multiple tokens share similarly low probabilities and thus reduce acceptance rates. Recently, relaxed speculative decoding with dynamic tree drafting was proposed to mitigate this ambiguity, demonstrating promising results in accelerating visual AR models. However, we observe that token selection ambiguity still negatively affects dynamic tree drafting, resulting in shallow draft trees and limited acceleration. To overcome this issue, we introduce LANTERN++, a refined framework that integrates static tree drafting with a tailored relaxed acceptance condition, allowing drafts to be selected independently of low-confidence predictions. This enables the acceptance of deeper sequences, improving decoding efficiency while preserving image quality. Extensive experiments on state-of-the-art visual AR models demonstrate that LANTERN++ significantly accelerates inference, achieving up to $\mathbf{\times 2.56}$ speedup over standard AR decoding while maintaining high image quality. The code is publicly available at https://github.com/jadohu/LANTERN.

Figures

Figures reproduced from arXiv: 2502.06352 by the authors.

Figure 1
Figure 1. Images generated by LANTERN++ on Lumina-mGPT ( [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Images generated by LANTERN++ on Lumina-mGPT with average step compression ratio [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Comparison of static tree structures. Top: The original static tree used in EAGLE-1. Bottom: The extended static tree used for both EAGLE-1 and LANTERN++ in our experiments, designed to match the scale of dynamic tree drafting while maintaining EAGLE-1’s structural principles. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Example of a shallow draft tree produced by dynamic tree drafting. Due to low drafter [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Comparison of images generated using additive relaxation ( [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SJD-PAC: Accelerating Speculative Jacobi Decoding via Proactive Drafting and Adaptive Continuation

    cs.CV 2026-03 accept novelty 6.0 of 10

    SJD-PAC combines proactive multi-path drafting and adaptive continuation to raise average acceptance length in Speculative Jacobi Decoding, delivering 3.8 imes lossless wall-clock speedup on Lumina-mGPT and Emu3.

Reference graph

Works this paper leans on

24 extracted references · 5 canonical work pages · cited by 1 Pith paper

  1. [1]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gr...

    1901

  2. [2]

    Lee, Deming Chen, and Tri Dao

    Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D. Lee, Deming Chen, and Tri Dao. Medusa: Simple LLM inference acceleration framework with multiple decoding heads. In Forty-first International Conference on Machine Learning, 2024. URL https://openreview.net/forum?id=PEpbUobfJv

    2024

  3. [3]

    Anole: An open, autoregressive, native large multimodal models for interleaved image-text generation, 2024

    Ethan Chern, Jiadi Su, Yan Ma, and Pengfei Liu. Anole: An open, autoregressive, native large multimodal models for interleaved image-text generation, 2024. URL https://arxiv.org/abs/2407.06135

    arXiv 2024

  4. [4]

    Laion coco: 600m synthetic captions from laion2b-en, 2022

    Christoph Chuhmann, Andreas Köpf, Richard Vencu, Theo Coombes, and Romain Beaumont. Laion coco: 600m synthetic captions from laion2b-en, 2022. URL https://laion.ai/blog/laion-coco/. September 27th, 2024

    work page 2022

  5. [5]

    Break the sequential dependency of LLM inference using lookahead decoding

    Yichao Fu, Peter Bailis, Ion Stoica, and Hao Zhang. Break the sequential dependency of LLM inference using lookahead decoding. In Forty-first International Conference on Machine Learning, 2024. URL https://openreview.net/forum?id=eDjvSFOkXw

    work page 2024

  6. [6]

    CLIPS core: A reference-free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPS core: A reference-free evaluation metric for image captioning. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (eds.), Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp.\ 7514--7528, Online and Punta Cana, D...

    doi:10.18653/v1/2021.emnlp-main.595 2021

  7. [7]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.), Advances in Neural Information Processing Systems, volume 30. Curran Associates,...

    work page 2017

  8. [8]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. URL https://openreview.net/forum?id=qw8AKxfYbI

    2021

Show all 24 references
  1. [9]

    Lantern: Accelerating visual autoregressive models with relaxed speculative decoding, 2025

    Doohyuk Jang, Sihwan Park, June Yong Yang, Yeonsung Jung, Jihun Yun, Souvik Kundu, Sung-Yub Kim, and Eunho Yang. Lantern: Accelerating visual autoregressive models with relaxed speculative decoding, 2025. URL https://arxiv.org/abs/2410.03355

    2025 arXiv

  2. [10]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas ...

    2023 arXiv

  3. [11]

    Fast inference from transformers via speculative decoding

    Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. In International Conference on Machine Learning, pp.\ 19274--19286. PMLR, 2023

    2023

  4. [12]

    Eagle: Speculative sampling requires rethinking feature uncertainty, 2024 a

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

    2024

  5. [13]

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

    Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. Eagle-2: Faster inference of language models with dynamic draft trees. arXiv preprint arXiv:2406.16858, 2024 b

    2024 arXiv

  6. [14]

    Lawrence Zitnick

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll \'a r, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars (eds.), Computer Vision -- ECCV 2014, p...

    2014

  7. [15]

    Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining, 2024

    Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining, 2024. URL https://arxiv.org/abs/2408.02657

    2024 arXiv

  8. [16]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=Bkg6RiCqY7

    2019

  9. [17]

    Autoregressive model beats diffusion: Llama for scalable image generation

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024

    2024 arXiv

  10. [18]

    Chameleon: Mixed-modal early-fusion foundation models

    Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024

    2024 arXiv

  11. [19]

    Llama: Open and efficient foundation language models, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language...

    2023 arXiv

  12. [20]

    Judging llm-as-a-judge with mt-bench and chatbot arena

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36: 0 46595--46623, 2023

    2023

  13. [21]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  14. [22]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  15. [23]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  16. [24]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.