Pith. sign in

REVIEW 4 major objections 5 minor 41 references

DAVET: Denoising-Aware Visual Evidence Trajectory Allocation for Diffusion Vision-Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Visual evidence demand in diffusion VLMs is step-dependent and backbone-specific, and a training-free allocation policy can cut conditioning cost 1.55x while losing under 2% quality.

desk verdict A plausible training-free allocation method for diffusion VLMs, with a real generality gap: the binary phase orientation is fitted on two benchmarks and not validated on the other two. read the letter →

arxiv 2608.01821 v1 pith:TCBOIWDF submitted 2026-08-03 cs.CV cs.LG

Yongkang Zhou , Xiang Xia , Cheng Yan , Fan Xu , Wuyang Zhang This is my paper
classification cs.CVcs.LG
keywords diffusionvision-languagemodelsvisualevidenceallocationtraining-freeaccelerationdenoising-awareschedulingtokencompressionenvelopequality-efficiencytrade-offmaskeddecoding
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

Diffusion vision-language models (dVLMs) condition every denoising step on the same visual tokens, so producing even a short answer repeatedly pays a large visual-context cost. This paper argues that those costs are misallocated: controlled interventions show visual evidence demand is steeply step-dependent, with LLaDA-V needing fine-grained evidence early in denoising and LaViDa needing it at the end. The authors treat visual evidence as a step-level budget and introduce DAVET, a training-free policy that builds an evidence envelope from a backbone-aligned phase trajectory plus a query-level operation demand, then spends within it using prediction confidence and churn. Across two backbones and four visual QA benchmarks, DAVET reports a 1.55x average speedup at a 1.86% average quality drop. If the picture holds, the recurring cost of visual tokens in diffusion generation is not fixed: it can be cut by spending evidence at the phase of denoising that actually consumes it.

What carries the argument

The evidence envelope — a per-step interval whose lower boundary is a normalized backbone-specific phase profile and whose upper boundary adds an operation-conditioned reserve — together with trajectory risk rt = 1 − alpha·ct−1(1 − churn), built from mean confidence and prediction churn. The risk term interpolates the step budget inside the envelope, and an evidence realizer with an ordered coarse/mixed/high hierarchy, precomputed from a single visual encoding, maps each budget to a view. The load-bearing separation is between when and how much evidence is supplied (the policy) and how the views are built (the realizer).

What would settle it

Run the Figure 1 pulse experiment on a held-out backbone or on DocVQA/TextVQA for the two evaluated backbones: if the high-evidence pulse's peak-quality phase is neither early nor late, or differs from the orientation DAVET assumes, the transfer claim fails directly. Symmetrically, the step-dependence claim itself falls if a search over all eight pulse placements yields quality retention within about 1% for every placement — that would mean demand is flat and DAVET's gains come from pure token compression, not timing.

Watch

Extended reading notes

Core claim

The central discovery: visual evidence demand in dVLMs is step-dependent and backbone-specific — a high-evidence pulse shifted across eight denoising phases retains most quality at Q2–Q3 for LLaDA-V (early grounding) and at Q8 for LaViDa (late verification). DAVET allocates a per-step evidence budget from a phase-conditioned lower trajectory, a query-parser-sized operation reserve, and trajectory risk (confidence and prediction churn), mapping the budget to one of three precomputed evidence views (coarse, mixed, high) of a single encoding, so pooling, pruning, or merging work equally as realizers. On two backbones across four VQA benchmarks it reports a 1.55x average speedup at a 1.86% avera

Load-bearing premise

The load-bearing premise is that the phase orientation read off two benchmarks (early grounding for LLaDA-V, late verification for LaViDa) holds for every other task and backbone — because choosing the wrong orientation costs 24–42% quality, and the paper offers no way to determine orientation for a new model without first running the pulse experiment.

Editorial extensions

If this is right

  • Visual conditioning in dVLMs becomes schedulable: the same average evidence budget yields better quality along a backbone-aligned trajectory than when spent uniformly, making step-dependent allocation the organizing principle rather than fixed visual-token compression.
  • The policy transfers across tasks with fixed per-backbone hyper-parameters, so calibration is one-time per backbone rather than per benchmark.
  • Pooling, pruning, and merging are interchangeable realizers: all stay within 0.51–1.60% of full-evidence quality under the same policy, so future token-compression methods inherit the allocation policy instead of defining it.
  • Phase orientation is the decisive calibration: the wrong orientation costs 24–42% relative quality at unchanged latency, so any application of DAVET must first determine whether the target backbone grounds early or verifies late.
  • The measured speedup comes almost entirely from reduced denoising time (view construction under 0.8 ms per sample), so the savings scale with the number of denoising steps and visual tokens.

Reading between the lines

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

  • The envelope machinery is realizer-agnostic and training-free, so it should port to any masked-diffusion multimodal model; the open question the paper leaves is whether a new backbone's phase orientation can be predicted from architecture statistics (connector depth, step count, visual-attention pattern) rather than measured with the pulse experiment.
  • The deterministic query parser is the least general component: it is hand-built around five operation categories and would likely mislabel compositional or multi-hop queries; a learned demand estimator fed by model confidence could replace it without touching the envelope, a point the paper itself flags.
  • The trajectory-risk signal (confidence times stability) is generic to masked decoding, so it could plausibly schedule other resources on the same envelope — when to commit answer positions, when to cache visual key-values, or how aggressively to prune in later steps — a combination the paper does not test.
  • The LaViDa DocVQA drop (6.83%) is the method's visible boundary and suggests dense document understanding wants fine-grained evidence across much of the trajectory; a per-task rather than per-backbone orientation selector is the natural next experiment.
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

4 major / 5 minor

Summary. The paper proposes DAVET, a training-free visual evidence allocation framework for diffusion vision-language models (dVLMs). DAVET treats visual evidence as a step-dependent resource: a phase-conditioned trajectory supplies the temporal structure, a query-derived operation demand sets the evidence reserve, and a trajectory-risk signal (mean confidence and prediction churn) selects the budget within the envelope. The resulting budget is realized through an ordered hierarchy of coarse/mixed/high evidence views built once from a single visual encoding. Evaluated on LLaDA-V and LaViDa across InfoVQA, ChartQA, DocVQA, and TextVQA, the paper reports a 1.55x average speedup with a 1.86% average relative quality drop. The authors also present controlled pulse interventions showing backbone-dependent early-grounding (LLaDA-V) and late-verification (LaViDa) patterns, ablations of the allocation components, a comparison across evidence realizers (pooling/pruning/merging), and hyperparameter sensitivity sweeps in the appendix.

Significance. The paper addresses a real and under-explored cost bottleneck in dVLMs: recurrent visual conditioning across all denoising steps. The conceptual separation of evidence allocation from evidence realization is useful and general, and the controlled phase-intervention experiments are a valuable diagnostic tool. The reported quality–efficiency trade-off, if it holds beyond the tested configuration, would be a meaningful step beyond static-resolution or fixed-ratio compression baselines. The manuscript is also commendably transparent: it includes a latency breakdown, a uniform-demand control, a reverse-orientation control, and full hyperparameter sweeps. These strengths make the empirical claims internally consistent. However, the central claim depends on several choices that are not fully validated outside the benchmarks used to derive them, in particular the backbone-specific phase orientation and the per-backbone hyperparameters. These issues are fixable but currently limit the generality of the headline numbers.

major comments (4)
  1. [§5.1, Table F] The backbone-specific phase orientation is inferred from pulse interventions on InfoVQA and ChartQA (Figure 1) and then applied unchanged to DocVQA and TextVQA. Table F shows that reversing the orientation collapses relative quality by 24.4–26.1% on LLaDA-V and 26.3–41.5% on LaViDa on the two benchmarks where the control was run. Since this binary choice is the single highest-stakes design decision, the central 1.55x/1.86% result depends on the untested assumption that the same orientation transfers across tasks for each backbone. Please provide pulse experiments or reverse-orientation controls on DocVQA and TextVQA, or otherwise justify that orientation is a backbone-level property independent of the task distribution.
  2. [§5.1 'DAVET configuration'] The paper reports one configuration per backbone, with different B (0.75 vs. 0.45) and α (1.0 vs. 0.5), and states that configurations are fixed across benchmarks. It does not explain how B, ρ, and α were selected; the sweeps in Appendix D are reported after the fact. If these values were chosen using the same four evaluation benchmarks, the headline averaged numbers reflect test-set selection rather than an a priori policy. Please specify a validation protocol (e.g., a held-out split or a small tuning set) and report the performance of the settings chosen by that procedure, or present the full trade-off frontier and describe a principled selection rule.
  3. [§4.2, Eq. (6)] The phase prior is restricted to two linear monotone profiles (early-grounding p_t = 1−(t−0.5)/T and late-verification p_t = (t−0.5)/T). The controlled pulse interventions in Figure 1 establish that high evidence is more useful in one window, but they do not establish that the optimal trajectory is a monotone ramp rather than, say, a pulse, a step, or a non-monotone shape. Because the reverse-orientation control shows the location of emphasis is decisive, the shape assumption is part of the central claim. Please justify the linear family or compare against alternative phase-shape hypotheses before concluding that DAVET captures the temporal demand pattern.
  4. [Appendix A, Table A and §5.4] The demand priors d(o) are fixed without a stated derivation, yet Table 3 shows that adding operation demand recovers 3.74–12.24 points on LaViDa, so this component is load-bearing for the overall result. The uniform-demand control in Table D shows the exact numerical values are not highly sensitive, but parser errors in classifying queries could still shift a non-trivial fraction of queries between categories. Please report parser accuracy on the evaluation benchmarks (or a sampled subset), or provide a sensitivity analysis over the demand-prior values to demonstrate that the reported behavior is robust to reasonable parser misclassification.
minor comments (5)
  1. [Figure 2] The overview figure is dense. Consider labeling the lower and upper envelope trajectories and the step indices more explicitly, and adding a small legend for the evidence views (coarse/mixed/high) so the mapping from budget to view is easier to follow.
  2. [Table 1] Static Mid on LLaDA-V/DocVQA reports a 0.98x speedup (slightly slower than Static High). The text treats this as a static control without acknowledging that a sub-1.0 speedup is within measurement noise or overhead; a brief note would avoid confusion.
  3. [Eq. (7)] The notation c̄_{t−1} is used for mean confidence, but the overline is not defined in the equation. Please state explicitly that c̄_{t−1} denotes the average over unresolved positions, consistent with the prose.
  4. [Appendix A, Table D] The column 'High' is not defined in the caption. State that it is the average number of denoising steps assigned to the high-evidence view; otherwise the reader has to infer this from the text.
  5. [§5.5 and Appendix D] The hyperparameter analysis is thorough, but the main text says only 'detailed results in Appendix D.' Since the selection of defaults is central to the reported claim, consider summarizing the boundary regimes (low B, high ρ, high α) and the resulting quality cliffs in the main text rather than only in the appendix.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the core allocation policy is derived from novel interventions and evaluated with independent test-set components.

full rationale

The paper's derivation chain is not circular. The central finding of step-dependent visual evidence demand is established by controlled pulse interventions (Figure 1) that hold total high-evidence exposure constant while shifting the pulse across denoising phases. This is an empirical measurement, not a definitional or self-referential step. The phase profile is then motivated by that measurement, and the operation-demand parser and trajectory-risk modulation are independent heuristic components. The final budgets are computed from explicit equations (Eq. 6–8) that do not reduce to the fitted inputs. The evaluation on InfoVQA/ChartQA does use the same benchmarks that informed the phase-orientation choice, which is a hyperparameter-selection caveat and could inflate those two results slightly. However, the method is also evaluated on DocVQA and TextVQA, which were not used in the orientation analysis, and the realizer-ablation and hyperparameter sweeps provide additional independent support. No load-bearing self-citation chains or uniqueness arguments are present. The main structural weakness is that the binary phase orientation is high-stakes and its transfer to new backbones/tasks is not independently validated, but that is a generality limitation, not circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim depends on several hand-set scalars (B, rho, alpha, phase orientation, demand priors, thresholds) and on domain assumptions about how well the phase profile, trajectory risk, and parser capture the true evidence demand. No new physical or model entities are introduced.

free parameters (6)
  • Average budget B = 0.75 (LLaDA-V), 0.45 (LaViDa)
    Sets the overall visual evidence scale. Chosen per backbone and appears calibrated on the evaluation benchmarks, as the paper says 'one calibrated configuration per backbone'.
  • Reserve fraction rho = 0.20 (both)
    Determines how much of B is conditionally allocated. Hand-tuned; sensitivity shown in Appendix D.
  • Risk strength alpha = 1.0 (LLaDA-V), 0.5 (LaViDa)
    Controls how strongly confidence and churn modulate reserve use. Hand-tuned per backbone.
  • Phase orientation = Early grounding for LLaDA-V, late verification for LaViDa
    Binary choice derived from the controlled interventions in Figure 1 on InfoVQA and ChartQA. If wrong, quality collapses (Appendix C).
  • Operation demand priors d(o) = [1.000, 0.941, 0.824, 0.647, 0.529]
    Hand-set values for five operation categories, normalized by the largest. Shared across backbones and benchmarks.
  • Evidence hierarchy thresholds = 1/3 and 2/3 for K=3
    Canonical thirds chosen to map the continuous budget to coarse/mixed/high views. Not justified beyond simplicity.
assumptions (5)
  • domain assumption The controlled-intervention phase analysis (Figure 1) correctly identifies backbone-specific visual evidence demand, and this pattern is stable across benchmarks.
    The entire method rests on the assumption that demand measured on InfoVQA and ChartQA transfers to DocVQA and TextVQA, and that the early/late orientation is a backbone property.
  • ad hoc to paper The linear phase family (early-grounding and late-verification profiles) is sufficient to represent the true temporal demand pattern.
    The paper tests only monotonically decreasing or increasing profiles, not more complex shapes such as U-shaped demand. Section 4.2.
  • domain assumption Trajectory risk, defined by mean confidence and prediction churn, correlates with the need for fine-grained visual evidence.
    The paper asserts that confident and stable predictions require less visual evidence, but provides no direct validation of this correlation beyond ablations. Equation (7) in Section 4.2.
  • ad hoc to paper The lexical operation parser maps query text to categories whose fixed demand priors reflect actual visual evidence requirements.
    The parser uses hand-crafted cues and hand-set priors. No learned estimator or empirical justification is given for the exact d(o) values. Appendix A.
  • domain assumption The ordered evidence hierarchy (coarse/mixed/high) provides a monotone evidence-cost interface, so lower budgets yield correspondingly lower latency with predictable quality impact.
    The paper assumes that three discrete views are enough to realize the continuous budget and that latency scales with view size. Section 4.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DAVET: Denoising-Aware Visual Evidence Trajectory Allocation for Diffusion Vision-Language Models." pith.science (2026). https://pith.science/paper/TCBOIWDF

@misc{pith2026260801821,
  author       = {Pith},
  title        = {Pith review of: DAVET: Denoising-Aware Visual Evidence Trajectory Allocation for Diffusion Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TCBOIWDF}},
  note         = {Machine review of arXiv:2608.01821}
}
abstract

Diffusion vision-language models (dVLMs) iteratively denoise masked responses while conditioning each denoising step on visual evidence, making visual conditioning a substantial recurring inference cost. Unlike autoregressive decoding, diffusion generation repeatedly revisits the entire response as uncertainty evolves. Our analysis reveals that visual evidence demand is strongly step-dependent, motivating adaptive allocation across denoising steps. Existing inference acceleration methods operate through decoding-side strategies or visual token compression via pruning and merging, but do not explicitly treat visual evidence as a resource whose demand evolves across the diffusion process. Therefore, we present Denoising-Aware Visual Evidence Trajectory Allocation (DAVET), a training-free framework that allocates visual evidence according to the evolving generation state. Starting from a phase-conditioned evidence trajectory, the proposed allocation policy uses operation demand to set an evidence reserve whose allocation at each denoising step is modulated by trajectory risk. DAVET realizes the resulting budgets through a hierarchy of evidence views constructed from a single visual encoding, separating when and how much evidence is needed from how the evidence views are constructed. Evaluated on two representative dVLMs, LLaDA-V and LaViDa, across multiple visual-understanding benchmarks, DAVET achieves an average speedup of 1.55$\times$ with an average relative performance drop of 1.86\%, showing that denoising-aware visual evidence allocation can reduce visual conditioning cost while largely preserving generation quality.

Figures

Figures reproduced from arXiv: 2608.01821 by the authors.

Figure 1
Figure 1. Backbone-dependent visual evidence demand [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of DAVET. During initialization, the image is encoded once, and the evidence realizer constructs an ordered [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Average quality–efficiency trade-offs across four [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references

  1. [1]

    Advances in Neural Information Processing Systems 36 , pages =

    Haotian Liu and Chunyuan Li and Qingyang Wu and Yong Jae Lee , title =. Advances in Neural Information Processing Systems 36 , pages =

  2. [2]

    CoRR , volume =

    Peng Wang and Shuai Bai and Sinan Tan and Shijie Wang and Zhihao Fan and Jinze Bai and Keqin Chen and Xuejing Liu and Jialin Wang and Wenbin Ge and Yang Fan and Kai Dang and Mengfei Du and Xuancheng Ren and Rui Men and Dayiheng Liu and Chang Zhou and Jingren Zhou and Junyang Lin , title =. CoRR , volume =

  3. [3]

    Making the

    Yash Goyal and Tejas Khot and Douglas Summers. Making the. Proceedings of the 30th

  4. [4]

    Proceedings of the 31st

    Peter Anderson and Xiaodong He and Chris Buehler and Damien Teney and Mark Johnson and Stephen Gould and Lei Zhang , title =. Proceedings of the 31st

  5. [5]

    Proceedings of the 17th European Conference on Computer Vision , pages =

    Geewook Kim and Teakgyu Hong and Moonbin Yim and JeongYeon Nam and Jinyoung Park and Jinyeong Yim and Wonseok Hwang and Sangdoo Yun and Dongyoon Han and Seunghyun Park , title =. Proceedings of the 17th European Conference on Computer Vision , pages =

  6. [6]

    Findings of the 61st Annual Meeting of the Association for Computational Linguistics , pages =

    Fangyu Liu and Julian Martin Eisenschlos and Francesco Piccinno and Syrine Krichene and Chenxi Pang and Kenton Lee and Mandar Joshi and Wenhu Chen and Nigel Collier and Yasemin Altun , title =. Findings of the 61st Annual Meeting of the Association for Computational Linguistics , pages =

  7. [7]

    Proceedings of the 33rd

    Ronghang Hu and Amanpreet Singh and Trevor Darrell and Marcus Rohrbach , title =. Proceedings of the 33rd

  8. [8]

    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 =

Show all 41 references
  1. [9]

    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 =

  2. [10]

    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...

  3. [11]

    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 =

  4. [12]

    Proceedings of the 39th

    Zebin You and Shen Nie and Xiaolu Zhang and Jun Zhou and Zhiwu Lu and Ji. Proceedings of the 39th

  5. [13]

    Advances in Neural Information Processing Systems 38 , pages =

    Shufan Li and Konstantinos Kallidromitis and Hritik Bansal and Akash Gokul and Yusuke Kato and Kazuki Kozuka and Jason Kuen and Zhe Lin and Kai. Advances in Neural Information Processing Systems 38 , pages =

  6. [14]

    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 =

  7. [15]

    CoRR , volume =

    Chengyue Wu and Shiyi Lan and Yonggan Fu and Sensen Gao and Jin Wang and Jincheng Yu and Jos. CoRR , volume =

  8. [16]

    CoRR , volume =

    Zhewen Wan and Tianchen Song and Chen Lin and Zhiyong Zhao and Xianpeng Lang , title =. CoRR , volume =

  9. [17]

    Beerel , title =

    Jingqi Xu and Jingxi Lu and Chenghao Li and Sreetama Sarkar and Souvik Kundu and Peter A. Beerel , title =. Proceedings of the 39th

  10. [18]

    Proceedings of the 40th Annual

    Shuochen Chang and Xiaofeng Zhang and Qingyang Liu and Li Niu , title =. Proceedings of the 40th Annual

  11. [19]

    Proceedings of the 39th

    Shufan Li and Jiuxiang Gu and Kangning Liu and Zhe Lin and Zijun Wei and Aditya Grover and Jason Kuen , title =. Proceedings of the 39th

  12. [20]

    CoRR , volume =

    Runpeng Yu and Xinyin Ma and Xinchao Wang , title =. CoRR , volume =

  13. [21]

    Advances in Neural Information Processing Systems 38 , pages =

    Ling Yang and Ye Tian and Bowen Li and Xinchen Zhang and Ke Shen and Yunhai Tong and Mengdi Wang , title =. Advances in Neural Information Processing Systems 38 , pages =

  14. [22]

    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 =

  15. [23]

    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 =

  16. [24]

    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 , pages =

  17. [25]

    Advances in Neural Information Processing Systems 37 , pages =

    Jiacheng Ye and Shansan Gong and Liheng Chen and Lin Zheng and Jiahui Gao and Han Shi and Chuan Wu and Xin Jiang and Zhenguo Li and Wei Bi and Lingpeng Kong , title =. Advances in Neural Information Processing Systems 37 , pages =

  18. [26]

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

    Shansan Gong and Ruixiang Zhang and Huangjie Zheng and Jiatao Gu and Navdeep Jaitly and Lingpeng Kong and Yizhe Zhang , title =. Proceedings of the 14th International Conference on Learning Representations , address =

  19. [27]

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

    Xinyou Wang and Zaixiang Zheng and Fei Ye and Dongyu Xue and Shujian Huang and Quanquan Gu , title =. Proceedings of the 41st International Conference on Machine Learning , pages =

  20. [28]

    CoRR , volume =

    Xiang Xia and Wuyang Zhang and Jiazheng Liu and Cheng Yan and Yanyong Zhang , title =. CoRR , volume =

  21. [29]

    Proceedings of the 18th European Conference on Computer Vision , pages =

    Chen, Liang and Zhao, Haozhe and Liu, Tianyu and Bai, Shuai and Lin, Junyang and Zhou, Chang and Chang, Baobao , title =. Proceedings of the 18th European Conference on Computer Vision , pages =

  22. [30]

    Proceedings of the 38th

    Xing, Long and Huang, Qidong and Dong, Xiaoyi and Lu, Jiajie and Zhang, Pan and Zang, Yuhang and Cao, Yuhang and He, Conghui and Wang, Jiaqi and Wu, Feng and Lin, Dahua , title =. Proceedings of the 38th

  23. [31]

    and Okuno, Tomoyuki and Nakata, Yohei and Keutzer, Kurt and Zhang, Shanghang , title =

    Zhang, Yuan and Fan, Chun-Kai and Ma, Junpeng and Zheng, Wenzhao and Huang, Tao and Cheng, Kuan and Gudovskiy, Denis A. and Okuno, Tomoyuki and Nakata, Yohei and Keutzer, Kurt and Zhang, Shanghang , title =. Proceedings of the 42nd International Conference on Machine Learning ...

  24. [32]

    CoRR , volume =

    Yulin Yuan and Hongshuo Zhao and Xiangming Meng , title =. CoRR , volume =

  25. [33]

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

    Moshe Kimhi and Nimrod Shabtay and Raja Giryes and Chaim Baskin and Eli Schwartz , title =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics , pages =

  26. [34]

    CoRR , volume =

    Huanxuan Liao and Zhongtao Jiang and Yupu Hao and Yuqiao Tan and Shizhu He and Ben Wang and Jun Zhao and Kun Xu and Kang Liu , title =. CoRR , volume =

  27. [35]

    CoRR , volume =

    Xiaoyu Liang and Chaofeng Guan and Jiaying Lu and Huiyao Chen and Huan Wang and Haoji Hu , title =. CoRR , volume =

  28. [36]

    Advances in Neural Information Processing Systems 38 , pages =

    Qiong Wu and Wenhao Lin and Yiyi Zhou and Weihao Ye and Zhanpeng Zeng and Xiaoshuai Sun and Rongrong Ji , title =. Advances in Neural Information Processing Systems 38 , pages =

  29. [37]

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

    Hengbo Xu and Shengjie Jin and Yanbiao Ma and Zhiwu Lu , title =. Proceedings of the 43rd International Conference on Machine Learning , address =

  30. [38]

    InfographicVQA , booktitle =

    Minesh Mathew and Viraj Bagal and Rub. InfographicVQA , booktitle =

  31. [39]

    Joty and Enamul Hoque , title =

    Ahmed Masry and Do Xuan Long and Jia Qing Tan and Shafiq R. Joty and Enamul Hoque , title =. Findings of the 60th Annual Meeting of the Association for Computational Linguistics , pages =

  32. [40]

    Minesh Mathew and Dimosthenis Karatzas and C. V. Jawahar , title =. Proceedings of the 21st

  33. [41]

    Proceedings of the 32nd

    Amanpreet Singh and Vivek Natarajan and Meet Shah and Yu Jiang and Xinlei Chen and Dhruv Batra and Devi Parikh and Marcus Rohrbach , title =. Proceedings of the 32nd

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.