Pith. sign in

REVIEW 3 major objections 4 minor 52 references

ET-Prune: Evidence-Aware Dynamic Budgeting for Visual Token Pruning in Text-Rich MLLMs

T0 review · 3 major / 4 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 A training-free token-pruning method that allocates visual tokens per question preserves OCR accuracy at about half the tokens, leading or tying all six core baseline comparisons.

desk verdict A coherent training-free pruner with a plausible new budgeting mechanism, but the paper never isolates the dynamic budgeter from its reader/safeguard, so the core claim rests on a missing control. read the letter →

arxiv 2608.01979 v1 pith:EC2YM236 submitted 2026-08-03 cs.CV cs.CL

Zizhong Ding , Junxian Li , Kai Liu , Shaoqiu Zhang , Xiao Xiao , Linghe Kong , Yulun Zhang This is my paper
classification cs.CVcs.CL
keywords visualtokenpruningmultimodallargelanguagemodelsdynamicbudgetingquestion-conditionedevidenceOCR-centricVQAtext-richimagestraining-freeaccelerationheterogeneity
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

ET-Prune claims that visual token pruning for text-rich multimodal models should allocate tokens per input, not per fixed ratio. It derives question-conditioned evidence from a decoder-side partial query-key block, protects text-like regions, and converts evidence uncertainty and density into a sample-specific token floor. At roughly half the visual tokens, this training-free framework leads or ties all six core backbone-benchmark comparisons among pruned methods, with observed gains up to +1.80 percentage points on OCRBench-v2 and a higher score than the unpruned model on MMBench v1.1. The result matters because fixed budgets either waste tokens on easy images or erase small answer-bearing text on hard ones.

What carries the argument

The framework's three carrying components: (1) a partial query-key reader that computes a softmax over all mask-permitted keys and slices the visual columns, yielding a question-conditioned distribution over active visual tokens; (2) a text-evidence safeguard that maps a lightweight contrast/edge prior to protected token support and applies a protected-first selection rule; (3) a risk-calibrated budgeter that combines normalized entropy H(a), text-like density D_p, and protected ratio R_p into a per-sample adjustment delta, producing endpoint K*. A progressive middle-layer schedule interpolates token counts from N to K* while maintaining nested selected sets. The reader's partial-QK slice ke

What would settle it

Run ET-Prune with the budget fixed at a constant r0 (set delta=0) while keeping the same reader and safeguard, and match the average retention to the full method; if the scores match, the dynamic budget contributes nothing, and if they drop, it is essential.

Watch

Extended reading notes

Core claim

The central discovery is that pruning for text-rich MLLMs should be framed as evidence allocation: the question determines which visual regions carry the answer, and the amount of evidence varies per image. By reading question-to-visual affinity from a partial QK block, the method scores tokens by answer relevance; a spatial safeguard forces retention of text-like components; and a risk-calibrated budgeter raises the retained floor when question evidence is diffuse, text-dense, or protected regions are prevalent. Three progressive middle-layer events then compact the visual sequence toward this floor. The observed point estimates show the method preserves more quality than representative pru

Load-bearing premise

The core claim rests on the linear risk model in Eq. 7, whose weights are tuned on a 500-example TextVQA subset, and on the unverified assumption that it is the sample-specific budget, rather than the evidence reader and safeguard, that produces the quality savings.

Editorial extensions

If this is right

  • Replacing a fixed global keep ratio with a per-example floor derived from question-conditioned evidence and text density preserves more OCR quality at comparable average token counts.
  • The method is training-free: model weights and decoding are untouched, so it can be applied to any frozen MLLM without fine-tuning.
  • At roughly half the visual tokens, the observed gains include a 1.80 pp improvement over the strongest pruned baseline on OCRBench-v2 for one 8B backbone and a 0.68 pp improvement for another, with average retention near 50%.

Reading between the lines

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

  • The paper does not isolate the dynamic budget from the reader and safeguard; a natural test is to fix the budget to the same average retention. If accuracy holds, the dynamic adjustment is unnecessary; if it drops, the budgeter is the causal engine.
  • The budgeter's weights are tuned on a 500-example TextVQA subset, so transferability to new resolutions or tokenizers is an open empirical question; an extension would be to learn the weights per image-scale class.
  • The question-conditioned partial-QK reader could be reused outside pruning, such as for attention-based explanation or test-time selection of key regions, since it already answers which visual tokens carry evidence for the question.
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

3 major / 4 minor

Summary. The paper proposes ET-Prune, a training-free visual-token pruning framework for text-rich multimodal LLMs. It combines a decoder-side partial-QK question-conditioned evidence reader, a lightweight text-region safeguard, a risk-calibrated dynamic budgeter that sets a per-image token floor, and progressive middle-layer compaction. The central empirical claim is that at roughly half visual-token retention, ET-Prune leads or ties all pruned baselines across six backbone-benchmark comparisons, with reported gains such as +1.80 pp on Qwen3-VL-8B OCRBench-v2 and +0.68 pp on InternVL3.5-8B, and a MMBench v1.1 score of 0.8467 versus 0.8437 for Vanilla at 54.45% average retention. The paper also reports transfer results on MMVP, RoboVQA, RealWorldQA, MME, and MMBench, and includes a LLaVA-1.5-7B compatibility diagnostic.

Significance. If the dynamic budgeter is genuinely responsible for the reported gains, ET-Prune is a useful training-free contribution: it addresses a real limitation of fixed-ratio pruning on text-rich inputs, is clearly specified, and the authors are transparent about single-pass point estimates and parameter tuning. The text-evidence safeguard and question-conditioned scoring are plausible mechanisms, and code is released. However, the core novelty is the per-image budget, and the current experiments do not isolate it from the reader/safeguard components. Given small margins on some comparisons and tuning on a TextVQA subset, the evidence for the dynamic-budgeting contribution is not yet load-bearing.

major comments (3)
  1. [Methodology, Eq. (7)-(8); Experiments, Table 3] The central claim concerns dynamic budgeting, but no experiment isolates the budgeter. Table 3 varies the target ratio R60/R50/R40 with the dynamic budgeter always active; it does not compare against a fixed-budget version of ET-Prune using the same reader and safeguard (e.g., setting w_h=w_d=w_r=0 so Δ=0 and r_eff=r0 at matched average retention). Without this control, the observed gains over FastV, PyramidDrop, SparseVLMs, IVC-Prune, and ZOO-Prune could be entirely due to the evidence reader and text safeguard, with the dynamic component neutral or harmful. Please add a fixed-budget ablation at comparable average retention and report both quality and FLOPs.
  2. [Experiments, Experimental Setup and Table 1] The budgeter weights and related hyperparameters are selected on a fixed 500-example TextVQA subset that is disjoint from the three core evaluation sets but drawn from the same distribution as a reported benchmark. Since the InternVL3.5-8B TextVQA gain is only +0.50 pp and the OCRBench-mini comparison is an exact tie, selection on this subset can materially inflate the TextVQA result. Please provide sensitivity analysis to the calibration subset (e.g., cross-validation or multiple disjoint subsets) and report the operating points of the budgeter on the evaluation sets.
  3. [Experiments, Table 1 and Abstract] All scores are deterministic single-pass point estimates with no significance bounds; the manuscript honestly disclaims statistical-significance claims. Nevertheless, the abstract states that ET-Prune 'leads or ties' in all six comparisons, and several margins are small. To support this claim, please include uncertainty estimates (bootstrap over samples, or multiple decoding runs) or explicitly soften the claim in the abstract and conclusions to 'observed point estimates lead or tie'. The current wording overstates the robustness of the empirical result.
minor comments (4)
  1. [Methodology, Eq. (8)] The clipping in Eq. (8) forces r_eff >= r0 for every image; since Eq. (1) constrains the average budget, clarify how the average retention is kept near the target when the floor is always at least r0. If r0 is a base ratio and Δ only increases retention, this should be stated explicitly.
  2. [Methodology, Eq. (10)] The notation L_s and K_s is used without fully defining the stage partition. Please define L_s as the number of decoder layers between pruning events and K_s as the visual token count after event s, or add a sentence explaining the indexing.
  3. [Experiments, Table 2] Table 2 includes the Vanilla row, but the bold/underline convention applies only to pruned rows. The caption should state this explicitly to avoid ambiguity, particularly since ET-Prune's MMBench score exceeds the Vanilla score.
  4. [Limitations] The limitations paragraph is candid about model-specific bookkeeping and the need for retuning at new resolutions/tokenizers. It would be helpful to state whether the dynamic budgeter's weights were retuned for the LLaVA-1.5-7B compatibility implementation, since that variant uses a pre-LLM selector and is not part of the decoder-side claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ET-Prune's claims are supported by held-out evaluation, and the dynamic budgeter is a fitted heuristic rather than a self-referential prediction.

full rationale

The paper's central claim is an empirical one: at roughly half the visual tokens, ET-Prune leads or ties pruned baselines across six backbone–benchmark comparisons. This claim is backed by measured accuracies on held-out evaluation sets. The dynamic budget weights (w_h, w_d, w_r in Eq. 7) are tuned on a fixed 500-example TextVQA subset that the paper explicitly states is disjoint from the three core evaluation sets, and the reported scores are observed point estimates from full-split passes, not outputs reconstructed from the fitted parameters. The budgeter determines a token floor per image, but it does not directly predict accuracy; the accuracy is obtained by running the frozen model on the pruned sequence. Thus there is no reduction of the central result to the fitted inputs by construction. The absence of a fixed-budget ablation of ET-Prune itself is an experimental isolation concern, not a circularity: it weakens attribution of gains to the dynamic budgeter, but it does not make any claimed number equal to an input by definition. Citations to prior work, including the authors' own G2TR reference, are contextual and not load-bearing; no uniqueness theorem or ansatz is imported from self-citations. The derivation chain consists of scoring, safeguarding, and budgeting equations whose outputs are actual model inputs, not the reported accuracies. No self-definitional step, renamed known result, or fitted-input-called-prediction was found.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

No new particles, forces, conserved quantities, or other entities are introduced. The 'evidence floor K*' is a per-input control variable, not a new entity. The paper's load-bearing structure is a set of hand-designed signals and weights, listed above.

free parameters (7)
  • base ratio r0
    Calibrated 'base rate' in Eq. 8; tuned on a 500-example TextVQA subset.
  • weights w_h, w_d, w_r
    Weights of entropy, text density, and protected ratio in Eq. 7; selected on the 500-example TextVQA validation subset.
  • max adjustment Δmax
    Caps the risk adjustment in Eq. 7; hand-set.
  • max ratio rmax
    Upper clip for the retention ratio in Eq. 8; hand-set.
  • minimum token guard B
    Lower bound on K* in Eq. 8; hand-set.
  • event layers and progress coefficients alpha_j
    Middle-layer pruning schedule (e.g., layers 8, 16, 24 in Figure 6); schedule adopted or adapted from PyramidDrop.
  • text-safeguard thresholds
    Local contrast and edge parameters that define the pixel mask T in Eq. 4; not specified in the main text.
assumptions (4)
  • domain assumption Question-to-visual partial QK attention at a decoder layer is a reliable question-conditioned evidence score.
    Equations 2-3 use only the question queries to read visual keys; no proof that this maps to answer evidence, only empirical support.
  • ad hoc to paper Normalized entropy of pi plus text coverage and protected ratio predicts the token budget needed to preserve evidence.
    Eq. 7-8 define an additive linear risk adjustment with tuned weights; no derivation links H(a), D_p, R_p to required capacity.
  • domain assumption Local contrast and edge responses identify text-like regions well enough to protect answer-bearing spans.
    Eq. 4 builds a pixel mask from local contrast and edges; Limitations admit 'this lightweight spatial prior is not a text detector.'
  • domain assumption Progressive middle-layer compaction preserves quality at reduced tokens.
    Borrowed from PyramidDrop (Xing et al., 2024) and assumed rather than re-derived for text-rich inputs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ET-Prune: Evidence-Aware Dynamic Budgeting for Visual Token Pruning in Text-Rich MLLMs." pith.science (2026). https://pith.science/paper/EC2YM236

@misc{pith2026260801979,
  author       = {Pith},
  title        = {Pith review of: ET-Prune: Evidence-Aware Dynamic Budgeting for Visual Token Pruning in Text-Rich MLLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EC2YM236}},
  note         = {Machine review of arXiv:2608.01979}
}
read the original abstract

Visual token pruning reduces the inference cost of multimodal large language models, but a fixed token ratio is poorly matched to text-rich inputs. In OCR-centric tasks, decisive evidence can be a small number, label, or field whose relevance is specified by the question; indiscriminate pruning can erase that evidence while retaining visually salient but irrelevant regions. We present ET-Prune, a training-free framework that casts pruning as evidence allocation. It derives question-conditioned evidence from a decoder-side partial query-key block, safeguards text-like spatial regions, and converts evidence uncertainty and density into a sample-specific token floor. Three progressive middle-layer events then move the sequence toward this budget, retaining more tokens for diffuse or text-dense evidence and pruning concentrated evidence more aggressively. At the observed point estimates from one deterministic pass per configuration, ET-Prune leads or ties among pruned methods in all six backbone-benchmark comparisons at roughly half tokens. On OCRBench-v2, it leads the strongest pruned baselines by 1.80 and 0.68 percentage points on Qwen3-VL-8B and InternVL3.5-8B, respectively, while retaining about half of the visual tokens; on MMBench v1.1, it reaches 0.8467 circular exact-matching accuracy versus 0.8437 for Vanilla at 54.45% average visual-token retention. These results show a favorable observed quality-cost trade-off for evidence-aware dynamic budgeting in text-rich multimodal inference.

Figures

Figures reproduced from arXiv: 2608.01979 by the authors.

Figure 1
Figure 1. Overview and quality–cost trade-off of ET-Prune. Left: question-conditioned evidence scoring, text-region safeguard [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of ET-Prune. A partial-QK reader scores question-conditioned evidence, a spatial safeguard protects text-like [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Motivating examples. (a) Full-token posterior [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Evidence heterogeneity across 15,194 Qwen3- [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Dynamic budget response on Qwen3-VL. Red: [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Qualitative TextVQA comparison. Top: ET-Prune [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 12 linked inside Pith

  1. [1]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Singh, Amanpreet and Natarajan, Vivek and Shah, Meet and Jiang, Yu and Chen, Xinlei and Batra, Dhruv and Parikh, Devi and Rohrbach, Marcus , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2019 , doi =

    2019

  2. [2]

    Scene Text Visual Question Answering , booktitle =

    Biten, Ali Furkan and Tito, Rub. Scene Text Visual Question Answering , booktitle =. 2019 , doi =

    2019

  3. [3]

    2024 , howpublished =

    Announcing Grok-1.5 , author =. 2024 , howpublished =

    2024

  4. [4]

    2019 International Conference on Document Analysis and Recognition (ICDAR) , pages =

    Mishra, Anand and Shekhar, Shashank and Singh, Ajeet Kumar and Chakraborty, Anirban , title =. 2019 International Conference on Document Analysis and Recognition (ICDAR) , pages =. 2019 , doi =

    2019

  5. [5]

    Mathew, Minesh and Karatzas, Dimosthenis and Jawahar, C. V. , title =. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages =. 2021 , doi =

    2021

  6. [6]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Chemvlm: Exploring the power of multimodal large language models in chemistry area , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  7. [7]

    arXiv preprint arXiv:2605.12309 , year=

    G ^2 TR: Generation-Guided Visual Token Reduction for Separate-Encoder Unified Multimodal Models , author=. arXiv preprint arXiv:2605.12309 , year=

    arXiv

  8. [8]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages =

    Mathew, Minesh and Bagal, Viraj and Tito, Rub. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages =. 2022 , doi =

    2022

Show all 52 references
  1. [9]

    Findings of the Association for Computational Linguistics: ACL 2022 , pages =

    Masry, Ahmed and Long, Do Xuan and Tan, Jia Qing and Joty, Shafiq and Hoque, Enamul , title =. Findings of the Association for Computational Linguistics: ACL 2022 , pages =. 2022 , doi =

    2022

  2. [10]

    Computer Vision -- ECCV 2020 , pages =

    Sidorov, Oleksii and Hu, Ronghang and Rohrbach, Marcus and Singh, Amanpreet , title =. Computer Vision -- ECCV 2020 , pages =. 2020 , doi =

    2020

  3. [11]

    Science China Information Sciences , volume =

    Liu, Yuliang and Li, Zhang and Huang, Mingxin and Yang, Biao and Yu, Wenwen and Li, Chunyuan and Yin, Xu-Cheng and Liu, Cheng-Lin and Jin, Lianwen and Bai, Xiang , title =. Science China Information Sciences , volume =. 2024 , doi =

    2024

  4. [12]

    arXiv preprint arXiv:2501.00321 , year =

    Fu, Ling and Kuang, Zhebin and Song, Jiajun and Huang, Mingxin and Yang, Biao and Li, Yuzhe and Zhu, Linghao and Luo, Qidi and Wang, Xinyu and Lu, Hao and Li, Zhang and Tang, Guozhi and Shan, Bin and Lin, Chunhui and Liu, Qi and Wu, Binghong and Feng, Hao and Liu, Hao and Huan...

    arXiv

  5. [13]

    Computer Vision -- ECCV 2024 , pages =

    Liu, Yuan and Duan, Haodong and Zhang, Yuanhan and Li, Bo and Zhang, Songyang and Zhao, Wangbo and Yuan, Yike and Wang, Jiaqi and He, Conghui and Liu, Ziwei and Chen, Kai and Lin, Dahua , title =. Computer Vision -- ECCV 2024 , pages =. 2024 , doi =

    2024

  6. [14]

    arXiv preprint arXiv:2306.17107 , year =

    Zhang, Yanzhe and Zhang, Ruiyi and Gu, Jiuxiang and Zhou, Yufan and Lipka, Nedim and Yang, Diyi and Sun, Tong , title =. arXiv preprint arXiv:2306.17107 , year =. 2306.17107 , archivePrefix =

    arXiv

  7. [15]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages =

    Ye, Jiabo and Hu, Anwen and Xu, Haiyang and Ye, Qinghao and Yan, Ming and Xu, Guohai and Li, Chenliang and Tian, Junfeng and Qian, Qi and Zhang, Ji and Jin, Qin and He, Liang and Lin, Xin and Huang, Fei , title =. Findings of the Association for Computational Linguistics: EMNL...

    2023

  8. [16]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages =

    Hu, Anwen and Xu, Haiyang and Ye, Jiabo and Yan, Ming and Zhang, Liang and Zhang, Bo and Zhang, Ji and Jin, Qin and Huang, Fei and Zhou, Jingren , title =. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages =. 2024 , doi =

    2024

  9. [17]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Li, Zhang and Yang, Biao and Liu, Qiang and Ma, Zhiyin and Zhang, Shuo and Yang, Jingxu and Sun, Yabo and Liu, Yuliang and Bai, Xiang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2024 , doi =

    2024

  10. [18]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Zhang, Jiaxin and Yang, Wentao and Lai, Songxuan and Xie, Zecheng and Jin, Lianwen , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2025 , doi =

    2025

  11. [19]

    arXiv preprint arXiv:2511.21631 , year =

    Shuai Bai and Yuxuan Cai and Ruizhe Chen and Keqin Chen and Xionghui Chen and Zesen Cheng and Lianghao Deng and Wei Ding and Chang Gao and Chunjiang Ge and Wenbin Ge and Zhifang Guo and Qidong Huang and Jie Huang and Fei Huang and Binyuan Hui and Shutong Jiang and Zhaohai Li a...

    arXiv

  12. [20]

    arXiv preprint arXiv:2508.18265 , year =

    Weiyun Wang and Zhangwei Gao and Lixin Gu and Hengjun Pu and Long Cui and Xingguang Wei and Zhaoyang Liu and Linglin Jing and Shenglong Ye and Jie Shao and Zhaokai Wang and Zhe Chen and Hongjie Zhang and Ganlin Yang and Haomin Wang and Qi Wei and Jinhui Yin and Wenhao Li and E...

    arXiv

  13. [21]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Liu, Haotian and Li, Chunyuan and Li, Yuheng and Lee, Yong Jae , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2024 , doi =

    2024

  14. [22]

    Computer Vision -- ECCV 2024 , pages =

    Chen, Liang and Zhao, Haozhe and Liu, Tianyu and Bai, Shuai and Lin, Junyang and Zhou, Chang and Chang, Baobao , title =. Computer Vision -- ECCV 2024 , pages =. 2024 , doi =

    2024

  15. [23]

    arXiv preprint arXiv:2410.17247 , year =

    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 =. arXiv preprint arXiv:2410.17247 , year =. 2410.17247 , archivePrefix =

    arXiv

  16. [24]

    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 ...

    2025

  17. [25]

    The Fourteenth International Conference on Learning Representations , year =

    Sun, Zhichao and Ma, Yidong and Liu, Gang and Chen, Yibo and Tang, Xu and Hu, Yao and Xu, Yongchao , title =. The Fourteenth International Conference on Learning Representations , year =

  18. [26]

    arXiv preprint arXiv:2509.24837 , year =

    Kim, Youngeun and Zhang, Youjia and Liu, Huiling and Jung, Aecheon and Lee, Sunwoo and Hong, Sungeun , title =. arXiv preprint arXiv:2509.24837 , year =. 2509.24837 , archivePrefix =

  19. [27]

    The Thirty-Ninth Annual Conference on Neural Information Processing Systems , year =

    Shao, Kele and Tao, Keda and Qin, Can and You, Haoxuan and Sui, Yang and Wang, Huan , title =. The Thirty-Ninth Annual Conference on Neural Information Processing Systems , year =

  20. [28]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Shang, Yuzhang and Cai, Mu and Xu, Bingxin and Lee, Yong Jae and Yan, Yan , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2025 , doi =

    2025

  21. [29]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages =

    Yang, Senqiao and Chen, Yukang and Tian, Zhuotao and Wang, Chengyao and Li, Jingyao and Yu, Bei and Jia, Jiaya , title =. Proceedings of the Computer Vision and Pattern Recognition Conference , pages =. 2025 , doi =

    2025

  22. [30]

    arXiv preprint arXiv:2411.10803 , year =

    Liu, Ting and Shi, Liangtao and Hong, Richang and Hu, Yue and Yin, Quanjun and Zhang, Linfeng , title =. arXiv preprint arXiv:2411.10803 , year =. 2411.10803 , archivePrefix =

    arXiv

  23. [31]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Cao, Jianjian and Ye, Peng and Li, Shengze and Yu, Chong and Tang, Yansong and Lu, Jiwen and Chen, Tao , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2024 , doi =

    2024

  24. [32]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Ye, Weihao and Wu, Qiong and Lin, Wenhao and Zhou, Yiyi , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2025 , doi =

    2025

  25. [33]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Lin, Zhihang and Lin, Mingbao and Lin, Luxi and Ji, Rongrong , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2025 , doi =

    2025

  26. [34]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages =

    Alvar, Saeed Ranjbar and Singh, Gursimran and Akbari, Mohammad and Zhang, Yong , title =. Proceedings of the Computer Vision and Pattern Recognition Conference , pages =. 2025 , doi =

    2025

  27. [35]

    arXiv preprint arXiv:2411.14228 , year =

    Zhu, Yuke and Xie, Chi and Liang, Shuang and Zheng, Bo and Guo, Sheng , title =. arXiv preprint arXiv:2411.14228 , year =. 2411.14228 , archivePrefix =

    arXiv

  28. [36]

    The Fourteenth International Conference on Learning Representations , year =

    Dong, Sixun and Hu, Juhua and Zhang, Mian and Yin, Ming and Fu, Yanjie and Qian, Qi , title =. The Fourteenth International Conference on Learning Representations , year =

  29. [37]

    arXiv preprint arXiv:2507.20630 , year =

    Li, Ao and Duan, Yuxiang and Zhang, Jinghui and Ma, Congbo and Xie, Yutong and Carneiro, Gustavo and Yaqub, Mohammad and Wang, Hu , title =. arXiv preprint arXiv:2507.20630 , year =. 2507.20630 , archivePrefix =

  30. [38]

    and Piergiovanni, A

    Ryoo, Michael S. and Piergiovanni, A. J. and Arnab, Anurag and Dehghani, Mostafa and Angelova, Anelia , title =. Advances in Neural Information Processing Systems , volume =. 2021 , url =

    2021

  31. [39]

    The Eleventh International Conference on Learning Representations , year =

    Bolya, Daniel and Fu, Cheng-Yang and Dai, Xiaoliang and Zhang, Peizhao and Feichtenhofer, Christoph and Hoffman, Judy , title =. The Eleventh International Conference on Learning Representations , year =

  32. [40]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    Cao, Qingqing and Paranjape, Bhargavi and Hajishirzi, Hannaneh , title =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2023 , doi =

    2023

  33. [41]

    arXiv preprint arXiv:1603.08983 , year =

    Graves, Alex , title =. arXiv preprint arXiv:1603.08983 , year =. 1603.08983 , archivePrefix =

    arXiv

  34. [42]

    and Grauman, Kristen and Feris, Rogerio , title =

    Wu, Zuxuan and Nagarajan, Tushar and Kumar, Abhishek and Rennie, Steven and Davis, Larry S. and Grauman, Kristen and Feris, Rogerio , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2018 , doi =

    2018

  35. [43]

    , title =

    Wang, Xin and Yu, Fisher and Dou, Zi-Yi and Darrell, Trevor and Gonzalez, Joseph E. , title =. Computer Vision -- ECCV 2018 , pages =. 2018 , doi =

    2018

  36. [44]

    Advances in Neural Information Processing Systems , volume =

    Hou, Lu and Huang, Zhiqi and Shang, Lifeng and Jiang, Xin and Chen, Xiao-Dong and Liu, Qun , title =. Advances in Neural Information Processing Systems , volume =. 2020 , url =

    2020

  37. [45]

    Advances in Neural Information Processing Systems , volume =

    Rao, Yongming and Zhao, Wenliang and Liu, Benlin and Lu, Jiwen and Zhou, Jie and Hsieh, Cho-Jui , title =. Advances in Neural Information Processing Systems , volume =. 2021 , url =

    2021

  38. [46]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Yin, Hongxu and Vahdat, Arash and. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2022 , doi =

    2022

  39. [47]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Meng, Lingchen and Li, Hengduo and Chen, Bor-Chun and Lan, Shiyi and Wu, Zuxuan and Jiang, Yu-Gang and Lim, Ser-Nam , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2022 , doi =

    2022

  40. [48]

    and Tay, Yi and Metzler, Donald , title =

    Schuster, Tal and Fisch, Adam and Gupta, Jai and Dehghani, Mostafa and Bahri, Dara and Tran, Vinh Q. and Tay, Yi and Metzler, Donald , title =. Advances in Neural Information Processing Systems , volume =. 2022 , doi =

    2022

  41. [49]

    and Lillicrap, Timothy and Humphreys, Peter C

    Raposo, David and Ritter, Sam and Richards, Blake A. and Lillicrap, Timothy and Humphreys, Peter C. and Santoro, Adam , title =. arXiv preprint arXiv:2404.02258 , year =. 2404.02258 , archivePrefix =

    arXiv

  42. [50]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Tong, Shengbang and Liu, Zhuang and Zhai, Yuexiang and Ma, Yi and LeCun, Yann and Xie, Saining , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2024 , doi =

    2024

  43. [51]

    Sermanet, Pierre and Ding, Tianli and Zhao, Jeffrey and Xia, Fei and Dwibedi, Debidatta and Gopalakrishnan, Keerthana and Chan, Christine and Dulac-Arnold, Gabriel and Maddineni, Sharath and Joshi, Nikhil J. and Florence, Pete and Han, Wei and Baruch, Robert and Lu, Yao and Mi...

    arXiv

  44. [52]

    arXiv preprint arXiv:2306.13394 , year =

    Fu, Chaoyou and Chen, Peixian and Shen, Yunhang and Qin, Yulei and Zhang, Mengdan and Lin, Xu and Yang, Jinrui and Zheng, Xiawu and Li, Ke and Sun, Xing and Wu, Yunsheng and Ji, Rongrong and Shan, Caifeng and He, Ran , title =. arXiv preprint arXiv:2306.13394 , year =. 2306.13...

    arXiv

Pith tools

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