Pith. sign in

REVIEW 4 major objections 3 minor 2 cited by

By protecting query-critical visual tokens before adding context, visual-token pruning retains 97.6% of accuracy with only 5.6% of the tokens, a result that holds across image and video vision-language models.

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 ↗

AnchorPrune prunes visual tokens by first selecting a protected query-relevance anchor and then greedily adding important, non-redundant context, preserving up to 97.6% of full-token accuracy with only 160 of 2,880 tokens.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A solid training-free pruning method whose ordered anchor-then-expand design is new and empirically consistent; main risks are unreported variance and the under-motivated negated-CLIP anchor. the 4 major comments →

arxiv 2607.07033 v4 pith:HOUWORDW submitted 2026-07-08 cs.CV cs.AI

AnchorPrune: Relevance-Anchored Contextual Expansion for Visual Token Pruning

Kyuan Oh , Bumsoo Kim This is my paper
classification cs.CV cs.AI
keywords visual token pruningvision-language modelstraining-free compressionquery relevancetoken diversitycontextual expansionefficient multimodal inference
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

AnchorPrune is a training-free visual-token pruning method built on a single ordering claim: query-critical evidence and supporting context are asymmetric, so pruning should first build a protected 'relevance anchor' from tokens ranked by query-conditioned priority, and only then expand with informative, non-redundant context. Existing methods that optimize relevance and diversity jointly can let the diversity term displace the small, non-substitutable regions the question actually needs; a separate anchor prevents that. The anchor size is chosen adaptively from how quickly novel tokens appear in the ranked sequence, and the remaining budget is spent on tokens that are both globally important and novel relative to the anchor. On LLaVA-NeXT-7B the method keeps 97.6% of full-token performance with only 160 of 2,880 visual tokens, and across image and video VLMs it beats training-free baselines by the widest margin at the most aggressive compression levels. If the paper is right, the ordering principle—not the particular scoring formula—is what carries the gains.

Core claim

The paper's central claim is that query-critical visual evidence and supporting context are asymmetric and should be selected in a fixed order, not jointly optimized. AnchorPrune first ranks tokens with an architecture-specific query-conditioned score (negated CLIP similarity for CLIP-aligned models, max post-projector cosine for others), builds a protected anchor whose size is set adaptively by counting novelty events in the ranked sequence, and only then fills the remaining budget with tokens that maximize importance-weighted novelty relative to the current selected set. This ordered design prevents the contextual expansion stage from displacing indispensable query tokens, which are treate

What carries the argument

The protected relevance anchor plus adaptive anchoring budget: tokens are ranked by an architecture-specific priority score (Eq. 7 or Eq. 9); the anchor size K_rel is chosen as the first position where P novel tokens (novelty = min cosine-distance to the initial anchor, threshold τ) have been seen, capped at half the total budget (Eqs. 12-16). Stage 2 then greedily adds tokens maximizing p_i · Δ(i;S): global importance prior p_i (CLS-attention or received-attention mass, Eqs. 18-20) times novelty to the current retained set (Eq. 23). The anchor moves the non-substitutable evidence out of competition with diversity.

Load-bearing premise

The paper's load-bearing premise is that the Stage-1 priority score ranks the truly query-critical tokens at the top (and the P-th novelty event marks when they have been adequately covered) — if that ranking is wrong, the protected anchor contains the wrong tokens and the Stage-2 expansion cannot recover them.

What would settle it

A concrete test: construct or find images where a small, query-critical region (e.g., a sign, a number, a face) is known to be necessary for the correct answer. Compute the Stage-1 score for all tokens. If, across many such examples, the rank of the ground-truth-necessary tokens is systematically below the rank of background tokens that the model would not use, the anchor will consist of the wrong tokens, and the claimed accuracy retention should fail on those examples. The claim would be settled by measuring the correlation between Stage-1 rank and ground-truth 'query-critical' masks on a ben

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

If this is right

  • At 160 of 2,880 visual tokens on LLaVA-NeXT-7B, AnchorPrune retains 97.6% of full-token accuracy, 4.7 points above the best training-free baseline at the same budget.
  • The ordering principle transfers across very different vision encoders: a CLIP-aligned image model, a non-CLIP image model, and a 16-frame video model all show their best retained performance with the same two-stage rule and shared hyperparameters (τ=0.2, P=3).
  • The ablation isolates the Stage-2 rule: starting from the same protected anchor, importance-weighted contextual expansion beats diversity-only, additive, and multiplicative relevance-diversity couplings by 3.7 points at 64 retained tokens, so the gains are attributed to the ordered design rather than the scoring.
  • Efficiency is not sacrificed for accuracy: with 32 tokens retained, AnchorPrune matches the fastest baselines in prefill latency (28.2 ms, 5.13× speedup) and FLOPs while achieving the highest MME score among pruning methods.

Where Pith is reading between the lines

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

  • Inference: The protected-anchor stage is a modular slot; any query-aware relevance score, including one trained or distilled for the task, could replace the negated CLIP similarity and likely inherit the same ordering benefit, since the ablations suggest ordering is the main driver.
  • Inference: Because the anchor is built per image unit/frame, temporally scattered query evidence in long videos might be under-protected; an extension that anchors on query-critical events across the temporal axis is a natural testable upgrade.
  • Inference: The adaptive stopping rule (P-th novelty event) is a proxy for 'enough query evidence has been seen'; a more continuous estimate of remaining evidence mass could reduce sensitivity to the threshold and patience settings.
  • Inference: The method is compatible with position-sensitive decoding, as shown by the raster-order restoration experiment; this suggests anchor construction could be made spatially coherent (e.g., entire text lines or objects) rather than per-token, which might further help OCR-heavy tasks.
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 presents AnchorPrune, a training-free visual-token pruning framework for vision-language models. It first constructs a protected relevance anchor using architecture-specific query-conditioned scores (negated CLIP patch–text similarity for CLIP-aligned models, Eq. 7; max post-projector cosine for non-CLIP models, Eq. 9), with the anchor size determined adaptively from the novelty profile of the relevance-ranked sequence (Eqs. 14–16). Stage 2 then expands the retained set through importance-weighted novelty relative to the current selection (Eq. 23). The method is evaluated on LLaVA-1.5-7B, LLaVA-NeXT-7B, Qwen2.5-VL-7B, and LLaVA-Video-7B under aggressive token budgets, reporting consistent improvements in relative retained performance over training-free baselines, with a headline result of 97.6% retained performance at 160 of 2,880 visual tokens on LLaVA-NeXT-7B.

Significance. If the claims hold, the paper offers a simple, architecture-aware principle—protect non-substitutable query-critical evidence before expanding with informative context—that improves accuracy–efficiency trade-offs without retraining or model modification. The work is well positioned as an empirical contribution: code is released, all methods are evaluated under matched retained-token budgets, the Stage-2 ablation in Table 5 isolates the expansion criterion, and the efficiency measurements in Table S8 include token-selection overhead. The main risk is that the Stage-1 anchor for the two CLIP-aligned backbones is built from a sign-flipped similarity score that the paper itself declines to interpret as calibrated semantic relevance, and several adaptive components were selected using MME ablations on the same backbones that appear in the headline tables. These issues are addressable with additional analysis, but they are load-bearing for the central ordering claim.

major comments (4)
  1. [§3.2, Eq. (7); Supp. C.3, Table S6] The protected relevance anchor for CLIP-aligned models is constructed from negated CLIP patch–text similarity. The paper explicitly disclaims that this score is calibrated semantic relevance, and Table S6 shows the sign choice is decisive: at 32 tokens on LLaVA-1.5-7B, raw similarity gives MME 1245.4 while negated similarity gives 1394.7. Because the Stage-2 objective in Eq. (23) contains no query term—p_i is CLS-attention importance and Δ(i;S) is visual novelty—a misranked Stage-1 anchor cannot be repaired by expansion. The ECLIP citation supports raw CLIP maps being background-heavy for visualization, but not the stronger claim that the negated map ranks query-critical, non-substitutable tokens first. I request a direct validation of Stage 1, e.g., recall of human-annotated query-relevant regions, comparison against alternative query-conditioned scores, or sign-flip ablations across th
  2. [§A.4, Supp. C.2, Tables S2, S5] The shared hyperparameters τ=0.2 and P=3 are selected from MME ablations on the same three backbones that later appear in the headline tables. Supp. C.2 shows P=3 is best on MME for all three architectures, but τ=0.2 is not best on LLaVA-NeXT-7B (τ=0.3 gives 1489.1 vs. 1481.1). This does not invalidate the approach, but it weakens the claim that the configuration is a single architecture-independent setting and introduces partial circularity for the MME rows in Tables 1–3. Please report sensitivity across all benchmarks rather than MME only, or fix the configuration on a held-out subset of benchmarks before presenting the headline numbers.
  3. [§4.2, Tables 1–4] No error bars, repeated runs, or significance tests are reported for any of the main results. Several aggregate differences are small—for example, Table 3 at 256 tokens shows 93.5 vs. 91.9 for DivPrune, and Table 4 at 512 tokens shows 94.1 vs. 93.4 for the next best method. Deterministic greedy selection and greedy decoding may reduce variance, but the paper does not state this. At minimum, the evaluation protocol should specify whether runs are deterministic or provide variance over multiple seeds so the reader can judge whether the headline gaps are meaningful.
  4. [§3.3, Eqs. (14)–(16)] The adaptive stopping rule is a heuristic whose validity is central to the ordered-design claim: it assumes that the P-th novelty event after the initial anchor marks sufficient coverage of query-critical evidence. The paper provides sensitivity to τ and P (Table S5), but no analysis of whether the resulting Krel actually tracks the amount or distribution of query-relevant evidence. Without this, the protected anchor size could be arbitrary relative to the true query-critical set, and the Stage-2 expansion budget would be misallocated. Please report the distribution of Krel for the evaluated benchmarks, or compare against an oracle anchor size on a subset of examples.
minor comments (3)
  1. [§4.3, Eq. (26)] The 'Efficiency Score' is a unitless product of prefill speedup and memory ratio. Its interpretation would be clearer if the two factors were also reported or plotted separately; as written, a method can achieve a high score either by being fast or by using less memory, which makes the single scalar hard to read.
  2. [Supp. C.3, Table S6] The sign-flip comparison is shown on only MME, TextVQA, and POPE. The text says the gain is 'not attributable to a uniform rescaling,' but the full benchmark table is needed to support the claim that the negated direction is consistently better for Stage 1, especially since the central mechanism depends on this choice.
  3. [§4.2, Tables 3–4; §A.5] The baseline comparison set is smaller for Qwen2.5-VL-7B and LLaVA-Video-7B (FastV, DivPrune, CDPruner). This is acknowledged in Supp. A.5, but the abstract and conclusion use 'consistently' across architectures; the cross-architecture claim would be stronger with at least one additional training-free baseline on these backbones.

Circularity Check

0 steps flagged

No significant circularity: AnchorPrune's central claim is an empirical comparison against external benchmarks, not a result forced by its own equations or by self-citation.

full rationale

The derivation chain is not circular. Stage 1 (Eqs. 6-11) ranks tokens by an architecture-specific heuristic score, and Stage 2 (Eqs. 18-23) greedily expands with importance-weighted novelty; the paper's claim that this protects query-critical evidence is then tested on external image/video benchmarks, with the headline LLaVA-NeXT number (97.6% at 160 tokens) computed by Eq. 25 from Table 2 rather than derived from the selection rule by construction. The negated-CLIP similarity in Eq. 7 is justified by an external citation [9], not by the authors' own prior work, and the paper explicitly disclaims calibrated semantic relevance ('without interpreting it as calibrated semantic relevance'), so any misranking risk is an assumption/correctness concern, not a self-definitional equivalence. The supplementary does show that tau, P, and the CLIP-similarity direction were selected using MME/POPE ablations on the same backbones and budgets that later appear in the main tables (Supp. Secs. C.2, C.3); this is a test-set selection / evaluation-leakage concern and a limitation of the evidence, but those values are not fitted constants disguised as predictions, and the central comparison against training-free baselines remains externally benchmarked. No equation in the paper reduces to a fitted parameter or to the target result, and there is no load-bearing self-citation chain. Hence no circularity under the stated rubric.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

All free parameters and axioms are method-level heuristics; the central contribution rests on the empirical validity of the Stage-1 relevance score and the heuristic adaptive budget. No new physical entities are introduced.

free parameters (4)
  • Novelty threshold tau = 0.2 (shared)
    Selected via MME sensitivity analysis (Sup. Table S5); tau=0.3 is slightly better on LLaVA-NeXT-7B but not adopted.
  • Patience parameter P = 3 (shared)
    Selected via MME sensitivity analysis (Sup. Table S5) from {1,3,5}; controls when anchor construction stops in Eq. 16.
  • Minimum anchor size Kmin schedule = 5-40 per unit (Sup. Table S2)
    Hand-set schedule per model and budget; defines the initial anchor in Eq. 13.
  • Maximum anchor size Kmax = floor(K/2)
    Design choice capping the anchor at half the budget in Eq. 12; no data-driven derivation.
axioms (5)
  • domain assumption Visual token pruning criteria such as relevance and diversity approximate the true conditional output distribution under a small budget.
    Sec. 2.1 defines the surrogate objective; no approximation bound is provided.
  • domain assumption Query-critical evidence is non-substitutable and concentrated in high-priority tokens, while contextual evidence is substitutable.
    Core premise of the ordered two-stage design (Sec. 1, Eqs. 5 and 21); if false, the anchor-first order gives no benefit.
  • domain assumption Negated CLIP patch-text similarity is a better localization signal than raw similarity for CLIP-aligned models.
    Imported from ECLIP [9] and validated only on MME/POPE in Sup. Table S6; Eq. 7 uses it without calibrated semantic interpretation.
  • ad hoc to paper The first P novelty events after the initial anchor mark sufficient coverage of query-critical evidence.
    Adaptive anchor budget (Eqs. 14-16) is a heuristic stopping rule; no evidence that it estimates the true size of query-critical evidence.
  • ad hoc to paper Global importance prior p_i (CLS attention or average attention mass) measures informativeness, and p_i * Delta(i;S) greedily approximates optimal contextual expansion.
    Stage 2 selection (Eq. 23) is a greedy heuristic; the multiplicative objective is validated only by ablation (Table 5), not derived.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of AnchorPrune: Relevance-Anchored Contextual Expansion for Visual Token Pruning." pith.science (2026). https://pith.science/paper/HOUWORDW

@misc{pith2026260707033,
  author       = {Pith},
  title        = {Pith review of: AnchorPrune: Relevance-Anchored Contextual Expansion for Visual Token Pruning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HOUWORDW}},
  note         = {Machine review of arXiv:2607.07033}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large vision-language models incur substantial inference costs because high-resolution inputs introduce thousands of visual tokens, many of which are redundant for a given query. Existing pruning methods often combine query relevance and token diversity, yet these objectives can conflict under aggressive compression: relevance-driven selection may overconcentrate the budget on correlated local evidence, while diversity-driven selection may suppress indispensable tokens or retain distinct but uninformative regions. We introduce AnchorPrune, a training-free framework that first constructs a protected relevance anchor and then expands it with complementary visual context. AnchorPrune adaptively determines the anchor size from the novelty profile of relevance-ranked tokens, preserving a compact set of query-critical evidence, and allocates the remaining budget through importance-weighted novelty to recover informative, non-redundant context relative to the anchor. This ordered design prevents contextual expansion from displacing indispensable query cues while improving overall visual coverage. AnchorPrune is lightweight, architecture-aware, and requires neither retraining nor model modification. Across image and video vision-language models and benchmarks, it consistently improves the accuracy-efficiency trade-off over training-free baselines, particularly under severe compression. On LLaVA-NeXT-7B, AnchorPrune preserves 97.6% of full-token performance using only 160 of 2,880 visual tokens. These results establish relevance-anchored contextual expansion as an effective principle for efficient multimodal inference. Code is available at https://github.com/MULTI-cau/AnchorPrune.

Figures

Figures reproduced from arXiv: 2607.07033 by Bumsoo Kim, Kyuan Oh.

Figure 1
Figure 1. Figure 1: Qualitative comparison on LLaVA-1.5-7B with 64 retained visual [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of AnchorPrune. Given an image and an instruction, Stage 1 ranks visual tokens using an architecture-specific query-conditioned priority signal and constructs a protected relevance anchor whose size is determined adaptively from the novelty profile of the ranked tokens. Stage 2 initializes the retained set with this anchor and allocates the remaining budget through importance-weighted novelty, sel… view at source ↗
Figure 3
Figure 3. Figure 3: Accuracy–efficiency trade-off on LLaVA-1.5-7B. [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

Review history (3 revisions) →

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. Beyond Accuracy: Auditing Spatial Provenance in Visual Token Pruning for OCR-Critical MLLM Inference

    cs.CV 2026-07 accept novelty 7.0

    For text-rich MLLMs, accuracy at a fixed token budget does not determine whether the surviving tokens retain spatial coverage of the OCR region that answers the query.

  2. Beyond Accuracy: Auditing Spatial Provenance in Visual Token Pruning for OCR-Critical MLLM Inference

    cs.CV 2026-07 conditional novelty 6.0

    Same-accuracy pruned vision-language prefixes can retain very different amounts of traceable OCR support (PosECR 0.620 vs 0.270 for Target vs Random at 30% budget), so pruning evaluation should report spatial provenan...

Reference graph

Works this paper leans on

30 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Alvar, S.R., Singh, G., Akbari, M., Zhang, Y.: Divprune: Diversity-based visual tokenpruningforlargemultimodalmodels.In:ProceedingsoftheComputerVision and Pattern Recognition Conference. pp. 9392–9401 (2025)

    2025

  2. [2]

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-vl technical report (2025)

    2025

  3. [3]

    In: European Conference on Computer Vision

    Chen, L., Zhao, H., Liu, T., Bai, S., Lin, J., Zhou, C., Chang, B.: An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In: European Conference on Computer Vision. pp. 19–35. Springer (2024)

    2024

  4. [4]

    In: Advances in Neural Information Processing Systems

    Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., Wu, Y., Ji, R., Shan, C., He, R.: Mme: A comprehensive evaluation bench- mark for multimodal large language models. In: Advances in Neural Information Processing Systems. vol. 38 (2025)

    2025

  5. [5]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Fu, C., Dai, Y., Luo, Y., Li, L., Ren, S., Zhang, R., Wang, Z., Zhou, C., Shen, Y., Zhang, M., et al.: Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 24108–24118 (2025)

    2025

  6. [6]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6904–6913 (2017)

    2017

  7. [7]

    In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition

    Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition. pp. 6700–6709 (2019)

    2019

  8. [8]

    In: European conference on computer vision

    Kembhavi, A., Salvato, M., Kolve, E., Seo, M., Hajishirzi, H., Farhadi, A.: A di- agram is worth a dozen images. In: European conference on computer vision. pp. 235–251. Springer (2016)

    2016

  9. [9]

    arXiv preprint arXiv:2209.07046 (2022)

    Li, Y., Wang, H., Duan, Y., Xu, H., Li, X.: Exploring visual interpretability for contrastive language-image pre-training. arXiv preprint arXiv:2209.07046 (2022)

    Pith/arXiv arXiv 2022

  10. [10]

    In: Proceedings of the 2023 conference on empirical methods in natural language processing

    Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, X., Wen, J.R.: Evaluating object hal- lucination in large vision-language models. In: Proceedings of the 2023 conference on empirical methods in natural language processing. pp. 292–305 (2023)

    2023

  11. [11]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tun- ing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 26296–26306 (2024)

    2024

  12. [12]

    io/blog/2024-01-30-llava-next/

    Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-next: Improved reasoning, ocr, and world knowledge (January 2024),https://llava-vl.github. io/blog/2024-01-30-llava-next/

    2024

  13. [13]

    Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? In: European conference on computer vision. pp. 216–233. Springer (2024)

    2024

  14. [14]

    Liu, Y., Li, S., Liu, Y., Wang, Y., Ren, S., Li, L., Chen, S., Sun, X., Hou, L.: Tem- pcompass: Do video llms really understand videos? In: Findings of the Association for Computational Linguistics: ACL 2024. pp. 8731–8772 (2024)

    2024

  15. [15]

    Science China Information Sciences67(12), 220102 (2024) AnchorPrune 17

    Liu, Y., Li, Z., Huang, M., Yang, B., Yu, W., Li, C., Yin, X.C., Liu, C.L., Jin, L., Bai, X.: Ocrbench: on the hidden mystery of ocr in large multimodal models. Science China Information Sciences67(12), 220102 (2024) AnchorPrune 17

    2024

  16. [16]

    Advances in neural information processing systems35, 2507– 2521 (2022)

    Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in neural information processing systems35, 2507– 2521 (2022)

    2022

  17. [17]

    Advances in Neural Information Processing Systems36, 46212–46244 (2023)

    Mangalam, K., Akshulakov, R., Malik, J.: Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems36, 46212–46244 (2023)

    2023

  18. [18]

    In: Findings of the association for computational linguistics: ACL 2022

    Masry, A., Do, X.L., Tan, J.Q., Joty, S., Hoque, E.: Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In: Findings of the association for computational linguistics: ACL 2022. pp. 2263–2279 (2022)

    2022

  19. [19]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Mathew, M., Karatzas, D., Jawahar, C.: Docvqa: A dataset for vqa on document images. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 2200–2209 (2021)

    2021

  20. [20]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

    2021

  21. [21]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Shang, Y., Cai, M., Xu, B., Lee, Y.J., Yan, Y.: Llava-prumerge: Adaptive token reduction for efficient large multimodal models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22857–22867 (2025)

    2025

  22. [22]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8317–8326 (2019)

    2019

  23. [23]

    arXiv preprint arXiv:2410.17247 (2024)

    Xing, L., Huang, Q., Dong, X., Lu, J., Zhang, P., Zang, Y., Cao, Y., He, C., Wang, J., Wu, F., et al.: Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247 (2024)

    Pith/arXiv arXiv 2024

  24. [24]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yang, S., Chen, Y., Tian, Z., Wang, C., Li, J., Yu, B., Jia, J.: Visionzip: Longer is better but not necessary in vision language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19792– 19802 (2025)

    2025

  25. [25]

    arXiv preprint arXiv:2308.02490 (2023)

    Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm- vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)

    Pith/arXiv arXiv 2023

  26. [26]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Yue, X., Ni, Y., Zhang, K., Zheng, T., Liu, R., Zhang, G., Stevens, S., Jiang, D., Ren, W., Sun, Y., et al.: Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9556–9567 (2024)

    2024

  27. [27]

    In: Findings of the Association for Computational Linguistics: NAACL 2025

    Zhang, K., Li, B., Zhang, P., Pu, F., Cahyono, J.A., Hu, K., Liu, S., Zhang, Y., Yang, J., Li, C., et al.: Lmms-eval: Reality check on the evaluation of large mul- timodal models. In: Findings of the Association for Computational Linguistics: NAACL 2025. pp. 881–916 (2025)

    2025

  28. [28]

    Advances in Neural Information Processing Systems38, 25438–25468 (2025)

    Zhang, Q., Liu, M., Li, L., Lu, M., Zhang, Y., Pan, J., She, Q., Zhang, S.: Be- yond attention or similarity: Maximizing conditional diversity for token pruning in mllms. Advances in Neural Information Processing Systems38, 25438–25468 (2025)

    2025

  29. [29]

    arXiv preprint arXiv:2410.04417 (2024) 18 K

    Zhang, Y., Fan, C.K., Ma, J., Zheng, W., Huang, T., Cheng, K., Gudovskiy, D., Okuno, T., Nakata, Y., Keutzer, K., et al.: Sparsevlm: Visual token sparsifica- tion for efficient vision-language model inference. arXiv preprint arXiv:2410.04417 (2024) 18 K. Oh and B. Kim

    Pith/arXiv arXiv 2024

  30. [30]

    Zhang, Y., Wu, J., Li, W., Li, B., Ma, Z., Liu, Z., Li, C.: Llava-video: Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713 (2024) Supplementary Material AnchorPrune: Relevance-Anchored Contextual Expansion for Visual Token Pruning Kyuan Oh and Bumsoo Kim† Chung-Ang University, Seoul, Korea {oka04108,bumsoo}@cau.ac.kr A Implemen...

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.