Pith. sign in

REVIEW 2 major objections 4 minor 73 references

Visual-token pruning can preserve answer accuracy while silently discarding the spatial evidence that supports the answer, so pruning audits must track geometric provenance, not just accuracy and compression.

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 ↗

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 provenance and cost.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The Qwen evidence is strong and the audit protocol is a real contribution; the cross-model framing outruns the construct-validity data, but the paper honestly scopes most of it. the 2 major comments →

arxiv 2608.00077 v2 pith:CFEG5JUG submitted 2026-07-29 cs.CV

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

Feixiang Liu , Qiang Qiu , Hao Zhang , Xinyue Wang This is my paper
classification cs.CV
keywords visual token pruningmultimodal large language modelsspatial provenanceOCR-critical inferenceevidence-risk audittoken traceabilityprefix compressionevaluation metrics
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

The paper argues that judging visual-token pruning by answer accuracy at a fixed token budget can conceal a distinct failure: the model still answers correctly even when no surviving token is traceable to the small OCR region that actually supports the answer. To make this failure measurable, the authors introduce spatial provenance — the geometric coverage of an annotated source region by the grid cells of retained tokens — and audit it jointly with answer behavior, interventional tests, and measured cost on three multimodal LLMs. On a locked, image-disjoint confirmation set, a target-conditioned selector that keeps 30% of visual tokens matches full-prefix accuracy (0.786 vs. 0.783) while positive-source coverage is 0.620, versus 0.270 and 0.318 for random and grid pruning at the same budget. The paper's conclusion is a reporting requirement: pruning evaluations should report surviving spatial support and realized efficiency, because the quality–risk–traceability frontier is model- and task-specific and cannot be transferred by keep ratio alone.

Core claim

The central claim is that answer quality and compression ratio do not determine what a pruned prefix still represents spatially. Using each token's grid cell as its provenance cell, the paper defines PosECR as the lineage-based geometric coverage of annotated answer-supporting regions, averaged over positive probes, with NegSRC tracking confusable near-miss source regions. At the same 30% budget on TextOCR-Hard, Qwen3-VL keeps 0.786 accuracy versus 0.783 for the full prefix, yet retains 0.620 of the supporting region's coverage while random and grid retention keep only 0.270 and 0.318; native VisionZip reaches 1.000 lineage coverage but only 0.846 anchor coverage, separating 'which tokens pa

What carries the argument

The load-bearing object is PosECR (positive evidence coverage ratio): for each compressed prefix, the area of the annotated answer-supporting region that is covered by the grid cells of the retained tokens' source lineage, averaged over positive probes (Eq. 2). Each token is assigned a provenance cell recording where it came from in the image grid — origin, not receptive field — and every compressed output carries a source-lineage set; a merged token can participate with source lineage while its representative anchor stays elsewhere, which the companion AnchorECR metric separates. Around this metric the paper builds a budgeted selection contract, S = P ∪ Top(s, K−|P|; rest), in which transpa

Load-bearing premise

The metric PosECR assumes a retained token's grid cell is the right unit for tracking whether a supporting region is still represented — but visual tokens are contextualized, so tokens outside the annotated box can still carry its content; zero PosECR proves that no locally traceable cell survives, not that the evidence is gone.

What would settle it

On hundreds of OCR-critical probes with a fourth MLLM, build paired prefixes at one budget whose retained cells either fully cover or fully avoid the annotated source box, and compare yes/no accuracy plus the margin change when covering cells are deleted. If the zero-coverage and full-coverage prefixes are behaviorally indistinguishable, and deleting covering cells never moves answers, then geometric provenance is not carrying the claimed hidden failure; the paper's own tables show LLaVA and InternVL construct-validity intervals already include zero, so this experiment decides whether the Qwen

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

If this is right

  • At matched budget and near-matched accuracy, prefixes can differ sharply in surviving support: on the same locked probes, positive-source coverage was 0.620 for target selection versus 0.270 random and 0.318 grid at 30% retention, so accuracy-based leaderboards conflate distinct prefixes.
  • Coverage and answer risk move on separate axes: raising PosECR can raise the hard-negative false-positive rate (InternVL soft evidence: 0.902 PosECR at 0.530 hFPR vs. grid's 0.474), so selectors must be chosen on a quality–risk–traceability frontier, not a single score.
  • Merging-based pruning must report lineage and anchor coverage separately; VisionZip's exhaustive lineage PosECR of 1.000 falls to 0.846 anchor coverage for Qwen and 0.563 for LLaVA, so 'represented somewhere in a merge' is not 'locally traceable.'
  • Operating points do not transfer across tasks: binary target verification is safe at 30% retention, but full-validation TextVQA/DocVQA generation needs roughly 70%, where Target+Grid trails Full by 0.033–0.093 and trails random on several model–task pairs.
  • Keep ratio understates real cost in both directions: a materialized 20%-retention Qwen prefix gave 4.32× batch-prefill speedup and 76.4% lower incremental peak memory, while online OCR-detector latency erased single-sample gains.

Where Pith is reading between the lines

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

  • The audit template extends beyond OCR: any task where a small annotated region supports an answer — charts, tables, faces, medical lesions — could use the same provenance-coverage readout, and the paper's geometry is annotation-agnostic.
  • Position-ID policy is a hidden confound in any pruning comparison: identical masks flipped 279/1000 LLaVA and 101/536 InternVL decisions when logical positions were preserved rather than compacted, implying that published pruning numbers are not portable unless position semantics are disclosed.
  • Zero PosECR is a lower bound on information loss, not a proof of it; a natural tightening is an information-recovery experiment that measures how much of the source region's content can be decoded from retained contextual tokens, moving from geometric traceability toward causal evidence.
  • The Qwen-only convergent validity suggests a deployment heuristic worth testing: for text-critical tasks, commission pruned prefixes by auditing surviving support on a small annotated probe set rather than tuning on accuracy, since accuracy is least sensitive exactly where provenance collapses.
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

2 major / 4 minor

Summary. The paper argues that visual-token pruning for OCR-critical MLLM inference cannot be evaluated by answer accuracy and compression ratio alone. It introduces a spatial-provenance audit centered on PosECR/NegSRC, geometric measures of whether retained (or merged) visual-token cells overlap annotation-defined source regions. Using TextOCR-Hard development and locked image-disjoint confirmation sets, it compares budgeted training-free selectors (Target, Random, Grid, Protected, Soft evidence) and external pruning methods across Qwen3-VL-8B, LLaVA-1.5-7B, and InternVL3.5-8B. The headline empirical result is on Qwen: at 30% retention, Target matches Full accuracy (paired difference +0.003, 95% CI [-0.014, +0.020]) while retaining 0.620 PosECR versus 0.270 for Random and 0.318 for Grid. The paper also reports interventions, wrong-image controls, human QC, position-policy checks, measured CUDA speedups/memory, and transfer results on OCRBench, TextVQA, and DocVQA. The central claim is that prefixes with similar budgets and accuracy can retain fundamentally different spatial support, and that pruning reports should therefore include provenance and realized cost.

Significance. If the audit methodology is accepted, this is a useful contribution to MLLM evaluation: it formalizes a failure mode that accuracy-comparison protocols miss, and it supplies a replicable contract for measuring it. The paper's strengths are substantial: a prespecified primary estimand on an image-disjoint locked confirmation set, paired image-cluster bootstraps, matched random/grid/shuffled-score controls, wrong-image and blank-image controls, deletion/restoration and occlusion interventions, human QC of the hard-negative construction, position-policy sensitivity analyses, and honest limitation statements. The efficiency measurements are grounded in actual shortened-prefix execution rather than FLOP counts. The main risk is not the internal statistics but the interpretation of PosECR as a behaviorally meaningful 'failure' signal on all three backbones, since the construct-validity evidence is strong only for Qwen. As a geometric provenance metric, the result is sound; as a cross-model 'hidden failure' claim, it currently overreaches.

major comments (2)
  1. [Abstract/Conclusion; Supplement 'Statistical and Construction Audits' (Table S9)] The abstract and conclusion claim a multi-backbone 'distinct failure' ('Across Qwen3-VL-8B, LLaVA-1.5-7B, and InternVL3.5-8B...'), but PosECR's construct validity is demonstrated only for Qwen. In Table S9, partial Spearman intervals are Qwen Delete [0.203,0.371]; LLaVA Occlusion CI is [-0.025,0.535]; InternVL Occlusion/Delete CIs are [-0.119,0.414]/[-0.075,0.096], all including zero. The Supplement itself states these results 'support convergent validity on Qwen and a geometry audit on all models, not a backbone-invariant causal metric.' The geometric separations in Table 1 are real for all backbones, but the 'hidden failure' framing implies a behavioral consequence that is not established for LLaVA and InternVL. Please either add per-backbone behavioral validation or limit the 'failure' claim to Qwen and describe the other rows as geometric provenance audits.
  2. [Problem Setup and Abstract] The paper defines PosECR over token-origin cells and explicitly warns that 'tokens outside R may still encode its information'; the Supplement reiterates that zero ECR does not prove information erasure. Given that, the Abstract's phrase 'an answer remains correct even when no retained token is locally traceable' should not be presented as a 'failure' without stating that traceability is a proxy for evidence, not evidence itself. The intervention evidence in Table S15 is also model-dependent: Qwen shows the expected positive/negative asymmetry, whereas LLaVA's evidence-kept vs removed logit-drop is +0.005 overall and InternVL's signs are mixed. The manuscript does acknowledge this in Limitations, but the abstract and conclusion carry the stronger interpretation. I recommend carrying the caveat into the headline claims, e.g., 'spatial provenance loss' rather than 'evidence failure'.
minor comments (4)
  1. [Table S16] The FastV adaptation predicts 'no' for every probe at every reported budget (positive accuracy 0.000, negative accuracy 1.000). Since FastV is not in the main-table comparisons, this is not load-bearing, but a footnote should state that these rows are a degenerate adaptation rather than a meaningful FastV operating point.
  2. [Abstract] The sentence 'same-budget Target, Random, and Grid retain sharply different positive-support coverage: 0.620, 0.270, and 0.318' omits the model (Qwen) and the retention budget (30%). Adding 'Qwen Target at 30% retention' would prevent misreading as a cross-model uniform result.
  3. [Table 1 and Table S5] For non-merging methods, AncECR equals lineage PosECR by construction. This is stated in the text, but the main table still lists both columns for those rows. A dash or a footnote would reduce redundancy and make the VisionZip distinction (lineage vs anchor) more visually salient.
  4. [Position-ID Policy Sensitivity] The supplement reports that preserving pre-pruning logical IDs flips 279/1000 LLaVA and 101/536 InternVL decisions. This is an important confound, but it is only discussed in the supplement. A one-sentence mention in the main text's Deployment Policy section would make the reporting contract more complete.

Circularity Check

0 steps flagged

No significant circularity: the paper is an empirical audit with locked confirmation, transparent selectors, and measured metrics.

full rationale

The paper does not derive its central claim from its own construction. PosECR (Eqs. 2-3) is defined as grid-cell lineage overlap and is measured independently of the Target selector's score (Eqs. 5-6), which uses only query-text similarity and token norm, not annotations. The main contrast (Qwen Target at 30%: accuracy 0.786 vs Full 0.783, PosECR 0.620 vs Random 0.270 / Grid 0.318) is an observed outcome on a locked image-disjoint set after selectors and budgets were frozen, with paired bootstrap intervals; it is not a fitted value or a renamed input. VisionZip's lineage PosECR of 1.000 is explicitly explained as exhaustive merge coverage and separated from AnchorECR, so no result is relabeled. The paper repeatedly disclaims that ECR is causal or backbone-invariant ('These results support convergent validity on Qwen and a geometry audit on all models, not a backbone-invariant causal metric'), and the construct-validity gaps for LLaVA/InternVL are validity limitations, not circular reductions. No load-bearing self-citation or imported uniqueness theorem appears. The selector constants were tuned on development, but the audited accuracy/provenance outcomes were not used to re-fit anything, so no fitted-input-called-prediction pattern applies.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim depends on metric validity and dataset construction more than on physical axioms. The paper is transparent about the proxy nature of PosECR and about backbone-specific construct validity. Free parameters are selector hyperparameters tuned on development; they do not by themselves create the Random/Grid low-coverage result.

free parameters (5)
  • relevance-norm mixture weights (0.85/0.15) = 0.85 relevance, 0.15 norm
    Eq. (6); one-factor sweep on development (Table S2). Affects which tokens are kept, hence all PosECR and accuracy points.
  • query top-k aggregation m=2 = 2
    Eq. (5); chosen over m=1,4 on development; affects target-conditioned scores.
  • soft-evidence boost beta=0.05 = 0.05
    Eq. (7); InternVL sweep beta in [0,0.2]; changes Soft evidence coverage and hFPR.
  • decision thresholds t = t=0 for Qwen/LLaVA; t=2.043 for InternVL common threshold
    InternVL threshold selected from Full-derived development scores; thresholds affect accuracy and hFPR rows.
  • retention budgets and grid reservation = 0.30/0.40/0.50; grid reserve 0.5K
    Budgets and grid reservation selected on development; main confirmation rows are at these fixed budgets.
axioms (5)
  • domain assumption Grid-induced provenance cell Bi records token origin; geometric overlap of retained cells with annotated boxes measures surviving spatial support.
    Eq. (2); paper acknowledges tokens outside R may still encode region info (citing Fan et al. 2026), so PosECR is a proxy, not a causal guarantee.
  • domain assumption TextOCR annotations are accurate enough to define positive support and confusable negative source regions.
    Dataset construction; human QC found 465/500 strict-valid confirmation pairs and kappa 0.445, so annotation reliability is partial.
  • domain assumption Likelihood-threshold scoring of yes/no continuations yields a valid accuracy and hFPR measure for these probes.
    Evaluation contract in experiments; no external validation of prompt template sensitivity.
  • domain assumption The locked image-disjoint confirmation split prevents operating-point adaptation.
    Split generated with a seed before inference and selectors frozen; however hyperparameters were still tuned on development, so this is a split lock, not full preregistration.
  • domain assumption Pruning materialization preserves semantics of retained tokens, with position handling controlled as a separate variable.
    Materialization section; position-ID policy sensitivity is tested because logical position IDs can flip decisions (279/1000 LLaVA, 101/536 InternVL).

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Accuracy: Auditing Spatial Provenance in Visual Token Pruning for OCR-Critical MLLM Inference." pith.science (2026). https://pith.science/paper/CFEG5JUG

@misc{pith2026260800077,
  author       = {Pith},
  title        = {Pith review of: Beyond Accuracy: Auditing Spatial Provenance in Visual Token Pruning for OCR-Critical MLLM Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CFEG5JUG}},
  note         = {Machine review of arXiv:2608.00077}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Visual-token pruning is usually judged by answer quality at a fixed retention budget. For text-rich multimodal large language models (MLLMs), this protocol can miss a distinct failure: an answer remains correct even when no retained token is locally traceable to the small OCR region that supports it. We turn this blind spot into an evidence-risk audit that couples answer behavior with geometric token-origin provenance, interventions, and realized cost; transparent training-free selectors isolate controlled operating points. On locked image-disjoint confirmation, Qwen Target at 30% retention has observed accuracy 0.786 versus 0.783 for Full (paired image-cluster difference +0.003, 95% CI [-0.014, +0.020]), yet same-budget Target, Random, and Grid retain sharply different positive-support coverage: 0.620, 0.270, and 0.318. Across Qwen3-VL-8B, LLaVA-1.5-7B, and InternVL3.5-8B, matched controls, interventions, detector tests, and external methods reveal model-specific quality-risk-traceability frontiers that accuracy alone does not expose. Materialized prefixes yield up to 4.32x batch-prefill speedup and 76.4% lower incremental peak memory; full-validation TextVQA and DocVQA further show that favorable target-verification points do not imply task-general compression. Visual-token pruning should therefore report surviving spatial provenance and realized cost alongside quality and compression.

Figures

Figures reproduced from arXiv: 2608.00077 by Feixiang Liu, Hao Zhang, Qiang Qiu, Xinyue Wang.

Figure 1
Figure 1. Figure 1: Evidence-risk audit overview. Box-free Target selects a fixed-budget prefix using target-conditioned relevance and [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

Review history (2 revisions) →

discussion (0)

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

Reference graph

Works this paper leans on

73 extracted references · 37 canonical work pages · 4 internal anchors

  1. [1]

    2023 , eprint=

    Visual Instruction Tuning , author=. 2023 , eprint=

    2023

  2. [2]

    2023 , eprint=

    Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond , author=. 2023 , eprint=

    2023

  3. [3]

    2024 , eprint=

    InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks , author=. 2024 , eprint=

    2024

  4. [4]

    2019 , eprint=

    Towards VQA Models That Can Read , author=. 2019 , eprint=

    2019

  5. [5]

    2021 , eprint=

    TextOCR: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text , author=. 2021 , eprint=

    work page 2021

  6. [6]

    2024 , eprint=

    OCRBench: On the Hidden Mystery of OCR in Large Multimodal Models , author=. 2024 , eprint=. doi:10.1007/s11432-024-4235-6 , url=

    doi:10.1007/s11432-024-4235-6 2024

  7. [7]

    2022 , eprint=

    FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness , author=. 2022 , eprint=

    2022

  8. [8]

    2023 , eprint=

    FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning , author=. 2023 , eprint=

    2023

  9. [9]

    2024 , eprint=

    An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models , author=. 2024 , eprint=

    2024

  10. [10]

    2026 , eprint=

    LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Models , author=. 2026 , eprint=

    work page 2026

  11. [11]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Boosting Multimodal Large Language Models with Visual Tokens Withdrawal for Rapid Inference , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2025 , doi=

    work page 2025

  12. [12]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    What Kind of Visual Tokens Do We Need? Training-Free Visual Token Pruning for Multi-Modal Large Language Models from the Perspective of Graph , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2025 , doi=

    work page 2025

  13. [13]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Fit and Prune: Fast and Training-free Visual Token Pruning for Multi-modal Large Language Models , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2025 , doi=

    work page 2025

  14. [14]

    2024 , eprint=

    TokenPacker: Efficient Visual Projector for Multimodal LLM , author=. 2024 , eprint=

    work page 2024

  15. [15]

    2025 , eprint=

    SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference , author=. 2025 , eprint=

    2025

  16. [16]

    2025 , eprint=

    PyramidDrop: Accelerating Your Large Vision-Language Models via Pyramid Visual Redundancy Reduction , author=. 2025 , eprint=

    2025

  17. [17]

    2026 , eprint=

    VisionZip: Longer is Better but Not Necessary in Vision Language Models , author=. 2026 , eprint=

    work page 2026

  18. [18]

    2025 , eprint=

    TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model , author=. 2025 , eprint=

    work page 2025

  19. [19]

    2024 , eprint=

    ATP-LLaVA: Adaptive Token Pruning for Large Vision Language Models , author=. 2024 , eprint=

    work page 2024

  20. [20]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Token Pruning in Multimodal Large Language Models: Are We Solving the Right Problem? , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=. 2025 , address=. doi:10.18653/v1/2025.findings-acl.802 , url=

    doi:10.18653/v1/2025.findings-acl.802 2025

  21. [21]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Stop Looking for ``Important Tokens'' in Multimodal Language Models: Duplication Matters More , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=. 2025 , address=. doi:10.18653/v1/2025.emnlp-main.505 , url=

    doi:10.18653/v1/2025.emnlp-main.505 2025

  22. [22]

    2025 , eprint=

    Feather the Throttle: Revisiting Visual Token Pruning for Vision-Language Model Acceleration , author=. 2025 , eprint=

    work page 2025

  23. [23]

    2025 , eprint=

    VFlowOpt: A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization , author=. 2025 , eprint=

    work page 2025

  24. [24]

    2025 , eprint=

    CROP: Contextual Region-Oriented Visual Token Pruning , author=. 2025 , eprint=

    work page 2025

  25. [25]

    Huang, Yihong and Ma, Fei and Shao, Yihua and Guo, Jingcai and Yu, Zitong and Cui, Laizhong and Tian, Qi , year=. N. 2602.02951 , archivePrefix=

    arXiv

  26. [26]

    2026 , eprint=

    How Many Visual Tokens Do Multimodal Language Models Need? Scaling Visual Token Pruning with F ^3 A , author=. 2026 , eprint=

    work page 2026

  27. [27]

    2021 , eprint=

    Learning Transferable Visual Models From Natural Language Supervision , author=. 2021 , eprint=

    2021

  28. [28]

    2022 , eprint=

    Flamingo: a Visual Language Model for Few-Shot Learning , author=. 2022 , eprint=

    2022

  29. [29]

    2023 , eprint=

    BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models , author=. 2023 , eprint=

    2023

  30. [30]

    2023 , eprint=

    InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning , author=. 2023 , eprint=

    2023

  31. [31]

    2023 , eprint=

    MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models , author=. 2023 , eprint=

    2023

  32. [32]

    2023 , eprint=

    Kosmos-2: Grounding Multimodal Large Language Models to the World , author=. 2023 , eprint=

    work page 2023

  33. [33]

    2021 , eprint=

    DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification , author=. 2021 , eprint=

    2021

  34. [34]

    2021 , eprint=

    Evo-ViT: Slow-Fast Token Evolution for Dynamic Vision Transformer , author=. 2021 , eprint=

    work page 2021

  35. [35]

    2021 , eprint=

    AdaViT: Adaptive Tokens for Efficient Vision Transformer , author=. 2021 , eprint=

    work page 2021

  36. [36]

    2022 , eprint=

    Token Merging: Your ViT But Faster , author=. 2022 , eprint=

    work page 2022

  37. [37]

    2019 , eprint=

    Scene Text Visual Question Answering , author=. 2019 , eprint=

    2019

  38. [38]

    2020 , eprint=

    DocVQA: A Dataset for VQA on Document Images , author=. 2020 , eprint=

    work page 2020

  39. [39]

    2020 , eprint=

    TextCaps: a Dataset for Image Captioning with Reading Comprehension , author=. 2020 , eprint=

    2020

  40. [40]

    2021 , eprint=

    InfographicVQA , author=. 2021 , eprint=

    2021

  41. [41]

    2022 , eprint=

    ChartQA: A Benchmark for Question Answering about Charts with Visual and Logical Reasoning , author=. 2022 , eprint=

    2022

  42. [42]

    2023 , eprint=

    Evaluating Object Hallucination in Large Vision-Language Models , author=. 2023 , eprint=

    2023

  43. [43]

    2023 , eprint=

    MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models , author=. 2023 , eprint=

    2023

  44. [44]

    2023 , eprint=

    MMBench: Is Your Multi-modal Model an All-around Player? , author=. 2023 , eprint=

    2023

  45. [45]

    2023 , eprint=

    SEED-Bench: Benchmarking Multimodal LLMs with Generative Comprehension , author=. 2023 , eprint=

    2023

  46. [46]

    2023 , eprint=

    What's ``up'' with vision-language models? Investigating their struggle with spatial reasoning , author=. 2023 , eprint=

    work page 2023

  47. [47]

    2024 , eprint=

    GSR-BENCH: A Benchmark for Grounded Spatial Reasoning Evaluation via Multimodal LLMs , author=. 2024 , eprint=

    2024

  48. [48]

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

    Mincheol Kwon and Minseung Lee and Seonga Choi and Miso Choi and Kyeongjin Oh and Hyunyoung Lee and Cheonyoung Park and Yongho Song and Seunghyun Park and Jinkyu Kim , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    work page

  49. [49]

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

    Qihui Zhu and Tao Zhang and Yuchen Wang and Shuangwu Chen and Xiaobin Tan and Jian Yang and Yang Liu and Yinfei Pan , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    work page

  50. [50]

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

    Yahong Wang and Juncheng Wu and Zhangkai Ni and Longzhen Yang and Yihang Liu and Chengmei Yang and Ying Wen and Lianghua He and Xianfeng Tang and Hui Liu and Yuyin Zhou , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    work page

  51. [51]

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

    Zhenkai Wu and Xiaowen Ma and Zhenliang Ni and Dengming Zhang and Han Shu and Xin Jiang and Xinghao Chen , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    work page

  52. [52]

    arXiv preprint arXiv:2511.21631 , year =

    Shuai Bai and Yuxuan Cai and Ruizhe Chen and Keqin Chen and Xionghui Chen and Zesen Cheng and others , title =. arXiv preprint arXiv:2511.21631 , year =

    Pith/arXiv arXiv

  53. [53]

    arXiv preprint arXiv:2508.18265 , year =

    Weiyun Wang and Zhangwei Gao and Lixin Gu and Hengjun Pu and Long Cui and Xingguang Wei and others , title =. arXiv preprint arXiv:2508.18265 , year =

    Pith/arXiv arXiv

  54. [54]

    Advances in Neural Information Processing Systems , volume =

    Jinhong Deng and Wen Li and Joey Tianyi Zhou and Yang He , title =. Advances in Neural Information Processing Systems , volume =. 2025 , url =

    work page 2025

  55. [55]

    Advances in Neural Information Processing Systems , volume =

    Qizhe Zhang and Mengzhen Liu and Lichen Li and Ming Lu and Yuan Zhang and Junwen Pan and Qi She and Shanghang Zhang , title =. Advances in Neural Information Processing Systems , volume =. 2025 , url =

    work page 2025

  56. [56]

    Advances in Neural Information Processing Systems , volume =

    Yangfu Li and Hongjian Zhan and Tianyi Chen and Qi Liu and Yu-Jie Xiong and Yue Lu , title =. Advances in Neural Information Processing Systems , volume =. 2025 , url =

    work page 2025

  57. [57]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Ziyang Wang and Mengwei Li and Hao Yin and Wenhao Liu and Zilei Wang , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2026 , doi =

    work page 2026

  58. [58]

    arXiv preprint arXiv:2602.13315 , year =

    Yifan Tan and Yifu Sun and Shirui Huang and Hong Liu and Guanghua Yu and Jianchen Zhu and Yangdong Deng , title =. arXiv preprint arXiv:2602.13315 , year =

    arXiv

  59. [59]

    TOPS: First-Principles Visual Token Pruning via Constructing Token Optimal Preservation Sets for Efficient MLLM Inference

    Tinghao Wang and Yichen Guo and Rui Huang and Zheng Lu and Qizhe Zhang and Chenxi Li and Yuan Zhang and Jiajun Cao and Zhirong Shen and Yaosong Du and Guangyan Gan and Wenya Wang and Lin William Cong and Shanghang Zhang , title =. arXiv preprint arXiv:2606.27161 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  60. [60]

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

    Chenxi Du and Yongheng Deng and Jiani Liu and Yujia Zhang and Xi Chen and Ju Ren , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2026 , url =

    work page 2026

  61. [61]

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

    Tong Xu and Hailong Shi and Xingyu Gao , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2026 , url =

    work page 2026

  62. [62]

    Kim , title =

    Joonmyung Choi and Sanghyeok Lee and Jongha Kim and Sehyung Kim and Dohwan Ko and Jihyung Kil and Hyunwoo J. Kim , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2026 , url =

    work page 2026

  63. [63]

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

    Tianxiao Gao and Shanwei Zhao and Shuo Fang and Shiai Zhu and Chenguang Ma , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2026 , url =

    work page 2026

  64. [64]

    arXiv preprint arXiv:2511.02650 , year =

    Tianfan Peng and Yuntao Du and Pengzhou Ji and Shijie Dong and Kailin Jiang and Mingchuan Ma and Yijun Tian and Jinhe Bi and Qian Li and Wei Du and Feng Xiao and Lizhen Cui , title =. arXiv preprint arXiv:2511.02650 , year =

    arXiv

  65. [65]

    arXiv preprint arXiv:2503.10501 , year =

    Xudong Tan and Peng Ye and Chongjun Tu and Jianjian Cao and Yaoxin Yang and Lin Zhang and Dongzhan Zhou and Tao Chen , title =. arXiv preprint arXiv:2503.10501 , year =

    Pith/arXiv arXiv

  66. [66]

    arXiv preprint arXiv:2605.00392 , year =

    Ben Wan and Yan Feng and Zihan Tang and Weizhe Huang and Yuting Zeng and Jia Wang and Tongxuan Liu , title =. arXiv preprint arXiv:2605.00392 , year =

    Pith/arXiv arXiv

  67. [67]

    When Does Visual Token Pruning Improve Calibration? The Role of Evidence Coverage in MLLMs

    Tan, Kaizhen and Feng, Yang and Du, Heqing and Hong, Hanzhe and Tao, Siru , title =. arXiv preprint arXiv:2604.12035 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  68. [68]

    arXiv preprint arXiv:2604.12358 , year =

    Jiwan Kim and Kibum Kim and Wonjoong Kim and Byung-Kwan Lee and Chanyoung Park , title =. arXiv preprint arXiv:2604.12358 , year =

    Pith/arXiv arXiv

  69. [69]

    arXiv preprint arXiv:2506.21873 , year =

    Tzu-Chun Chien and Chieh-Kai Lin and Shiang-Feng Tsai and Ruei-Chi Lai and Hung-Jen Chen and Min Sun , title =. arXiv preprint arXiv:2506.21873 , year =

    Pith/arXiv arXiv

  70. [70]

    Reroute, Don't Remove: Recoverable Visual Token Routing for Vision-Language Models

    Cheng-Yu Yang and Shao-Yuan Lo and Yu-Lun Liu , title =. arXiv preprint arXiv:2606.12412 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  71. [71]

    AnchorPrune: Relevance-Anchored Contextual Expansion for Visual Token Pruning

    Kyuan Oh and Bumsoo Kim , title =. arXiv preprint arXiv:2607.07033 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  72. [72]

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

    Yingqi Fan and Junlong Tong and Anhao Zhao and Xiaoyu Shen , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2026 , url =

    work page 2026

  73. [73]

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

    Guohao Sun and Yufei Wang and Sizhuo Ma and Yuege Xie and Yuting Cheng and Zhiqiang Tao and Jian Wang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2026 , url =

    work page 2026

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.