Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

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

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read When visual token pruning is guided by coverage rather than saliency, it can improve calibration while preserving accuracy, cutting ECE from 0.041 to 0.016 on POPE.

desk verdict Plausible POPE result with an honest body, but the metadata abstract overclaims and the restricted-verbalizer confidence measure is an unaddressed confounder. read the letter →

arxiv 2604.12035 v3 pith:YLPXIENM submitted 2026-04-13 cs.CV

Kaizhen Tan , Yang Feng , Heqing Du , Hanzhe Hong , Siru Tao , Xin Xu This is my paper
classification cs.CV
keywords visualtokenpruningcalibrationexpectederrormultimodallargelanguagemodelscoverage-basedselectionsaliency-basedevidencecoverageselectiveprediction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that how visual tokens are pruned in multimodal LLMs changes confidence quality, not just accuracy. On POPE with LLaVA-1.5-7B, pruning 576 visual tokens down to 128 using coverage-based selection lowers expected calibration error from 0.041 to 0.016 with accuracy essentially unchanged. The key claim is that the selection rule—coverage versus saliency—matters more than the compression budget. Saliency-based pruning becomes less calibrated than random pruning, and kept-set coverage correlates with accuracy but not with mean confidence, which accounts for the overconfidence drop. If right, pruning evaluations should report calibration alongside accuracy.

What carries the argument

The load-bearing object is the SCOPE selection score, score(v)=Δcov(v;S)·a(v)^α, where Δcov is the facility-location coverage gain of adding token v to the selected set and a(v) is the CLS attention score. Setting α=0 isolates pure coverage; larger α pulls selection toward saliency. The α-sweep is the controlled instrument that attributes calibration changes to the selection rule. A secondary mechanism is the gap power exponent p in the coverage objective, which the paper varies over {1.0, 1.2, 1.5, 2.0} and finds is not always optimal at its default p=1. Confidence is measured from first-token logits over a restricted verbalizer (yes/no or A–D), giving a consistent but narrow definition of

What would settle it

Run the same α-sweep with a full-sequence or free-form confidence measure on another MLLM (e.g., Qwen2-VL) and check whether pure coverage still beats random pruning at matched budgets; or directly compare FastV with pruned tokens removed versus zeroed—if zeroing reproduces the 50% accuracy collapse, the saliency-is-worse conclusion needs re-examination.

Watch

Extended reading notes

Core claim

The central discovery is an empirical dissociation between accuracy and calibration under visual token pruning. With SCOPE's pure-coverage setting (α=0) at K=128, expected calibration error on POPE drops from 0.041 to 0.016 while accuracy stays at 87.1% versus 86.9% unpruned. The α-sweep within SCOPE shows a monotone trend: lowering the saliency weight improves ECE at every tested budget with accuracy moving little, while random pruning shows no such dose-response. Across pruning conditions, kept-set coverage correlates strongly with accuracy (Spearman ρ=+0.89), not at all with mean confidence (ρ=−0.03), and inversely with overconfidence (ρ=−0.92). The authors interpret this as evidence cove

Load-bearing premise

The comparison rests on treating first-token, restricted-verbalizer confidence as a valid calibration measure across pruning conditions, on LLaVA-1.5-7B on POPE being representative enough to generalize 'selection rule matters,' and on the FastV baseline being a faithful remove-tokens implementation rather than the zeroing variant that the paper warns can collapse accuracy to chance.

Editorial extensions

If this is right

  • Moderate coverage-based pruning can improve calibration (ECE 0.041→0.016 on POPE) while preserving accuracy, so efficiency and reliability do not have to trade off.
  • Pruning evaluations that report only accuracy can miss large changes in confidence quality; the paper argues calibration metrics (ECE, Brier, AURC) should accompany accuracy.
  • Selection rule, not compression budget, drives calibration behavior: at the same budget, pure coverage beats saliency-only and random pruning on POPE.
  • The calibration gain transfers to selective prediction: abstaining on low-confidence samples at 80% coverage reaches 94.2% accuracy under coverage pruning versus 92.8% unpruned.
  • On ScienceQA-IMG, pruning reduces ECE with stable or slightly improved accuracy, but absolute ECE values are higher than on POPE and the paper does not claim coverage orders calibration there.

Reading between the lines

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

  • Editorial inference: if the evidence-coverage account is right, diversity-based pruning that maximizes the spread of retained tokens should also improve calibration, which is testable with existing diversity selectors.
  • Editorial inference: the FastV zeroing-versus-removal pitfall suggests published FastV calibration numbers could depend on implementation details; re-running FastV with true token removal is a direct check.
  • Editorial inference: the first-token, restricted-verbalizer confidence is a narrow proxy; sequence-level or free-form confidence measures might change the magnitude of the ECE gains while leaving the selection-rule ordering intact.
  • Editorial inference: the U-shaped calibration curve hints at an optimal evidence budget per task; finding that budget automatically via coverage thresholds would make the pruning-calibration benefit deployment-ready.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper studies how visual token pruning affects the calibration of multimodal large language models, measured by ECE, Brier score, and AURC. Using LLaVA-1.5-7B on POPE and ScienceQA-IMG, the authors compare SCOPE variants with different saliency weights (α=0, 0.5, 1), saliency-only pruning, random pruning, and FastV across token budgets (K=32–576). The main empirical claim is that on POPE, pure-coverage SCOPE at K=128 reduces ECE from 0.041 to 0.016 with no statistically significant accuracy drop, while saliency-based selection becomes less calibrated than random pruning. The paper also analyzes the gap power exponent p in the coverage objective and finds the default p=1 is not always optimal. The authors conclude that token pruning evaluation should include calibration and that the token selection rule, not merely the compression budget, drives calibration behavior.

Significance. If the main result holds, the paper makes a useful contribution by drawing attention to calibration as an evaluation axis for visual token pruning in MLLMs. The internal α-sweep is a clean, controlled comparison within the SCOPE framework, and the finding that moderate pruning can reduce ECE without accuracy loss is non-obvious and practically relevant. The paper is honest about its limitations, including the restricted-verbalizer confidence measure. However, the contribution is entirely empirical and currently rests on a single model-benchmark pair for the headline result; broader generalizations appear only in the abstract. The manuscript also does not mention code or data release, which limits reproducibility. The paper's significance will depend on whether the abstract's additional claims can be substantiated and whether the confidence measure is robust.

major comments (4)
  1. [Abstract vs. full text] The abstract claims kept-set coverage correlations (ρ=+0.89, −0.03, −0.92), generalization to GQA, LLaVA-NeXT, and Qwen2-VL, and a FastV zeroing-vs-removal pitfall. None of these analyses appear in the full text, which presents only POPE and ScienceQA-IMG results with LLaVA-1.5-7B. This is a load-bearing mismatch: the abstract is the primary statement of the paper's contributions, and unsupported claims must either be moved to the body and supported or removed from the abstract.
  2. [§3.2, Tables 1–3, §6 Limitations] All headline ECE numbers use a restricted-verbalizer, first-token confidence obtained by normalizing softmax probabilities over 'yes'/'no' (or A–D) only. The paper's own Limitations section concedes that 'other confidence definitions may reveal additional details.' The concern is that pruning may shift probability mass onto off-verbalizer tokens differently across selection rules, so the constructed confidence could be an artifact. No analysis of off-verbalizer mass or an alternative confidence measure is reported. Because this measure underlies every method comparison, the central claim that coverage-based pruning improves calibration is not yet robustly established.
  3. [§4.3, Table 3] FastV is reported to achieve 50.1% accuracy at K=128, which the abstract attributes to a zeroing-vs-removal implementation pitfall. The full text, however, neither describes the precise implementation nor verifies that it follows the original FastV procedure (which removes tokens, not zeroes them). If the 50.1% figure is an artifact of zeroing, then the conclusion that saliency-based FastV is catastrophically worse than coverage-based pruning is invalid. The paper must clarify the FastV implementation and, ideally, reproduce the published FastV pipeline to confirm the reported degradation.
  4. [§4.3, Table 3] The comparison between SCOPE α=0, saliency-only, and random pruning at K=128 is confounded by accuracy differences (87.1% vs. 84.4% vs. 83.6%). The claim that 'the selection rule matters' is supported by the α-sweep, where accuracy differences are small, but the external-baseline comparison could reflect an accuracy–calibration trade-off rather than a selection-rule effect. The paper should report calibration at matched accuracy levels, or use selective prediction metrics (e.g., coverage–risk curves) to separate these effects.
minor comments (5)
  1. [Table 1 vs. Table 2/3] The default SCOPE (α=1) ECE at K=128 is reported as 0.024 in Table 1 but 0.023 in Tables 2 and 3. Please reconcile the rounding.
  2. [Figure 2(b)] The full unpruned model is listed as ECE=0.040 in Figure 2(b), but the text and Tables 1/3 report 0.041. Please ensure consistency.
  3. [Title and Abstract] The arXiv metadata title is 'When Does Visual Token Pruning Improve Calibration? The Role of Evidence Coverage in MLLMs', while the full text title is 'Does Visual Token Pruning Improve Calibration? An Empirical Study on Confidence in MLLMs'. The two abstracts also differ substantially. The manuscript should have a single, consistent title and abstract.
  4. [§3.2] The explanation of ECE binning and bootstrap intervals is terse. Please specify the binning scheme details (e.g., equal-width vs. equal-frequency) and report confidence intervals for all key tables, not only Table 1.
  5. [§5] The gap power analysis is disconnected from the calibration theme. Consider moving it to an appendix or clearly indicating that it is a secondary analysis of accuracy, not calibration.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical measurement study with no derivation that reduces to its inputs.

full rationale

This paper does not present a derivation chain that could be circular. It compares empirical ECE/Brier/AURC outcomes across pruning selectors on fixed benchmarks, using a fixed confidence extraction rule (Section 3.2). The headline result—coverage-based SCOPE α=0 at K=128 lowers ECE from 0.041 to 0.016—is a measured outcome, not a consequence of how the metric is defined. The α-sweep varies a selection objective parameter and then measures calibration with an independent LLM forward pass; no parameter is fitted to the ECE values and then reported as a prediction. The FastV, saliency-only, and random baselines are externally computed comparisons, and the paper explicitly concedes in its Limitations section that 'our confidence definition is based on first-token answer probabilities over a restricted verbalizer set... other confidence definitions may reveal additional details.' That is a measurement-validity caveat, not a circular step: it does not make any reported equation equal to an input by construction. There are no load-bearing self-citations (the cited SCOPE and FastV works are by other author groups), no imported uniqueness theorem, and no ansatz smuggled in via citation. The gap-power analysis is a parameter sweep over 196 experiments, not a fitted prediction. The only related risk—that the restricted verbalizer confidence could confound cross-selector ECE comparisons—is an empirical confound, not a circularity, and would require additional off-verbalizer-mass evidence to substantiate. Thus the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central claim depends on three hand-chosen hyperparameters (α, p, K), on the validity of first-token confidence, and on one-model representativeness. The explanatory 'evidence coverage' mechanism is introduced but not independently measured, and the FastV comparison assumes a correct implementation the abstract itself calls a pitfall.

free parameters (3)
  • α (SCOPE saliency exponent) = 0.0 / 0.5 / 1.0 (sweep)
    The main calibration claim is the trend that ECE decreases as α decreases; values are hand-chosen endpoints, and the trend could be non-monotonic outside this range.
  • p (gap power exponent) = 1.0 / 1.2 / 1.5 / 2.0; p=1.5 highlighted
    Gap-power analysis sweeps four values and reports p=1.5 as best on MME at K=64 and K=192; this is data-dependent selection, not a pre-specified hypothesis.
  • K (token budget) = 32–576; K=128 selected as best
    The 'moderate pruning improves calibration' conclusion selects K=128 empirically on POPE; no theory predicts the optimum.
assumptions (4)
  • domain assumption First-token normalized verbalizer probabilities are a valid measure of model confidence for calibration comparisons.
    Section 3.2 defines confidence as max(P_yes,P_no)/(P_yes+P_no) and max P(c) for ScienceQA; if pruning changes tokenization or answer phrasing, this measure may not reflect true confidence.
  • domain assumption LLaVA-1.5-7B on POPE and ScienceQA-IMG is representative of MLLM visual token pruning generally.
    Main experiments use one model and two benchmarks; the abstract's claim that selector ordering generalizes requires this assumption.
  • domain assumption The SCOPE scoring rule is score(v)=Δcov(v;S)·a(v)^α, matching the open-source implementation.
    Section 3.1 says 'based on our reading of the open-source SCOPE implementation'; an implementation mismatch would invalidate the α-sweep.
  • domain assumption FastV's 'real, 2-pass' implementation correctly removes pruned tokens rather than zeroing them.
    The abstract flags zeroing-vs-removal as an evaluation pitfall; the body does not rule it out for the reported 50.1% accuracy collapse.
invented entities (1)
  • Evidence-coverage account of calibration
    purpose: Explains why coverage-based pruning reduces overconfidence: the retained token set still covers the image's evidence, so confidence tracks correctness.
    The abstract and Discussion invoke kept-set coverage as the mechanism, but the body never operationalizes or measures kept-set coverage; it is a post hoc explanatory construct, not a tested entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Does Visual Token Pruning Improve Calibration? The Role of Evidence Coverage in MLLMs." pith.science (2026). https://pith.science/paper/YLPXIENM

@misc{pith2026260412035,
  author       = {Pith},
  title        = {Pith review of: When Does Visual Token Pruning Improve Calibration? The Role of Evidence Coverage in MLLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YLPXIENM}},
  note         = {Machine review of arXiv:2604.12035}
}
abstract

Visual token pruning is widely used to reduce the inference cost of multimodal large language models (MLLMs), but it is usually evaluated only by accuracy. We study how pruning affects calibration, defined as the agreement between confidence and correctness, and show that the selection rule matters more than the token budget alone. On POPE with LLaVA-1.5, coverage-based pruning from 576 to 128 tokens reduces expected calibration error from 0.041 to 0.016 without a statistically significant accuracy loss. In contrast, attention-based selection preserves confidence while accuracy deteriorates, becoming less calibrated than random pruning at aggressive budgets. Across pruning conditions, kept-set coverage is strongly associated with accuracy (Spearman $\rho=+0.89$) but not with mean confidence ($\rho=-0.03$), producing a strong inverse relation with overconfidence ($\rho=-0.92$); controlled kept-set interventions support this evidence-coverage account. Two boundaries limit it: query-conditioned FastV is more overconfident than its coverage predicts, and on language-prior-dominated ScienceQA, coverage ceases to order calibration. The selector ordering otherwise generalizes to GQA and LLaVA-NeXT, and coverage beats random on Qwen2-VL, although the calibration gain over the unpruned model is task- and model-dependent. We also identify an evaluation pitfall in FastV: zeroing rather than removing pruned tokens can reduce accuracy to chance. Visual token pruning therefore changes confidence quality as well as efficiency, and calibration should be evaluated alongside accuracy when comparing pruning methods.

Figures

Figures reproduced from arXiv: 2604.12035 by the authors.

Figure 1
Figure 1. Study design. An image-question pair is encoded into V =576 visual tokens. We compare five token selection strategies: SCOPE with α=1 (default hybrid), SCOPE with α=0.5, SCOPE with α=0 (pure coverage), saliency-only pruning based on CLS attention, and random pruning. The selected K tokens are then passed to the LLM. We extract first-token class probabilities and evaluate calibration with ECE, Brier score, and AURC. … view at source ↗
Figure 2
Figure 2. (a) ECE heatmap across α and K on POPE. The lowest ECE at each tested token budget is achieved at α=0, and the best overall setting in this sweep is (α=0, K=128). (b) ECE decreases as the saliency weight is reduced for all three token budgets. For reference, we also show the random baseline, the saliency-only baseline at K=128, and the full unpruned model. 4.8 Selective Prediction We also evaluate selective predicti… view at source ↗
Figure 3
Figure 3. Main results on POPE. (a) ECE under default SCOPE across token budgets, with accuracy shown for reference. The best calibration is observed at K=128. (b) Comparison between SCOPE and random pruning at two token budgets. (c,d) Reliability diagrams for the full model (K=576) and SCOPE pruning at K=128. The high-confidence region shows a smaller confidence-accuracy gap after pruning [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Change in MME score relative to the default SCOPE [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

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 of 10

    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 of 10

    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

19 extracted references · 1 linked inside Pith · cited by 1 Pith paper

  1. [1]

    DivPrune: Diversity-based visual token pruning for large multimodal models

    Saeed Ranjbar Alvar, Gursimran Singh, Mohammad Akbari, and Yong Zhang. DivPrune: Diversity-based visual token pruning for large multimodal models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2025

    2025

  2. [2]

    AgilePruner: An empirical study of attention and diver- sity for adaptive visual token pruning in large vision-language models.arXiv preprint arXiv:2603.01236, 2026

    Changwoo Baek, Jouwon Song, Sohyeon Kim, and Kyeongbo Kong. AgilePruner: An empirical study of attention and diver- sity for adaptive visual token pruning in large vision-language models.arXiv preprint arXiv:2603.01236, 2026

    arXiv 2026

  3. [3]

    An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models

    Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. InProceedings of the European Conference on Computer Vision, 2024. 8

    2024

  4. [4]

    OTPrune: Distribution-aligned visual token pruning via optimal transport

    Xiwen Chen, Wenhui Zhu, Gen Li, Xuanzhao Dong, Yu- jian Xiong, Hao Wang, Peijie Qiu, Qingquan Song, Zhipeng Wang, Shao Tang, Yalin Wang, and Abolfazl Razi. OTPrune: Distribution-aligned visual token pruning via optimal transport. arXiv preprint arXiv:2602.20205, 2026

    arXiv 2026

  5. [5]

    Unveiling uncertainty: A deep dive into calibration and performance of multimodal large language models

    Zijun Chen, Wenbo Hu, Guande He, Zhijie Deng, Zheng Zhang, and Richang Hong. Unveiling uncertainty: A deep dive into calibration and performance of multimodal large language models. InProceedings of the 31st International Conference on Computational Linguistics, 2025

    2025

  6. [6]

    FLoC: Facility location-based efficient visual token compression for long video understanding

    Janghoon Cho, Jungsoo Lee, Munawar Hayat, Kyuwoong Hwang, Fatih Porikli, and Sungha Choi. FLoC: Facility location-based efficient visual token compression for long video understanding. InInternational Conference on Learning Representations, 2026

    2026

  7. [7]

    SCOPE: Saliency-coverage oriented token pruning for effi- cient multimodal LLMs

    Jinhong Deng, Wen Li, Joey Tianyi Zhou, and Yang He. SCOPE: Saliency-coverage oriented token pruning for effi- cient multimodal LLMs. InAdvances in Neural Information Processing Systems, 2025

    2025

  8. [8]

    Prune redundancy, preserve essence: Vision token compression in VLMs via synergistic importance-diversity

    Zhengyao Fang, Pengyuan Lyu, Chengquan Zhang, Guang- ming Lu, Jun Yu, and Wenjie Pei. Prune redundancy, preserve essence: Vision token compression in VLMs via synergistic importance-diversity. InInternational Conference on Learning Representations, 2026

    2026

Show all 19 references
  1. [9]

    Evaluating object hallucination in large vision-language models

    Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. InProceedings of the 2023 Con- ference on Empirical Methods in Natural Language Processing, 2023

    2023

  2. [10]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. InAdvances in Neural Information Processing Systems, 2023

    2023

  3. [11]

    LLaV A-NeXT: Improved rea- soning, OCR, and world knowledge

    Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaV A-NeXT: Improved rea- soning, OCR, and world knowledge. https://llava-vl. github.io/blog/2024-01-30-llava-next/ ,

    2024

  4. [12]

    Learn to explain: Multimodal reasoning via thought chains for science question answering

    Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. InAdvances in Neural Information Processing Systems, 2022

    2022

  5. [13]

    Better relia- bility compression: Model pruning with calibrated uncertainty estimation for mobile deep learning applications

    Runyu Ma, Songqing Chen, and Shuochao Yao. Better relia- bility compression: Model pruning with calibrated uncertainty estimation for mobile deep learning applications. In2025 IEEE 3rd International Conference on Mobility, Operations, Services and Technologies (MOST), 2025

    2025

  6. [14]

    Uncovering the hidden cost of model compression

    Diganta Misra, Muawiz Chaudhary, Agam Goyal, Bharat Run- wal, and Pin-Yu Chen. Uncovering the hidden cost of model compression. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2024

    2024

  7. [15]

    Investigat- ing calibration and corruption robustness of post-hoc pruned perception CNNs: An image classification benchmark study

    Pallavi Mitra, Gesina Schwalbe, and Nadja Klein. Investigat- ing calibration and corruption robustness of post-hoc pruned perception CNNs: An image classification benchmark study. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2024

    2024

  8. [16]

    CDPruner: Be- yond attention or similarity: Maximizing conditional diversity for token pruning in MLLMs

    Qizhe Zhang, Mengzhen Liu, Lichen Li, Ming Lu, Yuan Zhang, Junwen Pan, Qi She, and Shanghang Zhang. CDPruner: Be- yond attention or similarity: Maximizing conditional diversity for token pruning in MLLMs. InAdvances in Neural Informa- tion Processing Systems, 2025

    2025

  9. [17]

    VL- Uncertainty: Detecting hallucination in large vision- language model via uncertainty estimation.arXiv preprint arXiv:2411.11919, 2024

    Ruiyang Zhang, Hu Zhang, and Zhedong Zheng. VL- Uncertainty: Detecting hallucination in large vision- language model via uncertainty estimation.arXiv preprint arXiv:2411.11919, 2024

    2024 arXiv

  10. [18]

    Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, and Shanghang Zhang

    Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis A. Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, and Shanghang Zhang. Sparse- VLM: Visual token sparsification for efficient vision-language model inference. InProceedings of the 42nd In...

    2025

  11. [19]

    Calibrated self-rewarding vision lan- guage models

    Yiyang Zhou, Zhiyuan Fan, Dongjie Cheng, Sihan Yang, Zhaorun Chen, Chenhang Cui, Xiyao Wang, Yun Li, Linjun Zhang, and Huaxiu Yao. Calibrated self-rewarding vision lan- guage models. InAdvances in Neural Information Processing Systems, 2024. 9

    2024

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.