Pith. sign in

REVIEW 2 major objections 4 minor 299 references

Bayesian search over a continuous spatial-scale space lets MLLMs find tiny objects in huge images more accurately and with fewer queries than prior methods.

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 ↗

T0 review · grok-4.5

2026-07-12 04:16 UTC pith:QCTUHGEJ

load-bearing objection Solid training-free BO search for UHR MLLMs with a real continuous kernel and strong numbers; the RKHS residual assumption is untested but the empirical case still stands. the 2 major comments →

arxiv 2607.03184 v1 pith:QCTUHGEJ submitted 2026-07-03 cs.CV

BVS: Bayesian Visual Search with Multimodal Large Language Model for Fine-grained Perception

Geng Li , Yuxin Peng This is my paper
classification cs.CV
keywords Bayesian visual searchmultimodal large language modelsfine-grained perceptionultra-high-resolution imagesscale-aware kernelGP-UCBattention rolloutregret bounds
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.

Multimodal large language models still fail at fine-grained perception when images are ultra-high-resolution and the objects of interest occupy only a fraction of a percent of the pixels. Existing visual-search fixes either scan blindly and waste queries or lock onto a static prior that cannot correct its own mistakes. BVS treats the problem as global optimization of a relevance function over a continuous space of locations and zoom levels. It first builds a reasoning-aware prior by rolling attention only to the middle layers of the model, then uses a Gaussian process with a specially designed scale-aware kernel and upper-confidence-bound sampling to decide which crops to inspect next. Each local observation updates the posterior, so early biases are corrected and missing regions can still be discovered. The paper supplies a sub-linear regret guarantee under a standard RKHS residual assumption and shows clear gains over both training-free and training-based baselines on the hardest high-resolution perception benchmarks while using far fewer model calls.

Core claim

Perception of tiny objects in ultra-high-resolution scenes can be cast as Bayesian optimization of an unknown relevance function over a continuous spatial-scale manifold; an early-stop attention prior plus a scale-aware non-stationary kernel and GP-UCB acquisition yields both higher accuracy and a theoretically guaranteed sub-linear regret compared with discrete or static heuristics.

What carries the argument

Scale-Aware Non-stationary Kernel (SANK) together with GP-UCB: the kernel lets spatial correlation length grow with observation scale so that coarse and fine crops interact correctly, while GP-UCB balances exploration of unseen regions against exploitation of the current posterior.

Load-bearing premise

The residual relevance function after subtracting the attention prior must lie in the reproducing-kernel Hilbert space of the proposed kernel with a bounded norm; if the true residual is highly discontinuous or the prior is badly mis-calibrated, both the regret bound and the acquisition function lose their justification.

What would settle it

On a held-out set of ultra-high-resolution images, replace the continuous Bayesian loop with a matched number of random or pure-greedy crops; if accuracy and coverage remain statistically identical to BVS, the claimed advantage of posterior correction and the scale-aware kernel disappears.

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

If this is right

  • Tiny-object questions on multi-megapixel images become answerable with only a handful of MLLM forward passes rather than dozens.
  • Search no longer has to discretize the image into fixed grids or trust a single noisy prior; missing or noisy regions can still be recovered.
  • The same continuous spatial-scale formulation can be reused by any MLLM that can return a scalar relevance score for a crop.
  • Training-free methods can now match or exceed several trained visual-search agents while remaining model-agnostic.

Where Pith is reading between the lines

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

  • Because the method is training-free and only needs a scoring oracle, it can be dropped in front of future larger MLLMs without re-collecting search trajectories.
  • The early-stop attention construction may be useful as a cheap saliency prior even outside the Bayesian loop, for example in ordinary image cropping or captioning pipelines.
  • If the residual RKHS assumption fails systematically on certain long-tail object classes, one could replace the Matérn kernel with a deep-kernel GP or a non-GP acquisition rule while keeping the continuous manifold formulation.

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 proposes BVS, a training-free Bayesian optimization framework that treats fine-grained perception in ultra-high-resolution images as global optimization of a relevance function over a continuous spatial-scale manifold. It constructs a reasoning-aware prior via early-stop attention rollout of an MLLM, then iteratively refines a Gaussian-process posterior with a proposed scale-aware non-stationary Matérn-style kernel and GP-UCB acquisition, finally consolidating high-posterior regions into a compact visual input. Theoretical claims include positive-definiteness of the kernel (via Paciorek–Schervish) and a sub-linear regret bound under an RKHS residual assumption. Empirically, BVS is shown to outperform recent training-free (ZoomEye, RAP, ViCrop) and several training-based visual-search methods on V* Bench and HR-Bench 4K/8K across InternVL3.5 and Qwen3-VL 4B/8B models, with a favorable accuracy-vs-forward-pass trade-off and supporting ablations.

Significance. If the empirical gains hold under broader scrutiny, BVS offers a practical, training-free route that unifies prior guidance with posterior correction and demonstrably improves both accuracy and efficiency on established fine-grained UHR benchmarks. The continuous multi-scale formulation and scale-aware kernel are genuine technical contributions relative to discrete-grid or static-prior baselines. The Mercer argument is correctly grounded; the regret analysis follows the standard Srinivas et al. template and is therefore a useful formalization even if the RKHS assumption remains unverified. Reproducible public-benchmark comparisons and component ablations strengthen the case for adoption as an inference-time module.

major comments (2)
  1. [Section 3.5, Assumption 3.4, Theorem 3.5] Section 3.5, Assumption 3.4 and Theorem 3.5: the sub-linear regret guarantee and the validity of the GP-UCB intervals (Eq. 4) rest on the residual f−μ0 lying in the RKHS of the scale-aware kernel with bounded norm. MLLM relevance scores are produced by a discrete token-generation prompt (Appendix C.1) and are therefore typically discontinuous in continuous crop coordinates (s,x). No residual-norm estimates, posterior-calibration plots, leave-one-out checks, or other diagnostics are reported. Without evidence that the assumption holds, both the theoretical claim advertised in the abstract and the justification for using GP-UCB remain unsupported; the observed gains could be explained by the early-stop prior plus multi-scale sampling alone.
  2. [Section 3.4] Section 3.4: the acquisition step is written as zt+1=argmax αt(z) over the continuous manifold X=S×Ω, yet no optimizer, multi-start strategy, discretization, or gradient-based procedure is described. Continuous Bayesian optimization is non-trivial; without this detail the method is not fully reproducible and it is unclear whether the reported efficiency (≤5 forward passes) includes the cost of the continuous maximizer.
minor comments (4)
  1. [Section 4.3] Table 5 and Figure 2: the efficiency claims would be clearer if wall-clock time or total FLOPs (including acquisition maximization) were reported alongside forward-pass counts.
  2. [Section 3.2] Section 3.2 / Figure 4: the early-stop layer is fixed at L/2; a short sensitivity plot over stop-layer fraction would strengthen the claim that the choice is robust rather than dataset-specific.
  3. [Section 3.3 / Appendix B] Notation and typesetting: repeated spacing artifacts appear around Matérn (“Mat ´ern”), and the MIG exponent in the proof sketch of Theorem 3.5 is rendered ambiguously; both should be cleaned for camera-ready.
  4. [Section 4.4] Table 3 (MMStar): the modest overall drop is acknowledged, yet a brief discussion of when the consolidation step can harm pure-reasoning items would help readers decide when to apply BVS.

Circularity Check

0 steps flagged

No load-bearing circularity; empirical claims rest on external public benchmarks and the regret bound is standard GP-UCB under an explicit (non-fitted) RKHS residual assumption. Only a minor non-load-bearing self-citation of the authors' prior DyFo baseline appears.

full rationale

The derivation chain is self-contained and non-circular. The early-stop attention prior (Eq. 1, Section 3.2) is extracted from the MLLM's own intermediate layers and used only as a non-zero mean function μ0; it is never fitted to the evaluation labels. The Scale-Aware Non-stationary Kernel (Def. 3.1, Eq. 2) is proved PSD by direct reduction to the external Paciorek-Schervish non-stationary construction (Theorem 3.2, Appendix B.1); no parameter is estimated from the target benchmarks. The sub-linear regret claim (Theorem 3.5) is the classical Srinivas et al. GP-UCB bound applied to the residual f̃ = f - μ0 under the explicit Assumption 3.4 (∥f̃∥k ≤ B); the paper does not fit B or any other constant on the test data and then call the result a prediction. All accuracy numbers (Tables 1-2) are obtained by running the training-free procedure on the public V*, HR-Bench and MMStar suites against independently published baselines. The sole self-citation (DyFo, Li et al. 2025) appears only as one among several prior-free baselines in Related Work and is not invoked to justify any theorem, kernel property or experimental claim. Consequently no step reduces by construction to a fitted input or to an unverified self-citation chain.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central empirical claim rests on a handful of free hyper-parameters and on standard BO assumptions; the theoretical claim additionally rests on the residual lying in the RKHS of the new kernel. No new physical entities are postulated; the 'invented' objects are algorithmic constructs whose only evidence is the paper's own experiments.

free parameters (5)
  • confidence threshold τ = 0.2
    Default τ=0.2 used for semantic masking; chosen by hand and not ablated for sensitivity on the main tables.
  • maximum search steps T = 5
    Default T=5; efficiency claims depend on this budget.
  • prior strength θ
    Scales the attention heatmap into the GP mean μ0; value not reported.
  • scale length-scale function ℓ(s) and λs
    Define the non-stationary kernel; only required to be positive and monotonically increasing; concrete schedule left unspecified.
  • GP-UCB exploration schedule βt
    Set to the theoretical form 2B²+300γt log³(t/δ); B and the information-gain constant are not measured on the actual residual.
axioms (4)
  • domain assumption Residual relevance f−μ0 belongs to the RKHS of the scale-aware kernel with bounded norm (Assumption 3.4).
    Required for calibrated confidence intervals and the sub-linear regret statement of Theorem 3.5.
  • domain assumption MLLM crop scores are noisy observations of a latent continuous relevance function (yt=f(zt)+ε).
    Standard GP observation model; if scores are highly discontinuous the acquisition function can fail.
  • ad hoc to paper Early-stopping attention rollout at layer L/2 yields a useful reasoning-aware prior (Section 3.2).
    Justified by the observed 'attention divergence' phenomenon and by the ablation in Table 4, but not derived from first principles.
  • standard math Paciorek–Schervish non-stationary construction yields a valid PSD kernel when Σ(z) is block-diagonal with scale-dependent spatial length (Theorem 3.2).
    Direct application of a published theorem; proof sketch in Appendix B.1.
invented entities (2)
  • Scale-Aware Non-stationary Kernel (SANK) no independent evidence
    purpose: Models multi-scale spatial correlation so that coarse observations influence large neighborhoods while fine observations remain local.
    Defined in Eq. (2); positive-definiteness shown via Paciorek–Schervish; no independent evidence outside the paper's own experiments.
  • Early-stop Attention Rollout prior no independent evidence
    purpose: Extracts a reasoning-aware spatial prior from intermediate MLLM layers while avoiding shallow-layer divergence.
    Heuristic truncation at L/2; supported only by the authors' ablation and visualization (Figure 4).

pith-pipeline@v1.1.0-grok45 · 21435 in / 3303 out tokens · 32239 ms · 2026-07-12T04:16:36.450493+00:00 · methodology

0 comments
read the original abstract

While Multimodal Large Language Models (MLLMs) demonstrate impressive general capabilities, they struggle with fine-grained perception in ultra-high-resolution (UHR) images, particularly for tiny objects in cluttered scenes. Existing methods face a dilemma: they either rely on inefficient prior-free scanning, or depend on static prior-driven heuristics that lack posterior correction to rectify initial model biases. To address this, we propose BVS (Bayesian Visual Search), a framework that formulates perception as a global optimization problem over a continuous spatial-scale manifold. Specifically, BVS bridges prior guidance with posterior correction: it utilizes an early-stop attention rollout of MLLM to construct reasoning-aware priors, while employing a scale-aware non-stationary kernel and GP-UCB to dynamically rectify noise and recover missing information in the prior through iterative local observations. We provide theoretical guarantees via sub-linear regret bounds, and extensive experiments demonstrate that BVS significantly outperforms state-of-the-art baselines with a superior trade-off between accuracy and efficiency.

Figures

Figures reproduced from arXiv: 2607.03184 by Geng Li, Yuxin Peng.

Figure 1
Figure 1. Figure 1: Overview of the BVS. Our framework consists of three main stages: (1) Prior Attention Acquisition: extracting coarse regions of interest using early-stop attention rollout. (2) Active Search via Scale-Aware BO: an iterative Bayesian Optimization loop that progressively locates fine-grained visual evidence. It employs a Scale-Aware Non-stationary Kernel (SANK) to model multi-scale spatial dependencies and a… view at source ↗
Figure 2
Figure 2. Figure 2: Efficiency Analysis of BVS comparing with other training-free visual search methods [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cases study from V* Bench of BVS with Qwen3VL-8B. Qwen3-VL w/tool output is also reported to compare. Attention Rollout prior, which captures logical reasoning context with a rigorous posterior correction mechanism. A key contribution is the Scale-Aware Non-stationary Ker￾nel, which adapts the sampling strategy to dynamic ob￾servation scales, thereby preventing inefficient exploration caused by scale misma… view at source ↗
Figure 4
Figure 4. Figure 4: The Phenomenon of Attention Divergence. An attention rollout case study from V* Bench illustrating the evolution of attention maps across a full 28-layer Qwen2.5VL. Both insufficient depths (≤ 9) and excessive depths (≥ 19) fail to capture critical visual regions accurately. Relevant visual signals only become prominent within the intermediate layers, highlighting the necessity of our early-stopping mechan… view at source ↗

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

299 extracted references · 3 canonical work pages

  1. [1]

    Langley , title =

    P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

    2000

  2. [2]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    1980

  3. [3]

    M. J. Kearns , title =

  4. [4]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    1983

  5. [5]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    2000

  6. [6]

    Suppressed for Anonymity , author=

  7. [7]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    1981

  8. [8]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    1959

  9. [9]

    FirstName LastName , title =

  10. [10]

    FirstName Alpher , title =

  11. [11]

    Journal of Foo , volume = 13, number = 1, pages =

    FirstName Alpher and FirstName Fotheringham-Smythe , title =. Journal of Foo , volume = 13, number = 1, pages =

  12. [12]

    Journal of Foo , volume = 14, number = 1, pages =

    FirstName Alpher and FirstName Fotheringham-Smythe and FirstName Gamow , title =. Journal of Foo , volume = 14, number = 1, pages =

  13. [13]

    FirstName Alpher and FirstName Gamow , title =

  14. [14]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  15. [15]

    Advances in Neural Information Processing Systems , volume=

    Flamingo: a visual language model for few-shot learning , author=. Advances in Neural Information Processing Systems , volume=

  16. [16]

    Advances in Neural Information Processing Systems , volume=

    Visual instruction tuning , author=. Advances in Neural Information Processing Systems , volume=

  17. [17]

    2311.08294 , archivePrefix=

    Yuan-Fang Li Ye and Wen-Qi Wang and Kun-Hu Wang and Xiang-Rong Zhang and Zhen-Hui Liu and Jian-Wei Niu and Guang-Xuan Xiao and Wen-Gao Pu and Yu-Jie Zhang , year=. 2311.08294 , archivePrefix=

    Pith/arXiv arXiv

  18. [18]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Evaluating object hallucination in large vision-language models , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  19. [19]

    2303.11381 , archivePrefix=

    Zhengyuan Yang and Linjie Li and Jianfeng Wang and Kevin Lin and Ehsan Azarnasab and Faisal Ahmed and Zicheng Liu and Ce Liu and Michael Zeng and Lijuan Wang , year=. 2303.11381 , archivePrefix=

    Pith/arXiv arXiv

  20. [20]

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

    Visual-programming: Compositional visual reasoning without training , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  21. [21]

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

    ViperGPT: Visual inference via python execution for reasoning , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  22. [22]

    2304.09842 , archivePrefix=

    Pan Lu and Baolin Peng and Haoitsu Cheng and Michel Galley and Kai-Wei Chang and Ying Nian Wu and Yoon Kim and Jianfeng Gao , year=. 2304.09842 , archivePrefix=

    Pith/arXiv arXiv

  23. [23]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Making the V in VQA matter: Elevating the role of image understanding in visual question answering , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  24. [24]

    2306.13394 , archivePrefix=

    Chaoyou Fu and Peixian Chen and Yunhang Shen and Yulei Qin and Mengdan Zhang and Xu Lin and Zhenyu Mok and Qinghao Ye and Haoran Wei and Jingren Zhou and Ke Li and Jun-Jie-Jagger Huang , year=. 2306.13394 , archivePrefix=

    Pith/arXiv arXiv

  25. [25]

    2307.16125 , archivePrefix=

    Bohao Li and Yuanhan Zhang and Chunyuan Li and Ruoxi Chen and Chongjian Ge and Song-Chun Zhu and Jianfeng Gao and Yizhou Wang , year=. 2307.16125 , archivePrefix=

    Pith/arXiv arXiv

  26. [26]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  27. [27]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  28. [28]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  29. [29]

    arXiv preprint arXiv:2503.01222 , year=

    Retrieval-augmented perception: High-resolution image perception meets visual rag , author=. arXiv preprint arXiv:2503.01222 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint , year=

    How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites , author=. arXiv preprint , year=

  31. [31]

    LLaVA-NeXT: Improved reasoning, OCR, and world knowledge , url=

    Liu, Haotian and Li, Chunyuan and Li, Yuheng and Li, Bo and Zhang, Yuanhan and Shen, Sheng and Lee, Yong Jae , year=. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge , url=

  32. [32]

    arXiv preprint , year=

    Feast Your Eyes: Mixture-of-Resolution Adaptation for Multimodal Large Language Models , author=. arXiv preprint , year=

  33. [33]

    2023 , journal=

    CogVLM: Visual Expert for Pretrained Language Models , author=. 2023 , journal=

    2023

  34. [34]

    arXiv preprint , year=

    Phi-3 technical report: A highly capable language model locally on your phone , author=. arXiv preprint , year=

  35. [35]

    ai , author=

    Yi: Open foundation models by 01. ai , author=. arXiv preprint , year=

  36. [36]

    arXiv preprint , year=

    InternLM-XComposer2: Mastering Free-form Text-Image Composition and Comprehension in Vision-Language Large Model , author=. arXiv preprint , year=

  37. [37]

    2023 , journal=

    InternLM-XComposer: A Vision-Language Large Model for Advanced Text-image Comprehension and Composition , author=. 2023 , journal=

    2023

  38. [38]

    Introducing our Multimodal Models , url =

    Bavishi, Rohan and Elsen, Erich and Hawthorne, Curtis and Nye, Maxwell and Odena, Augustus and Somani, Arushi and Ta. Introducing our Multimodal Models , url =

  39. [39]

    Xu, Haiyang and Ye, Qinghao and Yan, Ming and Shi, Yaya and Ye, Jiabo and Xu, Yuanhong and Li, Chenliang and Bi, Bin and Qian, Qi and Wang, Wei and Xu, Guohai and Zhang, Ji and Huang, Songfang and Huang, Fei and Zhou, Jingren , booktitle =. m. 2023 , url =

    2023

  40. [40]

    2024 , journal=

    DeepSeek-VL: Towards Real-World Vision-Language Understanding , author=. 2024 , journal=

    2024

  41. [41]

    NeurIPS , title =

    Lauren. NeurIPS , title =. 2023 , volume =

    2023

  42. [42]

    arXiv preprint , year=

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

  43. [43]

    arXiv preprint , year=

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context , author=. arXiv preprint , year=

  44. [44]

    arXiv preprint , year=

    Gpt-4 technical report , author=. arXiv preprint , year=

  45. [45]

    OpenCompass: A Universal Evaluation Platform for Foundation Models , author=

  46. [46]

    CVPR , url=

    Monkey: Image resolution and text label are important things for large multi-modal models , author=. CVPR , url=

  47. [47]

    arXiv preprint , year=

    ConvLLaVA: Hierarchical Backbones as Visual Encoder for Large Multimodal Models , author=. arXiv preprint , year=

  48. [48]

    arXiv preprint , year=

    Beyond LLaVA-HD: Diving into High-Resolution Large Multimodal Models , author=. arXiv preprint , year=

  49. [49]

    arXiv preprint , year=

    Accelerating Transformers with Spectrum-Preserving Token Merging , author=. arXiv preprint , year=

  50. [50]

    ICML , year=

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models , author=. ICML , year=

  51. [51]

    ICLR , year=

    Analyzing and Mitigating Object Hallucination in Large Vision-Language Models , author=. ICLR , year=

  52. [52]

    ICCVW , year=

    Div8k: Diverse 8k resolution image dataset , author=. ICCVW , year=

  53. [53]

    arXiv preprint , year=

    Dragonfly: Multi-Resolution Zoom Supercharges Large Visual-Language Model , author=. arXiv preprint , year=

  54. [54]

    arXiv preprint , year=

    InfiMM-HD: A Leap Forward in High-Resolution Multimodal Understanding , author=. arXiv preprint , year=

  55. [55]

    arXiv preprint , year=

    Vary: Scaling up the vision vocabulary for large vision-language models , author=. arXiv preprint , year=

  56. [56]

    ICCV , year=

    Segment anything , author=. ICCV , year=

  57. [57]

    arXiv preprint , year=

    Mini-gemini: Mining the potential of multi-modality vision language models , author=. arXiv preprint , year=

  58. [58]

    CVPR , year=

    A convnet for the 2020s , author=. CVPR , year=

  59. [59]

    EMNLP , year=

    Evaluating Object Hallucination in Large Vision-Language Models , author=. EMNLP , year=

  60. [60]

    ECCV , year=

    A Diagram is Worth a Dozen Images , author=. ECCV , year=

  61. [61]

    CVPR , year=

    MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI , author=. CVPR , year=

  62. [62]

    C hart QA : A Benchmark for Question Answering about Charts with Visual and Logical Reasoning

    Masry, Ahmed and Long, Do and Tan, Jia Qing and Joty, Shafiq and Hoque, Enamul. C hart QA : A Benchmark for Question Answering about Charts with Visual and Logical Reasoning. ACL. 2022

    2022

  63. [63]

    WACV , year=

    Docvqa: A dataset for vqa on document images , author=. WACV , year=

  64. [64]

    arXiv preprint , year=

    InternLM-XComposer2-4KHD: A Pioneering Large Vision-Language Model Handling Resolutions from 336 Pixels to 4K HD , author=. arXiv preprint , year=

  65. [65]

    ICLR , year=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. ICLR , year=

  66. [66]

    ICLR , year=

    Large language models are not robust multiple choice selectors , author=. ICLR , year=

  67. [67]

    Token Merging: Your

    Bolya, Daniel and Fu, Cheng-Yang and Dai, Xiaoliang and Zhang, Peizhao and Feichtenhofer, Christoph and Hoffman, Judy , booktitle=. Token Merging: Your. 2023 , url=

    2023

  68. [68]

    TMLR , year=

    Unsupervised Dense Information Retrieval with Contrastive Learning , author=. TMLR , year=

  69. [69]

    arXiv preprint , year=

    MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models , author=. arXiv preprint , year=

  70. [70]

    Neural computation , volume=

    Backpropagation applied to handwritten zip code recognition , author=. Neural computation , volume=. 1989 , url=

    1989

  71. [71]

    CVPR , year=

    Deep residual learning for image recognition , author=. CVPR , year=

  72. [72]

    arXiv preprint , year=

    Are We on the Right Way for Evaluating Large Vision-Language Models? , author=. arXiv preprint , year=

  73. [73]

    ICML , year=

    Learning transferable visual models from natural language supervision , author=. ICML , year=

  74. [74]

    arXiv preprint , year=

    Llama 2: Open foundation and fine-tuned chat models , author=. arXiv preprint , year=

  75. [75]

    arXiv preprint , year=

    Llama: Open and efficient foundation language models , author=. arXiv preprint , year=

  76. [76]

    arXiv preprint , year=

    Qwen technical report , author=. arXiv preprint , year=

  77. [77]

    2024 , journal=

    InternLM2 Technical Report , author=. 2024 , journal=

    2024

  78. [78]

    2024 , journal=

    ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools , author=. 2024 , journal=

    2024

  79. [79]

    MMBench: Is Your Multi-modal Model an All-around Player? , year =

    Yuan, Liu and Haodong, Duan and Yuanhan, Zhang and Bo, Li and Songyang, Zhang and Wangbo, Zhao and Yike, Yuan and Jiaqi, Wang and Conghui, He and Ziwei Liu and Kai, Chen and Dahua, Lin , journal =. MMBench: Is Your Multi-modal Model an All-around Player? , year =

  80. [80]

    ICML , year=

    Mm-vet: Evaluating large multimodal models for integrated capabilities , author=. ICML , year=

Showing first 80 references.