Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

VISTA: Enhancing Vision-Text Alignment in MLLMs via Cross-Modal Mutual Information Maximization

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The cross-entropy loss that trains multimodal LLMs hides a vision-text alignment term that shrinks as text grows, and a weighted mutual-information regularizer called VISTA restores it.

desk verdict A cheap and plausibly useful alignment trick, buried under a theory that is internally contradictory and does not survive even a first read. read the letter →

arxiv 2505.10917 v2 pith:HP4JBRWK submitted 2025-05-16 cs.CV

Mingxiao Li , Na Su , Fang Qu , Zhizhou Zhong , Ziyang Chen , Yuan Li , Zhaopeng Tu , Xiaolong Li This is my paper · ORCID
classification cs.CV
keywords multimodallargelanguagemodelsvision-textalignmentmutualinformationmaximizationcross-entropylossexplicitobjectivevisualquestionansweringLLaVAmodalityimbalance
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

This paper tries to show that the ordinary cross-entropy loss used to train multimodal large language models contains an implicit vision-text alignment term that loses strength as the generated text gets longer, making the model lean on language and underuse the image. The proposed fix, VISTA, adds an explicit alignment loss that maximizes an estimate of the mutual information between each text token and the visual hidden state, weighted by position so the visual signal is not diluted at later tokens. The authors report that adding this single loss term to two LLaVA-style baselines improves average performance across more than a dozen visual question answering, reasoning, and perception benchmarks, without any additional trainable modules, architectural changes, or extra data. If the claim holds, a known modality imbalance in this model family can be reduced by changing only the training objective.

What carries the argument

The argument runs through the chain-rule decomposition of mutual information, $I(x_t; S^I, s_{<t}) = I(x_t; S^I) + I(x_t; s_{<t} \mid S^I)$, used to expose the implicit alignment hidden inside cross-entropy. The practical engine is the VISTA regularizer, $\mathcal{L}_{\mathrm{VISTA}} = \frac{1}{m}\sum_{t=1}^m \frac{t}{m}\|x_t - S^I_n\|_2^2$, which pulls each text token embedding toward the aggregated visual hidden state with a position-dependent weight. The weighting $f(t) = t/m$ is the component that keeps the visual term from being diluted at later tokens while preventing the regularizer from overwhelming the loss as sequence length changes.

What would settle it

Compute $I(x_t; s_{<t})$ for any finite-vocabulary distribution and check the Appendix A.3 inequality: since $I(x_t; s_{<t}) \le H(x_t) \le \log|\mathcal{V}|$, the lower bound $t\Delta H - \epsilon$ cannot hold for large $t$, so the ratio in Theorem 3.3 has no proven limit of zero. A direct alternative test is to train a model with and without VISTA and measure the estimated visual mutual information $I(x_t; S^I)$ at successive token positions on long answers; if it does not decay, the loss is helping for a reason other than the stated theorem.

Watch

Extended reading notes

Core claim

The central claim is that the autoregressive cross-entropy objective implicitly optimizes a sum of two mutual information terms: $I(x_t; S^I)$, the alignment between the next text token and the visual hidden state, and $I(x_t; s_{<t} \mid S^I)$, the text-only autoregressive dependency. The paper argues that the textual term grows with sequence length while the visual term is bounded by the entropy of a single token, so the relative contribution of vision to predicting the next token tends to zero as text length grows (Theorem 3.3). VISTA is the proposed countermeasure: an explicit alignment loss that estimates the visual mutual information with the squared Euclidean distance between the token embedding and the final visual hidden state, weighted by $f(t) = t/m$ so the regularizer stays balanced against the cross-entropy term. The claimed outcome is that the visual contribution no longer vanishes, and the model's visual grounding improves, most visibly on vision-heavy benchmarks such as MMStar and MME Cognition.

Load-bearing premise

The degradation theorem depends on the assumption that the previous text tokens keep adding new information about the next token without limit, so that $I(x_t; s_{<t})$ grows at least linearly with $t$; in fact this quantity is bounded by the entropy of a single token, a constant.

Editorial extensions

If this is right

  • Any MLLM trained with cross-entropy can adopt VISTA during pretraining or instruction tuning without new modules, parameters, or data.
  • Long outputs are where the correction should bind most strongly, since the $t/m$ weighting focuses alignment pressure on later tokens.
  • Vision-heavy benchmarks should benefit most; the reported gains on MMStar, MME Cognition, and DocVQA are the direct empirical signature of the claimed mechanism.
  • The choice of mutual information estimator matters: the L2 variant outperforms the cosine variant in the reported ablations, which the paper attributes to different optimization dynamics.

Reading between the lines

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

  • If the mechanism is right, other explicit visual regularizers, such as contrastive, reconstruction, or geometric losses, could be plugged into the same weighted objective, and the fixed $t/m$ schedule could be replaced by a learned or content-dependent weight.
  • The paper does not address uninformative images; when the visual features carry little signal, forcing alignment to them could hurt, so a data-dependent weight is a natural testable extension.
  • The linear-growth assumption in the proof is not needed for the proposed loss to be tested empirically; measuring the visual contribution at successive token positions on long-form vision tasks would directly show whether the decay the theory predicts actually occurs.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. VISTA proposes an information-theoretic analysis of the standard cross-entropy training objective for multimodal LLMs, claiming that it implicitly maximizes a sum of mutual information terms, that the visual alignment contribution vanishes as response length grows (Theorem 3.3), and that an explicit alignment loss with a linear-in-position weighting f(t)=t/m prevents this degradation. The method adds an L2-distance-based regularizer between text token embeddings and the final visual hidden state, requiring no extra parameters or data, and is evaluated on TinyLLaVA (3B) and LLaVA-v1.5 (7B) across over a dozen benchmarks, reporting consistent gains.

Significance. If valid, the theoretical claim would provide a principled justification for a parameter-free, generally applicable alignment regularizer, and the empirical results across two architectures and many benchmarks suggest genuine practical value. The paper also deserves credit for releasing code, specifying training details, and including a limitations section. However, the central information-theoretic derivations contain load-bearing errors: the cross-entropy loss is not the model's conditional entropy, and the proof of Theorem 3.3 relies on an unbounded growth assumption that contradicts the boundedness of mutual information. These errors undermine the paper's claimed theoretical foundation, and the L2 surrogate for mutual information is unvalidated. As a result, the paper's core explanatory contribution is not established, although the empirical recipe may still be useful.

major comments (3)
  1. [Section 3.1, Eq. (3)] The equation L_CE = -1/m sum H(x^T_t | S^I, s^T_<t) equates the training cross-entropy with the model's conditional entropy, but L_CE is an expectation over the data distribution p_data while H is the entropy of the model's predictive distribution q. In standard notation L_CE = H(p_data, q) = H(p_data) + KL(p_data || q), which is not equal to H(q); a model that always outputs a fixed token has zero conditional entropy but arbitrarily large cross-entropy. Therefore the decomposition in Eq. (4) does not describe the actual training objective, and Theorem 3.3, which builds on it, does not apply to cross-entropy training.
  2. [Appendix A.3 / Theorem 3.3] The proof lower-bounds I(x_t; s_<t) by t*DeltaH - epsilon, but mutual information is bounded above by H(x_t) <= log|V|, a constant independent of t. Thus the denominator I(x_t; S^I, s_<t) cannot grow unboundedly, and the claimed limit of the ratio to zero is not established. Lemma 3.1 bounds the entropy of the prefix, not the mutual information between the current token and the prefix; the step from H(x_<t) >= (t-1)DeltaH to I(x_t; s_<t) >= t*DeltaH - epsilon in Appendix A.3 is invalid.
  3. [Section 4.2, Eq. (9), and Appendix B] The squared Euclidean distance between x_t and S^I_n is introduced as a surrogate for I(x_t; S^I) without any derivation, bound, or validation; the paper does not show that minimizing this distance maximizes mutual information or that the approximation error is controlled. Since Theorem 3.3 is invalid, Appendix B's argument that f(t)=t/m prevents the visual proportion rho_VISTA(t) from vanishing is unsupported. The weighting function and the loss coefficient are thus free parameters chosen by heuristic reasoning, not by the theory.
minor comments (5)
  1. [Section 3.1, notation] The notation is inconsistent: Section 3.1 says 'The embedding dimension is V,' but Lemma 3.2 uses log|V| for the vocabulary size; V should denote vocabulary size throughout.
  2. [Appendix C.3] The text for RefCOCO+ is duplicated verbatim in two consecutive paragraphs, which should be corrected.
  3. [Abstract and Section 5.2] The abstract states 'MME Cognition (+8.5%)' and Section 5.2 states '+8.52% relative gain,' but the table reports the absolute increase from 348.20 to 377.85; please specify that the percentage is relative and clarify the aggregation used in Tables 1-3.
  4. [Figures 2 and 3] The captions say 'Darker colors indicate higher semantic alignment,' but the color maps do not include a color scale; please add a scale or describe the normalization.
  5. [Section 5.1 / Appendix D] The main text should state that the TinyLLaVA baselines are derived from the authors' own runs on the LLaVA dataset rather than official numbers, as this is only mentioned in Appendix D and affects comparability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's theoretical derivation is partially invalid, but its claims are not equivalent to their inputs by construction.

full rationale

I walked the derivation chain from Eq. (1) through Theorem 3.3 and Appendix B. The paper does not fit any parameter to the benchmarks it later predicts, and it does not invoke a load-bearing self-citation. The weighting f(t)=t/m is fixed a priori (Sec. 4.2), and Appendix B motivates it as a design choice; Appendix E explicitly concedes that the weighting is not guaranteed optimal and that the mutual-information estimate is a tractable approximation. The theoretical chain does contain serious mathematical errors: Eq. (3) identifies the cross-entropy loss with the model's conditional entropy, although cross-entropy is a cross quantity between the data and model distributions; Eq. (4) drops the marginal entropy H(x_t) when equating entropy minimization with mutual-information maximization; and Appendix A.3 asserts I(x_t; s_<t) >= t*DeltaH - epsilon, which is impossible because I(x_t; s_<t) <= H(x_t) <= log|V|. These are incorrect inferences, not circular reductions: the theorem's conclusion is not assumed as a premise, the 'implicit objective' is not a fitted quantity, and the empirical benchmark gains are independent of the flawed proof. A mistaken justification for a method is a correctness risk, not a circularity. The one self-citation (Li et al., 2025) appears only in a related-work list and is not load-bearing. No step reduces, by the paper's own equations, to its own inputs.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central derivation leans on the false linear-growth premise for mutual information, and the L2 surrogate for mutual information is unvalidated. The only free choices are the functional form of f(t) and the unit coefficient on the auxiliary loss; neither is fitted to test data, but neither is grounded by a correct theory.

free parameters (2)
  • weighting function f(t) = t/m (fixed by design, not tuned)
    The linear-in-t weighting is chosen by hand in Section 4.2 and Appendix B; it determines each token's contribution to the alignment loss and is not derived from a valid theorem.
  • alignment loss coefficient = 1
    Equation 10 adds the VISTA term with an implicit coefficient of 1; no sensitivity analysis or search over this coefficient is reported.
assumptions (4)
  • domain assumption Non-degenerate autoregressive text model: every conditional entropy H(x_k | x_<k) is at least DeltaH > 0.
    Invoked in Lemma 3.1 to obtain linear growth of prefix entropy; reasonable for natural language but not verified for the specific training datasets.
  • domain assumption Decoder-only LLM last hidden state S^I_n summarizes the whole visual input.
    Used in Section 4.2 to reduce all image tokens to one vector for the alignment loss; this is an operational approximation, not a theorem.
  • ad hoc to paper Squared Euclidean distance between x_t and S^I_n is a valid surrogate for mutual information I(x_t; S^I).
    Equation 9 simply states this; no derivation or empirical validation is provided.
  • ad hoc to paper I(x_t; s_<t) grows at least linearly with t.
    This is the key premise of Appendix A.3; it is false because mutual information is upper bounded by H(x_t) ≤ log|V|, a constant.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VISTA: Enhancing Vision-Text Alignment in MLLMs via Cross-Modal Mutual Information Maximization." pith.science (2026). https://pith.science/paper/HP4JBRWK

@misc{pith2026250510917,
  author       = {Pith},
  title        = {Pith review of: VISTA: Enhancing Vision-Text Alignment in MLLMs via Cross-Modal Mutual Information Maximization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HP4JBRWK}},
  note         = {Machine review of arXiv:2505.10917}
}
read the original abstract

Current multimodal large language models (MLLMs) face a critical challenge in modality alignment, often exhibiting a bias towards textual information at the expense of other modalities like vision. This paper conducts a systematic information-theoretic analysis of the widely used cross-entropy loss in MLLMs, uncovering its implicit alignment objective. Our theoretical investigation reveals that this implicit objective has inherent limitations, leading to a degradation of cross-modal alignment as text sequence length increases, thereby hindering effective multimodal information fusion. To overcome these drawbacks, we propose Vision-Text Alignment (VISTA), a novel approach guided by our theoretical insights. VISTA introduces an explicit alignment objective designed to maximize cross-modal mutual information, preventing the degradation of visual alignment. Notably, VISTA enhances the visual understanding capabilities of existing MLLMs without requiring any additional trainable modules or extra training data, making it both efficient and practical. Our method significantly outperforms baseline models across more than a dozen benchmark datasets, including VQAv2, MMStar, and MME, paving the way for new directions in MLLM modal alignment research.

Figures

Figures reproduced from arXiv: 2505.10917 by the authors.

Figure 1
Figure 1. VISTA framework. We prevent target degradation caused by implicit alignment of the target by introducing the explicit alignment loss LVISTAof the text token and vision hidden state. Abstract Current multimodal large language models (MLLMs) face a critical challenge in modality alignment, often exhibiting a bias towards textual information at the expense of other modalities like vision. This paper conducts a systemat… view at source ↗
Figure 2
Figure 2. Visualization of the semantic similarity between prompt text embeddings and image [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Semantic similarity visualization between MLP-generated prompt embeddings and image [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. MS-DETR: Towards Effective Video Moment Retrieval and Highlight Detection by Joint Motion-Semantic Learning

    cs.CV 2025-07 conditional novelty 6.0 of 10

    MS-DETR improves moment retrieval and highlight detection by disentangling motion and semantic video features, sharing task information between the two tasks, and training on generated auxiliary captions.

  2. Fine-Grained Zero-Shot Object Detection

    cs.CV 2025-07 conditional novelty 6.0 of 10

    The authors define fine-grained zero-shot object detection, build a 1,432-species bird benchmark (FGZSD-Birds), and show their hierarchical MSHC detector outperforms prior ZSD models on that benchmark.

  3. From Web Search towards Agentic Deep Research: Incentivizing Search with Reasoning Agents

    cs.IR 2025-06 conditional novelty 5.0 of 10

    The paper frames agentic deep research as the successor to web search and proposes, without derivation, a test-time scaling law for reasoning and search.

Reference graph

Works this paper leans on

22 extracted references · 7 canonical work pages · cited by 3 Pith papers

  1. [1]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923,

    arXiv

  2. [4]

    Locality alignment improves vision- language models

    Ian Covert, Tony Sun, James Zou, and Tatsunori Hashimoto. Locality alignment improves vision- language models. arXiv preprint arXiv:2410.11087,

    arXiv

  3. [5]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929,

    arXiv 2010

  4. [9]

    * indicates equal contribution

    URL https://github.com/EvolvingLMMs-Lab/lmms-eval. * indicates equal contribution. Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Bench- marking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023a. Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-im...

    arXiv

  5. [10]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13 , pages 740–755. Springer,

    work page 2014

  6. [11]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296–26306, 2024a. Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, ...

    arXiv

  7. [14]

    Metamorph: Multimodal understanding and generation via instruction tuning

    Shengbang Tong, David Fan, Jiachen Zhu, Yunyang Xiong, Xinlei Chen, Koustuv Sinha, Michael Rabbat, Yann LeCun, Saining Xie, and Zhuang Liu. Metamorph: Multimodal understanding and generation via instruction tuning. arXiv preprint arXiv:2412.14164,

    arXiv

  8. [15]

    Reconstructive visual instruction tuning

    Haochen Wang, Anlin Zheng, Yucheng Zhao, Tiancai Wang, Zheng Ge, Xiangyu Zhang, and Zhaoxiang Zhang. Reconstructive visual instruction tuning. arXiv preprint arXiv:2410.09575 , 2024a. Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perc...

    arXiv

Show all 22 references
  1. [16]

    Accessed: 2025-01-

    URL https://x.ai/blog/grok-1.5v. Accessed: 2025-01-

    2025

  2. [18]

    Lmms-eval: Reality check on the evaluation of large multimodal models, 2024a

    Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Lmms-eval: Reality check on the evaluation of large multimodal models, 2024a. URL https://arxiv.org/abs/2407.12772. Xiaofeng Zh...

    arXiv

  3. [19]

    Minigpt-4: En- hancing vision-language understanding with advanced large language models

    13 Enhancing Vision-Text Alignment in MLLMs via Cross-Modal MIM Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: En- hancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592,

    arXiv

  4. [20]

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Yuchen Duan, Hao Tian, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479,

    arXiv

  5. [21]

    B.4 Implications for the Weighting Function This analysis shows that by choosing f (t) to grow linearly with t, we effectively prevent the modal alignment ratio from vanishing, maintaining a stable and meaningful alignment signal throughout the sequence. Moreover, normalizing ...

    2019

  6. [26]

    Mmt-bench: A comprehensive multimodal benchmark for evaluating large vision-language models towards multitask agi

    Kaining Ying, Fanqing Meng, Jin Wang, Zhiqian Li, Han Lin, Yue Yang, Hao Zhang, Wenbo Zhang, Yuqi Lin, Shuo Liu, et al. Mmt-bench: A comprehensive multimodal benchmark for evaluating large vision-language models towards multitask agi. arXiv preprint arXiv:2404.16006,

    arXiv

  7. [256]

    During fine-tuning, the batch size is halved to 16 to accommodate increased model complexity and dataset scale

    Mixed precision (bf16) and gradient checkpointing are enabled to optimize memory footprint and training stability. During fine-tuning, the batch size is halved to 16 to accommodate increased model complexity and dataset scale. Checkpoints are saved every 500 steps, retaining a...

  8. [2014]

    A diagram is worth a dozen images

    11 Enhancing Vision-Text Alignment in MLLMs via Cross-Modal MIM Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherland...

    2016

  9. [2019]

    Kimi-vl technical report

    Kimi Team, Angang Du, Bohong Yin, Bowei Xing, Bowen Qu, Bowen Wang, Cheng Chen, Chenlin Zhang, Chenzhuang Du, Chu Wei, et al. Kimi-vl technical report. arXiv preprint arXiv:2504.07491,

    arXiv

  10. [2020]

    Multimodal autore- gressive pre-training of large vision encoders

    Enrico Fini, Mustafa Shukor, Xiujun Li, Philipp Dufter, Michal Klein, David Haldimann, Sai Aitharaju, Victor Guilherme Turrisi da Costa, Louis B´ethune, Zhe Gan, et al. Multimodal autore- gressive pre-training of large vision encoders. arXiv preprint arXiv:2411.14402,

    arXiv

  11. [2021]

    Ocr-vqa: Visual question answering by reading text in images

    12 Enhancing Vision-Text Alignment in MLLMs via Cross-Modal MIM Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDA...

    2019

  12. [2023]

    Referitgame: Referring to objects in photographs of natural scenes

    Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787–798,

    2014

  13. [2024]

    Gramian multimodal representation learning and alignment

    Giordano Cicchetti, Eleonora Grassucci, Luigi Sigillo, and Danilo Comminiello. Gramian multimodal representation learning and alignment. arXiv preprint arXiv:2412.11959,

    arXiv

  14. [2025]

    Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024a

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024a. Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Y...

    arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.