acceptodds
Under review as a conference paper at ICLR 2027

A STRONG BASELINE FOR VISION ENCODER EVALUATION IN MULTIMODAL LARGE LANGUAGE MODELS

Abstract

Evaluating vision encoders requires metrics that reliably predict their downstream performance in multimodal large language models (MLLMs). Although recent studies have shown that cross-modal metrics can better capture such performance, unimodal metrics remain the dominant choice in practice. In this work, we revisit cross-modal evaluation of vision encoders through large-scale experiments. We identify important limitations in both the experimental design and methodological formulation of prior approaches. After addressing these limitations and introducing simple improvements, we propose RAVEL, a training-free method based on cross-modal nearest-neighbor retrieval. Despite its simplicity, RAVEL achieves state-of-the-art performance across our experiments, outperforming prior methods by a substantial margin. Our results demonstrate that simple cross-modal metrics, when evaluated under a careful and comprehensive setup, can provide a strong basis for evaluating vision encoders for MLLMs.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Large Language Models, Visual Encoder Selection

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_W58wOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper revisits the problem of evaluating vision encoders without training complete MLLMs. The authors propose RAVEL, a training-free method based on cross-modal nearest-neighbor agreement, enhanced with stabilized PCA whitening and patch-level similarity matching. They conduct an extensive evaluation involving 70 vision encoders, three language backbones, and 210 trained MLLMs. Experiments show that RAVEL achieves substantially higher correlations with downstream MLLM performance than existing unimodal and cross-modal evaluation methods.

Critical strengths
  1. Comprehensive and well-controlled evaluation. The paper evaluates a diverse collection of 70 vision encoders across three language backbones, using a consistent multimodal training pipeline. This provides convincing evidence for the effectiveness of RAVEL and reveals important limitations of existing cross-modal evaluation metrics.
  2. Simple and effective method with useful empirical insights. RAVEL achieves substantial improvements over existing methods without requiring additional model training. The analysis of PCA whitening and patch-level matching is particularly informative, showing how representation geometry and feature granularity affect encoder evaluation.
Critical weaknesses
  1. Limited evaluation of selection among similarly strong encoders. The main results emphasize correlations across a heterogeneous pool of 70 encoders, where performance differences can be substantial. However, practical encoder selection often involves choosing among a small set of competitive candidates. Although the paper includes a selection-based analysis, it remains unclear how reliably RAVEL distinguishes encoders with similar downstream performance.
  2. Limited generalization across MLLM configurations. All experiments use three relatively small language models, ranging from 1.5B to 1.7B parameters, and the same LLaVA-style training pipeline. It remains unclear whether the strong correlations generalize to larger language models or different multimodal architectures and training strategies. Additionally, much of the improvement comes from PCA whitening, a well-established technique, which somewhat limits the methodological novelty.
Questions for authors
  1. Could the authors report ranking correlations restricted to the top-performing encoders, such as the top 10 or top 20, and evaluate Top-1 selection regret on randomly sampled competitive candidate pools?
  2. Does RAVEL maintain its effectiveness with larger language models or different multimodal training recipes? Even a smaller-scale experiment with a different architecture would help clarify its generalizability.
  3. How sensitive is RAVEL to the distribution of the paired image-text data used for evaluation? Would the ranking remain stable when using a different dataset from the one involved in multimodal training?
  4. Given that PCA whitening substantially improves several existing metrics, could the authors further analyze the practical benefits of patch-level matching over whitened baselines, particularly when selecting among strong encoders?
5 comments
AuthorsOpenReview ↗

Response to Reviewer W58w (1/4)

We thank the reviewer for the careful assessment and for recognizing the breadth of the evaluation and the usefulness of the representation analyses.

W1: The evaluation across a heterogeneous pool may not establish how reliably RAVEL distinguishes similarly strong encoders in a small, competitive candidate set.

Response. Thank you for raising this important point! We agree that evaluating on a large, diverse encoder pool and a smaller, more competitive pool emphasizes different aspects of encoder evaluation. The former primarily assesses robustness and generalizability across diverse encoder configurations, while the latter focuses on the ability to distinguish between encoders with similar downstream performance.

Our paper primarily focuses on evaluating the generalizability of different metrics across a diverse collection of vision encoders. This contrasts with prior studies, which typically consider a limited number and diversity of encoders, potentially leading to conclusions that do not generalize well across different encoder configurations, as demonstrated in Figure 1.

Meanwhile, we would like to emphasize that the difficulty of an evaluation setting is also metric-dependent. Although RAVEL typically achieves correlations above 0.80 in our main experiments, most existing methods achieve correlations around or below 0.50. This suggests that our large-scale evaluation setting is already challenging for existing metrics, despite RAVEL's strong performance.

Nevertheless, we agree that evaluating RAVEL on a smaller, more competitive encoder subset would provide complementary insights into its ability to distinguish among high-performing encoders. To this end, we conduct additional controlled experiments on the top 20 encoders ranked by their actual downstream MLLM performance, creating a more challenging setting where the differences between candidate encoders are potentially smaller.

The corresponding results are presented in our response to Q1, in Tables R3.1a and R3.1b.

AuthorsOpenReview ↗

Response to Reviewer W58w (2/4)

Q1: Could the authors report ranking correlations restricted to the top-performing encoders, such as the top 10 or top 20, and evaluate Top-1 selection regret on randomly sampled competitive candidate pools?

Response. Thank you for this suggestion! We are happy to conduct additional experiments to evaluate RAVEL from a selection-based perspective. However, we would appreciate clarification on the precise definition of "top-1 selection regret," as the intended formulation is not entirely clear to us.

In the meantime, we report an alternative selection-based metric in Table R3.1b, which measures the average downstream performance of the final MLLM using the top-1 encoder selected by each evaluation metric.

In fact, we have already explored this evaluation approach in Appendix K ("ROBUSTNESS UNDER A SELECTION-BASED METRIC"). We chose to emphasize ranking correlations in the main paper because repeated top-1 selection over randomly sampled encoder subsets, followed by averaging, captures information closely related to ranking correlations. We therefore consider correlation-based measures a more comprehensive and straightforward way to evaluate the overall effectiveness of encoder rankings.

Nevertheless, we agree that selection-based evaluations provide a useful complementary perspective and are happy to report additional metrics if the reviewer could clarify the intended definition of "top-1 selection regret."

More specifically, our evaluation protocol for constructing the top-20 encoder pools and computing the selection-based metric is as follows:

  • Constructing the top-20 encoder pools. For each language model backbone, we compute the unweighted average of the final MLLM's performance across 11 downstream benchmarks. We then select the 20 encoders with the highest average scores to form a backbone-specific pool of high-performing encoders, using their actual downstream performance as the ground truth.
  • Computing the selection-based metric. For each top-20 pool, we enumerate all 15,504 possible five-encoder subsets ((205)=15,504\binom{20}{5}=15{,}504), rather than relying on random sampling. Within each subset, each evaluation method selects the encoder with the highest predicted score. We then record the selected encoder's actual downstream performance, measured by the unweighted average across the 11 benchmarks. Finally, we average these scores across all 15,504 subsets to obtain the selection-based performance for each evaluation method.

We report both ranking correlations and selection-based performance within the top-20 encoder pools in Tables R3.1a and R3.1b, respectively. RAVEL consistently achieves the best results under both evaluation criteria across all three language model backbones, demonstrating its effectiveness in distinguishing among similarly high-performing encoders.

Table R3.1a. Correlations within the Top-20 Pools.

MethodsQwen3-1.7B SpearmanQwen3-1.7B PearsonQwen2.5-1.5B SpearmanQwen2.5-1.5B PearsonSmolLM2-1.7B SpearmanSmolLM2-1.7B Pearson
Linear probing0.5080.4940.5700.5390.5680.365
KNN0.7610.5720.7360.6040.5570.321
Mid-Training Loss0.3640.3110.2680.330-0.014-0.011
AC Policy-0.031-0.0130.2560.1990.3930.316
Alignment probing0.7290.5080.5680.4700.3410.268
RSA-0.296-0.338-0.538-0.571-0.223-0.137
CCA-0.724-0.695-0.553-0.537-0.405-0.414
MutualNN-0.048-0.062-0.083-0.096-0.0330.039
GW0.8310.7000.6860.6750.4480.344
RAVEL0.8530.7190.7860.7250.5870.545

Table R3.1b. Top-1 Selection Metric within the Top-20 Pools.

MethodsQwen3-1.7B Top-1 scoreQwen2.5-1.5B Top-1 scoreSmolLM2-1.7B Top-1 score
Linear probing50.2752.7643.27
KNN50.9553.1243.21
Mid-Training Loss50.3052.3442.61
AC Policy50.9052.9343.15
Alignment probing51.6253.5343.10
RSA48.1151.3042.50
CCA47.5350.6741.81
MutualNN47.8350.7041.45
GW52.2353.5043.33
RAVEL52.2953.6844.03
AuthorsOpenReview ↗

Response to Reviewer W58w (3/4)

W2: Evaluation uses only three 1.5B–1.7B language models and one LLaVA-style pipeline, leaving generalization to larger models, other architectures, and training recipes unclear. Much of the gain comes from established PCA whitening, raising concerns about methodological novelty.

Q2: Does RAVEL maintain its effectiveness with larger language models or different multimodal training recipes? Even a small-scale experiment with a different architecture would help clarify its generalizability.

Response. Thank you for the comment!

Regarding novelty, we would like to clarify that our contributions lie in adapting and applying two simple yet effective methods to the important task of evaluating vision encoders for MLLMs, which has not been explored in prior work on this problem. We do not claim that the underlying methods themselves are fundamentally novel. Rather, as detailed in Sections 3.2 and 3.3, we introduce necessary modifications to make these methods applicable and effective in the encoder-MLLM evaluation setting.

While the general principles behind these methods have been explored in other domains, their adaptation to encoder-MLLM evaluation represents a novel contribution. We believe these contributions are valuable to the broader research community, particularly in facilitating the development and selection of vision encoders for next-generation MLLMs. We will make this point clearer in the next version of our paper.

Regarding more diverse evaluation settings, thank you for this suggestion! To further evaluate the generalizability of RAVEL, we conduct additional experiments using the publicly released checkpoints from Law of Vision Representation in MLLMs (LAWs) [1], with Vicuna-v1.5-7B as the language model backbone. The LAWs collection comprises 13 vision encoder configurations, including diffusion-based encoders and combinations of features from two different vision encoders. Notably, both the language model backbone and the vision encoder configurations differ substantially from those evaluated in our original experiments, providing a complementary setting for assessing RAVEL's generalizability. Importantly, we directly apply RAVEL to this new evaluation setting using the same hyperparameters as in our original experiments, without any additional hyperparameter tuning or adaptation.

As shown in Table R3.2, RAVEL achieves the highest correlations under both evaluation measures, demonstrating its ability to generalize to larger language model backbones and diverse vision encoder architectures, including diffusion-based and combined representations. These results further highlight RAVEL's robustness and transferability across substantially different model configurations. We will add these results to the next version of our paper.

Table R3.2. Correlations on the LAWs collection.

MethodsSpearmanPearson
Linear probing0.8450.800
KNN0.8040.867
Alignment probing0.8170.903
MutualNN0.6870.665
GW0.7550.602
RAVEL0.8510.923

Q3: How sensitive is RAVEL to the distribution of the paired image-text data used for evaluation? Would the ranking remain stable when using a different dataset from the one involved in multimodal training?

Response. Thank you for raising this question!

To investigate RAVEL's sensitivity to the data used for metric estimation, we conduct additional experiments using 1,000 image-text pairs from DOCCI [2], an image captioning dataset that does not overlap with our downstream evaluation benchmarks. We deliberately choose DOCCI to introduce a substantial distribution shift from the data used in our original experiments, thereby providing a more rigorous assessment of RAVEL's robustness to unseen evaluation data. Specifically, we replace only the paired image-text data used for metric estimation, while keeping the original vision encoder collection and their corresponding downstream MLLM performance scores unchanged. This allows us to isolate the effect of changing the evaluation data source. Since this change only affects cross-modal methods that rely on paired image-text data, we restrict this comparison to those methods.

We compute both ranking correlations for each of the three language model backbones. RAVEL consistently achieves the highest correlations under both measures across all three backbones, demonstrating that its effectiveness generalizes to unseen evaluation data.

Table R3.3. Correlations using 1,000 DOCCI data pairs.

MethodsQwen3-1.7B SpearmanQwen3-1.7B PearsonQwen2.5-1.5B SpearmanQwen2.5-1.5B PearsonSmolLM2-1.7B SpearmanSmolLM2-1.7B Pearson
RSA0.1150.1230.0440.1010.5110.592
CCA0.2300.2020.3150.1780.4120.169
GW0.3050.1910.3060.1640.3250.235
MutualNN-0.1310.100-0.0510.1510.1690.327
RAVEL0.7560.7730.8600.8360.8600.879
Reviewer_zKQvOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RAVEL, which predicts the downstream performance of combinations of visual encoders and target language models via cross-modal nearest-neighbor consistency on paired image-text samples. Building upon MutualNN, the method incorporates stabilized PCA whitening and symmetric matching of image patch sets, while pooled representations are adopted on the text side. The authors construct 210 MLLMs using 70 encoder configurations and three language models of approximately 1.5–1.7B parameters. Taking the average score across 11 tasks as the reference, they compare correlations, component contributions, dataset sizes, and computational costs.

Critical strengths
  1. Training full MLLMs one by one for visual encoder screening incurs high costs, whereas evaluating visual representations in isolation may not reflect their utility when paired with the target language model. This paper shifts the evaluation target to the compatibility between visual and linguistic representations, aligning the research problem directly with practical model selection requirements. It also reminds future work to focus on how representations are consumed, rather than only their performance on standalone visual tasks.
  2. RAVEL requires no downstream performance labels for candidate encoders and does not mandate training a full MLLM for each combination. Stabilized whitening and patch-set matching make the main steps of the method easy to comprehend and implement. Its value lies in achieving consistent prediction improvements with low design complexity, offering a clear starting point for practical model selection and future method comparisons.
Critical weaknesses

AVEL extends the cross-modal nearest-neighbor consistency framework. Its main improvements combine existing whitening techniques and local set matching. While results verify these designs are effective, the paper does not fully distinguish its new evaluation principles from the careful implementation and tuning of existing metrics.

Questions for authors

Please address the key concerns raised above.

1 comment
AuthorsOpenReview ↗

Response to Reviewer zKQv

We thank the reviewer for the constructive feedback and for recognizing the value of our work.

W1: While results verify these designs are effective, the paper does not fully distinguish its new evaluation principles from the careful implementation and tuning of existing metrics.

Response.

Thank you for your comment! We appreciate your recognition of the contributions of our experimental results. We would like to reiterate that our primary contributions lie in developing simple yet effective methods for improving the predictive performance of metrics in encoder-LLM evaluation setups, which have not been systematically explored in prior work.

Regarding the implementation of baseline methods, we would like to clarify that we primarily relied on existing implementations, making only the necessary adjustments to accommodate our experimental setup. We devoted substantial effort to ensuring that these baselines achieve competitive performance consistent with published results. For example, under our controlled FLOPs budget, we had to adapt training-based baselines to different dataset sizes. For linear probing, we incorporated batch normalization and tuned the learning rate separately for each vision encoder. To verify that linear probing was not adversely affected by the limited training data, we adopted the five-shot ImageNet-1K evaluation setting following ViT [1]. As shown in Table R2.1, our implementation achieves better performance than the official implementation [2].

For alignment probing, AC policy, and KNN, we directly used the official implementations from [4], [5], and [2], respectively. Similarly, for RSA, CCA, and MutualNN, we used the official implementations from [7], [8], and [9], respectively, with additional modifications to RSA following the recommendations of Li et al. [6].

For GW, since Li et al. [6] did not release an official implementation of their proposed metric, we developed our own implementation using the Python Optimal Transport package [3]. We also contacted the authors several months ago to request implementation details but have not yet received a response. Nevertheless, we carefully examined the formulation described in their paper and made every effort to ensure that our implementation was consistent with the proposed method.

We will clarify these implementation details in the revised paper to improve transparency and reproducibility.

Table R2.1. Five-shot ImageNet-1K classification accuracy.

ImplementationTop-1 accuracyTop-5 accuracy
Official implementation25.80%37.70%
Our implementation27.83%43.08%

Reference

[1] Alexey Dosovitskiy et al. An Image Is Worth 16x16 Words: Transformers for Image Recognition at Scale. ICLR, 2021. arXiv:2010.11929.

[2] Maxime Oquab et al. DINOv2: Learning Robust Visual Features without Supervision. TMLR, 2024. arXiv:2304.07193.

[3] Rémi Flamary et al. POT: Python Optimal Transport. Journal of Machine Learning Research, 22(78):1–8, 2021.

[4] Le Zhang, Qian Yang, and Aishwarya Agrawal. Assessing and Learning Alignment of Unimodal Vision and Language Models. CVPR, 2025. arXiv:2412.04616.

[5] Shijia Yang et al. Law of Vision Representation in MLLMs. COLM, 2025. arXiv:2408.16357.

[6] Muyang Li et al. Rethinking Model Selection in VLM Through the Lens of Gromov-Wasserstein Distance. CVPR, 2026. arXiv:2605.01325.

[7] Minyoung Huh et al. Position: The Platonic Representation Hypothesis. ICML, 2024. arXiv:2405.07987.

[8] https://github.com/scikit-learn/scikit-learn

[9] Nikolaus Kriegeskorte, Marieke Mur, and Peter Bandettini. Representational similarity analysis—connecting the branches of systems neuroscience. Frontiers in Systems Neuroscience, 2:4, 2008. doi:10.3389/neuro.06.004.2008.

Reviewer_o6uHOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies training-free prediction of how well a vision encoder will perform when paired with a given language model in an MLLM. The authors first show that existing cross-modal metrics (MutualNN, Gromov–Wasserstein) correlate well with downstream performance on small, homogeneous candidate pools, but degrade as the pool grows more heterogeneous, at which point unimodal metrics become competitive again.

They propose RAVEL, which keeps the retrieval-and-overlap formulation of MutualNN and adds two changes: (i) stabilized, L2-normalized PCA whitening fitted independently per representation space, and (ii) a symmetric Chamfer-style patch-level similarity for constructing visual neighborhoods.

The method is evaluated on 210 MLLMs built from 70 vision encoders (continuous and discrete) and three ~1.5B LLMs, trained with a LLaVA-style two-stage recipe and evaluated on 11 benchmarks. RAVEL reaches Spearman correlations of 0.82–0.90, substantially above all baselines, including methods that require training. Further analyses cover the post-processing choices, hyperparameter and sample-size sensitivity, the failure of reconstruction metrics (TokBench) for discrete tokenizers, and why GW can achieve low cost without preserving image–text correspondences.

Critical strengths

Large-scale and diverse ground truth. Training and evaluating 210 MLLMs across 70 encoders that span language-supervised, self-supervised, reconstruction-based, distilled, and discrete tokenizers is substantially more comprehensive than prior studies. The released checkpoints and per-benchmark results (Table 7) would be a useful resource for the community.

Simple method with large, consistent gains. RAVEL is training-free, easy to implement, and outperforms both training-free and training-based baselines across all three LLMs. The ablations (Table 2, Fig. 3) clearly show that stabilized whitening accounts for most of the gain, and Table 5 shows that this conditioning also benefits other metrics (RSA +0.68, MutualNN +0.33). This is a practically useful finding in its own right.

Informative diagnostic analyses. The pool-size sensitivity analysis (Fig. 1), the demonstration that free GW couplings almost entirely discard observed pairings (PairAcc 0.0077, Appendix H), and the evidence that reconstruction fidelity is neither necessary nor sufficient for MLLM utility (Fig. 6) offer useful insight into why existing metrics fail.

Critical weaknesses

Possible confounding factors. Downstream performance in Table 7 is strongly stratified by encoder family (SigLIP2/MetaCLIP at the top, SSL and discrete tokenizers lower) and by input resolution. RAVEL's high correlation over 70 heterogeneous encoders may therefore mostly reflect family or resolution differences rather than genuine encoder–LLM compatibility. This is a particular concern because the patch-level Chamfer score depends on the number of patches. No within-family correlations or controls for resolution are reported.

Questions for authors
  1. Could the authors report within-family correlations (e.g., among the SigLIP2 or MetaCLIP 2 variants), or partial correlations controlling for resolution and parameter count?
  2. Could the authors evaluate RAVEL on held-out encoders or a larger LLM (e.g., 7B) that were not used for any design decisions?
2 comments
AuthorsOpenReview ↗

Response to Reviewer o6uH

We thank the reviewer for the thoughtful assessment and for recognizing the value of our evaluation and analyses.

W1: No within-family correlations or controls for resolution are reported.

Q1: Could the authors report within-family correlations (e.g., among the SigLIP2 or MetaCLIP 2 variants), or partial correlations controlling for resolution and parameter count?

Response. Thank you for your comment! We would like to clarify that some of the results in our paper already focus on encoder subsets with specific characteristics. For example, Table 3 compares discrete vision encoders, where RAVEL substantially outperforms reconstruction-based metrics specifically designed to evaluate such tokenizers.

Nevertheless, we agree that more controlled comparisons would provide stronger evidence of RAVEL's generalizability across different encoder configurations. To address this concern, we conduct additional experiments on five encoder subsets:

  • Encoder families: Encoders within the SigLIP2 and MetaCLIP2 families (15 encoders each). Other families are excluded due to their limited number of encoders.
  • Input resolution: Encoders with an input resolution of 224 × 224 (31 encoders).
  • Model size (<500M): Encoders with fewer than 500M parameters (55 encoders).
  • Model size (≥500M): Encoders with at least 500M parameters (15 encoders).

For each subset, we report the average correlation across three language model backbones. For brevity, Table R1.1 presents only Spearman correlations.

As shown in Table R1.1, RAVEL consistently achieves the highest Spearman correlation across all five subsets, demonstrating its robustness across different encoder families, input resolutions, and model scales. These results further strengthen our original findings by showing that RAVEL's effectiveness is not limited to comparisons across heterogeneous encoder configurations, but also holds within more controlled encoder subsets.

We will incorporate these additional experimental results and the corresponding discussion into the revised paper.

Table R1.1. Spearman correlations for different encoder subsets.

MethodsSigLIP2 (15 vision encoders)MetaCLIP2 (15 vision encoders)Input Resolution 224 × 224 (31 vision encoders)Model Size < 500M parameters (55 vision encoders)Model size >= 500M parameters (15 vision encoders)
Linear probing0.7190.7870.6160.4200.677
KNN0.6010.7960.6080.4170.715
Mid-Training Loss-0.302-0.2610.2760.4170.174
AC Policy0.4550.5430.6360.5660.586
Alignment probing0.6540.7070.8740.6300.631
RSA-0.0590.2440.247-0.0840.408
CCA-0.3130.4790.5380.1650.504
MutualNN-0.6250.2900.5710.3670.574
GW0.7040.5470.2660.4420.433
RAVEL0.7310.8560.9060.8320.736

Q2: Could the authors evaluate RAVEL on held-out encoders or a larger LLM (e.g., 7B) that were not used for any design decisions?

Response. Thank you for this suggestion! To further evaluate the generalizability of RAVEL, we conduct additional experiments using the publicly released checkpoints from Law of Vision Representation in MLLMs (LAWs) [1], with Vicuna-v1.5-7B as the language model backbone. The LAWs collection comprises 13 vision encoder configurations, including diffusion-based encoders and combinations of features from two different vision encoders. Notably, both the language model backbone and the vision encoder configurations differ substantially from those evaluated in our original experiments, providing a complementary setting for assessing RAVEL's generalizability. Importantly, we directly apply RAVEL to this new evaluation setting using the same hyperparameters as in our original experiments, without any additional hyperparameter tuning or adaptation.

As shown in Table R1.2, RAVEL achieves the highest correlations under both evaluation measures, demonstrating its ability to generalize to larger language model backbones and diverse vision encoder architectures, including diffusion-based and combined representations. These results further highlight RAVEL's robustness and transferability across substantially different model configurations. We will add these results to the next version of our paper.

Table R1.2. Correlations on the LAWs collection.

MethodsSpearmanPearson
Linear probing0.8450.800
KNN0.8040.867
Alignment probing0.8170.903
MutualNN0.6870.665
GW0.7550.602
RAVEL0.8510.923

Reference

[1] Yang et al., Law of Vision Representation in MLLMs, COLM 2025. arXiv:2408.16357.

Thank you for your reply. This resolves my question, and I will maintain my positive score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.