Are We on the Right Way for Evaluating Large Vision-Language Models?

Large vision-language models (LVLMs) have recently achieved rapid progress, sparking numerous studies to evaluate their multi-modal capabilities. However, we dig into current evaluation works and identify two primary issues: 1) Visual content is unnecessary for many samples. The answers can be directly inferred from the questions and options, or the world knowledge embedded in LLMs. This phenomenon is prevalent across current benchmarks. For instance, GeminiPro achieves 42.9% on the MMMU benchmark without any visual input, and outperforms the random choice baseline across six benchmarks over 24% on average. 2) Unintentional data leakage exists in LLM and LVLM training. LLM and LVLM could still answer some visual-necessary questions without visual content, indicating the memorizing of these samples within large-scale training data. For example, Sphinx-X-MoE gets 43.6% on MMMU without accessing images, surpassing its LLM backbone with 17.9%. Both problems lead to misjudgments of actual multi-modal gains and potentially misguide the study of LVLM. To this end, we present MMStar, an elite vision-indispensable multi-modal benchmark comprising 1,500 samples meticulously selected by humans. MMStar benchmarks 6 core capabilities and 18 detailed axes, aiming to evaluate LVLMs' multi-modal capacities with carefully balanced and purified samples. These samples are first roughly selected from current benchmarks with an automated pipeline, human review is then involved to ensure each curated sample exhibits visual dependency, minimal data leakage, and requires advanced multi-modal capabilities. Moreover, two metrics are developed to measure data leakage and actual performance gain in multi-modal training. We evaluate 16 leading LVLMs on MMStar to assess their multi-modal capabilities, and on 7 benchmarks with the proposed metrics to investigate their data leakage and actual multi-modal gain.

Paper

References (67)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer sNEM6/10 · confidence 4/52024-06-13

Summary

This paper investigates the evaluation of large vision-language models (LVLMs) and the currently used benchmarks. Within the paper two primary issues are identified: the lack of need for visual information, and data leakage. Based on these issues a new compiled benchmark is proposed MMStar that includes a set of multi-modal samples validated by humans, and a variety of models are evaluated on this new benchmark.

Strengths

With the rapid progress of the field of LLM and LVLMs, it is crucial that we have reliable benchmarks for evaluation - this work explores existing benchmarks, identifies potential issues therein, and proposes a new benchmark which contributes to the quality of evaluation in the field. Additionally, the work proposes new metrics for evaluation and performs a benchmark evaluation of various models.

Weaknesses

1) Random Chance An issue in this work is that it insufficiently accounts for random guessing and data bias. As shown in [27], the ground-truth chance for it being answer A or B on MMbench are 26.4% (due to having questions with less than 4 options), and moreover, certain models may prefer certain answers. LLMs may get correct answers even when they require visual input. Given the 26.4% baseline for MMbench from [27], it is also surprising that the random choice value reported in this paper, for MMbench, is 0.0 in both Table 1 and Table 2. Moreover, an additional baseline based on majority class may also be necessary here. For the results in Figure 2, which require 6 out of 8 models to get a hit it is unlikely that random guessing has a big influence. However, for the results in Table 1 this is not the case. For instance, it appears that all MMBench results in Table 1 are below the 26.4% from [27] - which means all LLM do worse than always answering A (or B). For other benchmarks there may similarly be data biases in which answer is more frequent. Relatedly, it may be the case that such data biases about which options are more frequently chosen are more pronounced for multi-modal questions, i.e., when asked about colour the answer is more often grey across all datasets and settings - which means LLM do not learn this data bias as their training data does not include such questions, but LVLM may learn this because the same bias is present in their training data. While this is unlikely to fully explain the phenomenon observed in Table 2, it may explain part of it and not be directly related to data leakage. 2) Not all benchmarks The issues identified with existing benchmarks do not hold evenly across the benchmarks tested. In particular, MMBench and slightly lesser, MathVista seem to do pretty well with respect to these issues. This also appears to be reflected in the construction of MMStar, where after Manual Review the proportion of questions from these two existing benchmarks jumps considerable - in the end making up almost half of the MMStar benchmark. Which raises questions whether there is any benefit of using MMStar versus simply using MathVista and MMBench. 3) Manual review The manual review description is insufficiently clear. From the three criteria applied, only the first one is somewhat clear - for the other two it is unclear how the 'experts' judged this. I would expect description (can be in appendix) of agreement rates between these experts for these criteria, as well as further description of what these criteria entail. 4) New metrics The newly proposed metrics MG and ML are somewhat unclear in what they measure. If the 1500 questions in MMStar all require visual input as determined by the manual review, then what additional information does the MG metric give? Given the discussion above about chance, it appears the MG metric is more of a correction for random guessing. The ML metric similarly doesn't account for random guessing, or the aforementioned potential for data biases.

Questions

The paper raises a number of interesting questions and points out limitations of (some) existing benchmarks. Unfortunately, the data leakage issue appears to be not fully disentangled from the potential of data biases. I would appreciate input from the authors on this point. Additionally, I would be interested in discussion on random guessing within the paper, and whether the ideas presented in [27] may address this - and then subsequently how this influences the findings in the paper. Also taking into account the majority class versus random guess.

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

I believe question 9 in the paper checklist has been answered incorrectly - or at least the justification given is not valid. Given that the paper discusses a benchmark that may include images of people or copyrighted material it is crucial that the authors affirm whether the work has been done in accordance to the ethical guidelines. Even if the data is a compilation of existing datasets - by selecting and combining information from these the new resulting dataset may be biased in ways the original datasets were not (e.g., by selecting only those images containing users from a certain demographic).

Reviewer JQaJ7/10 · confidence 4/52024-06-25

Summary

The current benchmarks used to evaluate Vision Language Models (VLMs) contain several flaws. In particular, a lot of questions can be answered without looking at the image at all. These benchmarks still being hard, the best proprietary models without looking at the images can obtain better scores than strong VLM baselines (looking at the images). As a result, the authors create MMStar, a difficult benchmark aiming at evaluating the capability of the vision-language tasks. They manually review their benchmark, and do several ablations to confirm its importance.

Strengths

- The authors provide a benchmark that is hard to be good at without looking at the images. This is a problem for some questions in MMMU and MathVista currently. - The authors manually reviewed the examples of the benchmark. - The dataset is nicely divided into 6 subtasks, evaluating different aspects. The fact that there is exactly the same number of examples in each of these subtasks is appreciated. - The fact that each question is a MCQ, instead of an open-ended question that would be difficult to evaluate due to the different output formats of the models, is also appreciated.

Weaknesses

- In the released dataset, the choices are directly integrated into the prompt. It would be good to also add a column with only the original question, and another column containing the list with the possible options, so that researchers could evaluate their models with the prompts they used during their fine-tuning. - As the authors mentioned, it would be useful to also create a test set for this benchmark. - It would have probably made more sense to publish this in the Datasets and Benchmarks track.

Questions

I personally noticed many hallucinations in the SEED benchmark, with some Q/A pairs that are simply false. Since the Q/A pairs from SEED represent 28.3% of your dataset, I am worried that you would have such incorrect pairs in your benchmark. Can you confirm that this was removed during the manual filtering?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Reviewer bz7V7/10 · confidence 3/52024-07-13

Summary

In this paper, the authors examine current benchmarks for large vision-language models (LVLMs) and identify two main problems: 1) many samples do not require visual content, and 2) there is unintentional data leakage in LLM and LVLM training. To address these issues, they developed a multimodal benchmark called MMStar, consisting of 1,500 samples, and proposed two metrics to measure data leakage and performance gain in LVLMs’ multimodal training. They conducted empirical evaluations on 16 LVLMs to report their performance on MMStar.

Strengths

1. The paper is well-organized and easy to follow. 2. The motivation behind the study is clear, and the empirical analysis is thorough. 3. Data curation for MMStar is comprehensively explained. 4. The proposed performance metrics are intuitive and effectively presented.

Weaknesses

1. The authors only consider multiple-choice questions for the MMStar benchmark. Including a wider variety of well-curated questions without choices would be great. 2. Similar to Figure 2, the authors should provide the LLM Hit Rate for the MMStar benchmark. 3. What is the percentage distribution of the 1,500 samples across the four difficulty categories?

Questions

Please respond to the points of weakness I mentioned above.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes

Reviewer btYn6/10 · confidence 4/52024-07-15

Summary

The authors have identified two primary concerns with the benchmarks commonly used for large vision-language models (LVLMs). Firstly, many samples do not require visual content to answer the questions. Secondly, they noted unintentional data leakage during LVLM training. They assessed eight large language models (LLMs) across six widely-used multi-modal LLM benchmarks, demonstrating that LLMs can correctly answer a significant portion of questions without visual input. To more reliably evaluate LVLM performance, they developed a new benchmark by meticulously filtering data from six existing benchmarks with three requirements: 1) visual dependency, 2) minimal data leakage and 3) multi-modal capability for resolutions. Additionally, they designed two metrics: Multi-modal Gain, to quantify the improvement from multi-modal training, and Multi-modal Leakage, to assess the extent of potential data leakage. Using this new benchmark and the two metrics, they provide a comprehensive comparison of state-of-the-art LVLMs.

Strengths

- The paper is well-structured and clearly articulated, facilitating ease of comprehension. - The evaluation process is meticulously designed, and the conclusions drawn from it are convincing. - The findings presented in this paper meaningfully impact multi-modal large language model (LLM) research. Researchers have depended heavily on benchmarks without thoroughly examining their quality. The authors question the reliability of evaluations based on these benchmarks. Without a reliable benchmark, it is impossible to faithfully measure actual multi-modal gain. They developed a new benchmark, MMStar, which facilitates more reliable evaluations. - Using the MMStar benchmark, the authors evaluated two closed-source and fourteen open-source large vision-language models (LVLMs), with the results presented in Table 3. As expected, GPT4 emerged as the top performer in five out of six tasks. Additionally, they underscored the efficacy of smaller-scale models by highlighting that TinyLLaVa, a model with 3 billion parameters, outperformed some larger competitors with 7 billion and 13 billion parameters, thereby emphasizing the potential of smaller-scale LVLMs.

Weaknesses

- The proposed metrics, Multi-modal Gain and Multi-modal Leakage, are dependent on the base LLM utilized in the large vision-language models. This dependency complicates the use of these metrics for directly comparing the multi-modal gain across different LVLMs. - The manual review step aggressively reduces the MMStar benchmark from 11,607 samples to 1,500 samples. The explanation provided in Section 3.1 for this reduction is somewhat vague and lacks clear, objective criteria for filtering. I am curious about the rationale behind such an aggressive reduction by nearly tenfold. Is this reduction due to a scarcity of data meeting the three specified criteria mentioned between line 187 to 189, or are there other reasons for this decision?

Questions

- Please clarify the goal of using proposed metrics - Please clarify the decision to aggressively reduce the size of final benchmark

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors properly discuss the limitation in the paper

Reviewer JQaJ2024-08-07

Answer to authors

Thank you for answering the questions.

Authorsrebuttal2024-08-07

Thanks for the appreciation

Thank you for your prompt response and appreciation. Your suggestions have indeed helped us improve the quality of this work.

Reviewer sNEM2024-08-12

Thank you for the detailed reply and clarifications, with this my concerns are addressed, I will update my score to a weak accept.

Authorsrebuttal2024-08-12

Official Comment by Authors

Dear Reviewer sNEM: Thank you for your time and patience in reviewing our submission and rebuttal. Your appreciation for our work, along with the questions and suggestions you provided, has greatly helped us improve the quality of our work. Best regards and thanks, Paper 2491 Authors

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC