Venue
Neural Information Processing Systems
Domain
artificial intelligence and machine learning
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.
This paper investigates the evaluation processes of Large Vision-Language Models (LVLMs) and identifies two major issues: (1) Many evaluation samples do not require visual content for correct answers, thus undermining the true assessment of multi-modal capabilities; (2) There is unintentional data leakage during training, where models can answer visual-dependent questions without visual inputs, indicating memorization of training data. To address these issues, the authors propose the MMStar benchmark, consisting of 1,500 carefully curated samples designed to ensure visual dependency and minimize data leakage. The benchmark evaluates LVLMs on six core capabilities across 18 detailed axes. Additionally, two new metrics are introduced to assess multi-modal gain and leakage. The performance of 16 leading LVLMs is evaluated on MMStar, revealing that even the best performing model scores below 60% on average, highlighting the ongoing challenges in LVLM evaluation.
This paper employs the following methods:
- Vision-language model benchmarking
- GPT-4V
- GeminiPro
- Sphinx-X-MoE
- LLaMA-70B
- InternLM2-20B
- Yi-VL-34B
- Mixtral-8x7B
- Deepseek-67B
- LLaVA series
- Qwen-7B
The following datasets were used in this research:
- MMMU
- ScienceQA
- AI2D
- SEED
- MMBench
- MathVista
- Accuracy
- multi-modal gain (MG)
- multi-modal leakage (ML)
- MMStar benchmark demonstrates the inadequacy of existing evaluations in assessing LVLM capabilities.
- First place in MMStar benchmark is GPT-4V with 57.1% accuracy.
The authors identified the following limitations:
- The study indicates that issues persist in current benchmarking methods, potentially affecting the accuracy of model evaluations.
- Number of GPUs: None specified
- GPU Type: NVIDIA A100
vision-language models
benchmark evaluation
data leakage
multi-modal capabilities