TP-Eval: Tap Multimodal LLMs' Potential in Evaluation by Customizing Prompts

TL;DR

A new evaluation framework named TP-Eval is introduced, which introduces a prompt customization method to reduce evaluation biases and tap models' potential, and some well-designed modules for prompt customization tailored to the scenario of MLLM evaluation are proposed.

Abstract

Recently, multimodal large language models (MLLMs) have received much attention for their impressive capabilities. The evaluation of MLLMs is becoming critical to analyzing attributes of MLLMs and providing valuable insights. However, current benchmarks overlook the problem of prompt sensitivity - minor prompt variations may lead to significant performance fluctuations. Thus, inappropriate prompts may obscure the models' capabilities, underestimating the models' performance. Moreover, different models have different preferences for different prompts, and thus, using the same prompt for all models will cause evaluation bias. This paper analyzes this deficiency in existing benchmarks and further introduces a new evaluation framework named TP-Eval, which introduces a prompt customization method to reduce evaluation biases and tap models' potential. TP-Eval will rewrite the original prompts to different customized prompts for different models. In particular, we propose some well-designed modules for prompt customization tailored to the scenario of MLLM evaluation. Extensive experiments demonstrate the effectiveness of our approach to uncovering models' capabilities, and TP-Eval should benefit the community in developing more comprehensive and convincing MLLM evaluation benchmarks.

Figures (7)

• Figure 1: (a) shows underestimation caused by unsuitable prompts in MMT-Bench, (b) shows our proposed evaluation framework resolving this by customizing prompts.
• Figure 2: The overview of our automatic prompt customization structure.
• Figure 3: Results of different models on MMT-S (L2-category). Accuracy improvement is calculated by accuracy using the optimized prompt divided by accuracy using the original prompt. Three models showed varying improvement across different task types, while performance gains differ between models, highlighting the underestimation and bias introduced by original prompts and the effectiveness of our method.
• Figure 4: Overall performance with different prompt methods on MMMU with LLaVA. In most cases, the results after optimization surpass those achieved with the initial prompts, and they generally outperform the original questions as well.
• Figure 5: Result of applying optimized prompts to other models. Applying customized prompts from one model to another yields performance changes that differ from each model's inherent characteristics.