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

Yuxuan Xie, Tianhua Li, Wenqi Shao, Kaipeng ZhangSubmission to ICLR 2025 · Rejected

TL;DR

TP-Eval is a new evaluation framework that addresses "prompt sensitivity" in Multimodal Large Language Models (MLLMs) by automatically customizing prompts for each specific model. By moving away from "one-size-fits-all" benchmark prompts, researchers uncovered significant hidden potential in popular models, proving that current benchmarks often underestimate model capabilities due to suboptimal instruction design.

What Problem Does This Paper Address?

Evaluating MLLMs (models that process both text and images) is notoriously tricky. Currently, researchers use benchmarks like MMT-Bench or MMMU, which provide a static set of questions and images. However, MLLMs are highly sensitive to how a question is phrased.

Impact of prompt variations on task accuracy.

As shown above, a minor tweak—like adding "focus on visual cues"—can cause a model's accuracy to jump significantly. The authors identify two major deficiencies in current evaluation paradigms:

  1. Underestimation: Inappropriate prompts obscure a model's true capabilities, making it seem less capable than it actually is.
  2. Evaluation Bias: Different models have different "preferences." A prompt that works perfectly for LLaVA might cause DeepSeek to fail. Using the exact same prompt for all models creates an unfair comparison, as the benchmark might inadvertently favor the prompt style of one model's training data over another.

Methods

The authors introduce TP-Eval, an automatic prompt optimization (APO) framework tailored specifically for the multimodal setting. The process follows an iterative loop to refine and customize instructions.

Overview of the TP-Eval framework.

The framework consists of several key components:

  1. The Scorer: The model under evaluation (the "Target Model") processes a task using a candidate prompt. An "Answer Analyzer" (powered by GPT-4o mini) extracts the choice from the raw response to calculate an accuracy score.
  2. Error Introspection: This is a key novelty. Instead of just looking at a numerical score, the system asks an LLM to analyze why the Target Model got specific examples wrong. This qualitative feedback (e.g., "the model ignored the background details") acts as a "pseudo-gradient" to guide the next iteration.
  3. The Optimizer: Using a "meta-prompt," an optimizer takes the current prompt, the accuracy score, the introspection notes, and a few examples to generate a new, improved prompt.
  4. Semantic Constraint: To prevent the optimizer from changing the core meaning of the task (semantic drift), the framework uses BERT-based cosine similarity and a "decaying edit distance" to ensure the new prompt stays true to the original question.
  5. Iteration & Re-ranking: The loop runs for several rounds (typically 16), and a final re-ranking step selects the optimal prompt based on a weighted score of accuracy and semantic consistency.

Datasets

The researchers evaluated their framework on two major MLLM benchmarks:

  • MMT-Bench (MMT-S): A subset of 83 tasks covering 19 categories of general capabilities.
  • MMMU: A massive multi-discipline benchmark requiring college-level subject knowledge and reasoning.

They used a small "validation set" (roughly 20 samples per task) to find the optimal prompt and then tested that prompt on the remaining "test set."

Key Insights and Ideas

  • Prompts are Model-Specific: The "optimal" prompt for InternVL actually decreased the performance of LLaVA and DeepSeek when applied to them. This confirms that there is no such thing as a universally perfect prompt; customization is necessary for fair benchmarking.
  • Introspection is More Powerful than Scores: Simply telling an optimizer "this prompt got 60% accuracy" isn't very helpful. Telling it "the model failed because it didn't look at the helmet's visibility" allows the optimizer to add specific instructions that fix the behavior.
  • Hidden Potential: Many models are much smarter than their benchmark scores suggest. The "gap" between a model's score on a standard benchmark and its score with a customized prompt represents the "potential" that is currently being tapped by TP-Eval.

Key Results

The researchers tested three open-source models: LLaVA-1.5-7B, DeepSeek-VL-7B, and Mini-InternVL-4B.

  • Significant Gains: On MMT-S, LLaVA’s overall score improved by 4%, with some specific tasks seeing a performance jump of 25.1%.
  • Mini-InternVL saw the highest ratio of improvement in specific tasks, with a 40.4% gain in tasks where the prompt was optimized.
  • MMMU Success: Even on the difficult MMMU benchmark, adding a customized domain-specific prefix (e.g., "Answer the questions about [Subject]") significantly boosted scores across all categories compared to the original questions.
  • Zero-Shot Capability: The authors demonstrated that they could use "In-Context Learning" (using optimized prompts from other tasks) to improve performance even when no training data is available for a new task.

Limitations and Future Work

  • Overfitting Risk: Because the validation sets in MLLM benchmarks are often very small (sometimes only 20 images), there is a risk that the optimizer creates a prompt that is too specific to those 20 images, failing to generalize to the broader test set.
  • Computational Cost: The iterative process requires multiple calls to MLLM APIs, which can be expensive and time-consuming when scaling to hundreds of benchmark tasks.
  • Future Directions: The authors suggest exploring more robust introspection mechanisms and investigating how to make the prompt optimization process more efficient to reduce the number of required iterations.

Summary updated

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.
BibTeX
@misc{xie2024tp,
  author = {Yuxuan Xie and Tianhua Li and Wenqi Shao and Kaipeng Zhang},
  title = {{TP-Eval: Tap Multimodal LLMs' Potential in Evaluation by Customizing Prompts}},
  year = {2024},
  eprint = {2410.18071},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.18071},
}
Figure