MemeArena: Automating Context-Aware Unbiased Evaluation of Harmfulness Understanding for Multimodal Large Language Models

TL;DR

This work presents MemeArena, an agent-based arena-style framework for evaluating multimodal harmfulness understanding in memes. By simulating diverse interpretive contexts and aggregating judge-agent opinions, it delivers context-aware, unbiased rankings of target multimodal LLMs (mLLMs) beyond traditional binary detection benchmarks. The methodology comprises context simulation and task formulation, iterative multi-view fusion to generate value-aligned guidelines, and pairwise judgment with Elo and Bradley-Terry-based ranking. Experiments across HarM, FHM, and MAMI with 15 mLLMs show strong alignment with human preferences and reduced evaluation biases, supporting MemeArena as a scalable auditing tool for safety-focused multimodal reasoning.

Abstract

The proliferation of memes on social media necessitates the capabilities of multimodal Large Language Models (mLLMs) to effectively understand multimodal harmfulness. Existing evaluation approaches predominantly focus on mLLMs' detection accuracy for binary classification tasks, which often fail to reflect the in-depth interpretive nuance of harmfulness across diverse contexts. In this paper, we propose MemeArena, an agent-based arena-style evaluation framework that provides a context-aware and unbiased assessment for mLLMs' understanding of multimodal harmfulness. Specifically, MemeArena simulates diverse interpretive contexts to formulate evaluation tasks that elicit perspective-specific analyses from mLLMs. By integrating varied viewpoints and reaching consensus among evaluators, it enables fair and unbiased comparisons of mLLMs' abilities to interpret multimodal harmfulness. Extensive experiments demonstrate that our framework effectively reduces the evaluation biases of judge agents, with judgment results closely aligning with human preferences, offering valuable insights into reliable and comprehensive mLLM evaluations in multimodal harmfulness understanding. Our code and data are publicly available at https://github.com/Lbotirx/MemeArena.

Figures (15)

• Figure 1: An overview of previous evaluation paradigms and our proposed MemeArena framework.
• Figure 2: An illustration of context simulation, task formulation and multi-view fusion. We first simulate diverse interpretive contexts, and formulate context-specific tasks to enable target mLLMs' perspective-specific analyses, which are then iteratively and adaptively refined by the evaluator panel into a multi-view fused guideline.
• Figure 3: An illustration of the Elo rankings under MemeArena and LLM-as-a-Judge guideline settings. The order of target mLLMs in each row is ranked by the joint voting. See Figure \ref{['fig:rakings2']} for the ranking visualization of other settings.
• Figure 4: The impact of discussion rounds.
• Figure 5: An illustration of the Elo rankings under human-written guideline and w/o guideline settings. The order of target mLLMs in each row is ranked by the joint voting results under the corresponding settings.