SimpleVQA: Multimodal Factuality Evaluation for Multimodal Large Language Models

TL;DR

SimpleVQA introduces the first bilingual visual question answering benchmark focused on factuality for multimodal LLMs. It provides 2,025 high-quality Q&A pairs across 9 tasks and 9 domains, with strict guidelines, a five-step data pipeline, and a human-in-the-loop quality system, evaluated against 18 multimodal and 8 text-only LLMs using an LLM-as-a-judge framework. The work reveals substantial factuality gaps in current models, highlights the distinct roles of visual understanding and internalized knowledge, and presents atomic-question analyses (CFQ) to diagnose error sources. By offering a static, easily evaluable, and cross-linguistic benchmark, SimpleVQA aims to drive more trustworthy and reliable multimodal AI systems with broad practical impact across domains and languages.

Abstract

The increasing application of multi-modal large language models (MLLMs) across various sectors have spotlighted the essence of their output reliability and accuracy, particularly their ability to produce content grounded in factual information (e.g. common and domain-specific knowledge). In this work, we introduce SimpleVQA, the first comprehensive multi-modal benchmark to evaluate the factuality ability of MLLMs to answer natural language short questions. SimpleVQA is characterized by six key features: it covers multiple tasks and multiple scenarios, ensures high quality and challenging queries, maintains static and timeless reference answers, and is straightforward to evaluate. Our approach involves categorizing visual question-answering items into 9 different tasks around objective events or common knowledge and situating these within 9 topics. Rigorous quality control processes are implemented to guarantee high-quality, concise, and clear answers, facilitating evaluation with minimal variance via an LLM-as-a-judge scoring system. Using SimpleVQA, we perform a comprehensive assessment of leading 18 MLLMs and 8 text-only LLMs, delving into their image comprehension and text generation abilities by identifying and analyzing error cases.

Figures (14)

• Figure 1: An example from our proposed SimpleVQA.
• Figure 2: An overview of the data construction process of SimpleVQA.
• Figure 3: Calibration of LLMs based on their stated confidence. The x-axis represents the confidence level of the LLMs, and the y-axis represents the accuracy.
• Figure 4: F-score results for eight different models across nine task categories.
• Figure 5: Nine task categories SimpleVQA smaples of SimpleVQA.