M$^3$-VQA: A Benchmark for Multimodal, Multi-Entity, Multi-Hop Visual Question Answering

2026-04-27

AI Summary

Existing visual question answering (VQA) benchmarks struggle to evaluate multimodal large language models’ capacity for fine-grained multi-entity understanding and complex multi-hop reasoning. To address this gap, this work proposes the first fine-grained VQA evaluation benchmark specifically designed for multimodal, multi-entity, multi-hop reasoning, integrating a structured knowledge base with explainable evidence chains. The benchmark requires models to identify multiple entities across visual and textual modalities and perform sequential or parallel reasoning across heterogeneous, multi-source documents. Experiments on 16 state-of-the-art multimodal large language models reveal a significant performance drop without external knowledge, while accuracy markedly improves when precise evidence is provided. Furthermore, reasoning-aware retrieval strategies substantially outperform heuristic approaches, demonstrating the framework’s effectiveness in enhancing task complexity and evaluation depth.

Abstract

We present M$^3$-VQA, a novel knowledge-based Visual Question Answering (VQA) benchmark, to enhance the evaluation of multimodal large language models (MLLMs) in fine-grained multimodal entity understanding and complex multi-hop reasoning. Unlike existing VQA datasets that focus on coarse-grained categories and simple reasoning over single entities, M$^3$-VQA introduces diverse multi-entity questions involving multiple distinct entities from both visual and textual sources. It requires models to perform both sequential and parallel multi-hop reasoning across multiple documents, supported by traceable, detailed evidence and a curated multimodal knowledge base. We evaluate 16 leading MLLMs under three settings: without external knowledge, with gold evidence, and with retrieval-augmented input. The poor results reveal significant challenges for MLLMs in knowledge acquisition and reasoning. Models perform poorly without external information but improve markedly when provided with precise evidence. Furthermore, reasoning-aware agentic retrieval surpasses heuristic methods, highlighting the importance of structured reasoning for complex multimodal understanding. M$^3$-VQA presents a more challenging evaluation for advancing the multimodal reasoning capabilities of MLLMs. Our code and dataset are available at https://github.com/CASIA-IVA-Lab/M3VQA.

Problem

• Visual Question Answering
• Multimodal Reasoning
• Multi-Hop Reasoning
• Entity Understanding
• Multimodal Large Language Models

Innovation

• multi-hop reasoning
• multimodal entity understanding
• knowledge-based VQA
• retrieval-augmented reasoning
• multimodal large language models