RBench-V: A Primary Assessment for Visual Reasoning Models with Multi-modal Outputs

TL;DR

RBench-V is a new benchmark designed to evaluate the multi-modal reasoning abilities of vision-language models through complex questions requiring image manipulation and auxiliary reasoning, revealing current models' significant limitations.

Contribution

This paper introduces RBench-V, the first benchmark focused on assessing models' multi-modal reasoning with outputs, filling a gap in existing evaluation methods.

Findings

Current models achieve only 25.8% accuracy on RBench-V.

Models significantly underperform compared to human scores of 82.3%.

RBench-V reveals substantial challenges in multi-modal reasoning for state-of-the-art models.

Abstract

The rapid advancement of native multi-modal models and omni-models, exemplified by GPT-4o, Gemini, and o3, with their capability to process and generate content across modalities such as text and images, marks a significant milestone in the evolution of intelligence. Systematic evaluation of their multi-modal output capabilities in visual thinking processes (also known as multi-modal chain of thought, M-CoT) becomes critically important. However, existing benchmarks for evaluating multi-modal models primarily focus on assessing multi-modal inputs and text-only reasoning while neglecting the importance of reasoning through multi-modal outputs. In this paper, we present a benchmark, dubbed RBench-V, designed to assess models' vision-indispensable reasoning abilities. To construct RBench-V, we carefully hand-pick 803 questions covering math, physics, counting, and games. Unlike previous benchmarks that typically specify certain input modalities, RBench-V presents problems centered on multi-modal outputs, which require image manipulation such as generating novel images and constructing auxiliary lines to support the reasoning process. We evaluate numerous open- and closed-source models on RBench-V, including o3, Gemini 2.5 Pro, Qwen2.5-VL, etc. Even the best-performing model, o3, achieves only 25.8% accuracy on RBench-V, far below the human score of 82.3%, highlighting that current models struggle to leverage multi-modal reasoning. Data and code are available at https://evalmodels.github.io/rbenchv