Discriminative Visual Process Rewards for Scaling Thinking at Test-Time with Images

The “thinking with images” paradigm has led multimodal large language models to generate intermediate visual steps—such as cropping, annotation, spatial localization, and sketches—to enhance high-resolution perception and complex reasoning. However, existing multimodal Process Reward Models (PRMs) evaluate only textual reasoning and cannot judge the correctness of these visual steps, creating a key gap when visual reasoning is essential for solving tasks. We propose Discriminative Visual Process Reward Model (DiscPRM), a multimodal PRM that jointly evaluates textual and visual intermediate steps by modeling visual reasoning trajectories, image operations, and text-image consistency. To support this, we build VTReward-100K, a dataset of step-by-step visual reasoning sequences with supervision. Experiments show that using DiscPRM for Best-of-N process supervision substantially improves multimodal reasoning performance on tasks requiring visual intermediate steps, achieving over 5% gains across benchmarks. We further introduce VABench, the first benchmark for evaluating PRMs on visual reasoning error detection. We hope this work can provide foundational support for advancing the emerging direction of visual–textual process reward.

Many AI systems can now answer questions about images, but they often need to do more than simply look at the whole image at once. For difficult visual problems, it can be helpful for the system to take intermediate steps, such as zooming in on important regions, marking objects, locating areas, or drawing simple sketches before giving the final answer. However, current methods for judging an AI system’s reasoning process mainly focus on text and cannot reliably tell whether these visual steps are correct. In this paper, we propose DiscPRM, a method that can evaluate both the written reasoning and the visual steps produced by an AI system. It checks whether each step is useful, whether the image operations are appropriate, and whether the text matches the visual evidence. To train and test this ability, we build a new dataset of visual reasoning processes and introduce a benchmark for detecting errors in visual reasoning. Our experiments show that DiscPRM helps AI systems choose better solutions from multiple attempts and improves performance on tasks that require visual reasoning. We hope this work will support future AI systems that can reason more reliably with both language and images.