Multimodal RewardBench 2: Evaluating Omni Reward Models for Interleaved Text and Image

TL;DR

This paper introduces Multimodal RewardBench 2, a comprehensive benchmark for evaluating reward models on multimodal understanding and generation tasks involving interleaved text and images, highlighting current model performances and areas for improvement.

Contribution

The paper presents MMRB2, the first extensive benchmark for reward models on multimodal tasks, including a diverse set of tasks, expert-annotated data, and analysis of current reward model performances.

Findings

Gemini 3 Pro achieves 75-80% accuracy.

GPT-5 and Gemini 2.5 Pro reach 66-75% accuracy.

Open-source Qwen3-VL-32B performs comparably to Gemini 2.5 Flash.

Abstract

Reward models (RMs) are essential for training large language models (LLMs), but remain underexplored for omni models that handle interleaved image and text sequences. We introduce Multimodal RewardBench 2 (MMRB2), the first comprehensive benchmark for reward models on multimodal understanding and (interleaved) generation. MMRB2 spans four tasks: text-to-image, image editing, interleaved generation, and multimodal reasoning ("thinking-with-images"), providing 1,000 expert-annotated preference pairs per task from 23 models and agents across 21 source tasks. MMRB2 is designed with: (1) practical but challenging prompts; (2) responses from state-of-the-art models and agents; and (3) preference pairs with strong human-expert consensus, curated via an ensemble filtering strategy. Using MMRB2, we study existing judges for each subtask, including multimodal LLM-as-a-judge and models trained with human preferences. The latest Gemini 3 Pro attains 75-80% accuracy. GPT-5 and Gemini 2.5 Pro reach 66-75% accuracy, compared to >90% for humans, yet surpass the widely used GPT-4o (59%). The best performing open-source model Qwen3-VL-32B achieves similar accuracies as Gemini 2.5 Flash (64%). We also show that MMRB2 performance strongly correlates with downstream task success using Best-of-N sampling and conduct an in-depth analysis that shows key areas to improve the reward models going forward.