多模态奖励基准 2：评估面向交错图文的全能奖励模型
Multimodal RewardBench 2: Evaluating Omni Reward Models for Interleaved Text and Image
为视觉生成模型（文生图、图像编辑、交错生成）提供了首个统一的奖励模型评测基准，可直接用于评估和选择生成模型的偏好对齐信号，并揭示当前多模态奖励模型与人类判断的差距。
原文依据 We introduce Multimodal RewardBench 2 (MMRB2), the first comprehensive benchmark for reward models on multimodal understanding and (interleaved) generation.
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论奖励模型（RM）对于训练大语言模型（LLM）至关重要，但对于处理交错图像和文本序列的全能模型而言，其研究仍不充分。我们提出了多模态奖励基准 2（MMRB2），这是首个针对多模态理解与（交错）生成的奖励模型的综合基准。MMRB2 涵盖四项任务：文本到图像、图像编辑、交错生成和多模态推理（“用图像思考”），每项任务提供 1,000 个由专家标注的偏好对，这些偏好对来自 23 个模型和智能体，覆盖 21 个源任务。MMRB2 的设计遵循：（1）实用但具有挑战性的提示；（2）来自最优模型和智能体的回复；（3）通过集成过滤策略整理的、具有强人类专家共识的偏好对。利用 MMRB2，我们研究了每个子任务现有的评判者，包括多模态 LLM 作为评判者以及使用人类偏好训练的模型。GPT-5 和 Gemini-2.5-Pro 等顶级评判者达到 66-75% 的准确率，而人类超过 90%，且这些顶级评判者优于常用的 GPT-4o（59% 准确率）。表现最佳的开源模型 Qwen3-VL-32B 达到了与 Gemini-2.5-Flash 相似的准确率（64%）。我们还表明，MMRB2 的性能与下游任务成功强烈相关，并进行了深入分析，指出了未来改进奖励模型的关键领域。
英文原文摘要 1360 字符
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. Top judges like GPT-5 and Gemini-2.5-Pro reach 66-75% accuracy, compared to >90% for humans, and outperform the commonly used GPT-4o (59% accuracy). 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, and conduct an in-depth analysis that shows key areas to improve the reward models going forward.