[论文] ViRA: Toward Ultra-High-Resolution Visual Agents via Cross-Input Relat...
论文概要
研究领域: CV
作者: Huawei Lin, Jinting Liu, Tongyan Wang, Xuyang Chen, Ruihua Shi, Zhichao Xu, Jiaxin Qin, Guoshenghui Zhao, Baijian Yang, Yingjie Victor Chen, Tan Yu, Weijie Zhao
发布时间: 2026-11-16
arXiv: 2611.21605
中文摘要
视觉智能体通过放大图像的相关区域来回答超高分辨率照片的问题。然而,现有智能体往往依赖常识而非实际读取证据来作答,在最终答案上做强化学习(RL)无法纠正这一点,因为猜测和阅读获得相同奖励。我们提出跨输入相对策略优化(CrossRPO),在对比集上训练智能体。对比集对同一图像的多个版本提出同一个问题,这些版本仅在一个包含证据的小区域上不同且答案各异,因此只有读取证据才能解出整个集合。CrossRPO 奖励解出整个集合的行为,我们证明它不会强化基于常识作答的集合。我们通过编辑 2100 万至 1.2 亿像素图像中的一个小物体构建了用于训练的 ViRA-Train 和用于评估的 ViRA-Bench,并拍摄了真实变更的集合 ViRA-Bench-Real 以排除编辑痕迹。我们仅用约 3500 条 ViRA-Train 数据训练了两个视觉推理智能体 ViRA-9B 和 ViRA-27B。CrossRPO 在最大分辨率的 ViRA-Bench 上达到 63.3% 的集合准确率,而 GRPO 为 59.3%。ViRA-9B 是所有基准上同尺寸最准确的开源模型,ViRA-27B 在多数基准上超越了最高思考级别的 Claude Haiku 5.5 和 GPT-6-Luna。
原文摘要
Visual agents answer questions about ultra-high-resolution photos by zooming into the relevant regions. However, existing agents often answer from common sense instead of reading the evidence. Reinforcement learning (RL) on the final answer cannot correct this, because a guess earns the same reward as reading. We propose Cross-Input Relative Policy Optimization (CrossRPO), which trains agents on contrast sets. A contrast set asks one question about several versions of an image that differ only in a small region with the evidence and have different answers, so only reading the evidence solves the whole set. CrossRPO rewards solving the whole set, and we prove that it does not reinforce a set answered from common sense. We build contrast sets for training (ViRA-Train) and evaluation (ViRA-Be...
自动采集于 2026-11-18
#论文 #arXiv #CV #小凯