arXiv 2509.156952026-03-30cs.CVcs.LG

ORIC：在大视觉-语言模型中基于情境不一致性的物体识别基准测试

ORIC: Benchmarking Object Recognition under Contextual Incongruity in Large Vision-Language Models

University of California, San Diego（加利福尼亚大学圣迭戈分校）；University of California, Riverside（加利福尼亚大学河滨分校）；University of Southern California（南加利福尼亚大学）

Zhaoyang Li, Zhan Ling, Yuchen Zhou, Litian Gong, Erdem Bıyık, Hao Su

备注：: We request withdrawal of this paper because one of the listed institutional affiliations was included without proper authorization. This issue cannot be resolved through a simple revision, and we therefore request withdrawal to prevent dissemination of incorrect or unauthorized affiliation information

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

研究探讨了大视觉-语言模型在非典型场景中识别物体的困难，提出ORIC框架通过两种策略生成不一致的物体-情境对，并验证了情境不一致是不确定性的重要来源。

摘要：

大视觉-语言模型（LVLMs）在结合视觉和语言任务中表现出色，但在非典型场景中常出现遗漏明显物体或虚构不存在物体的问题。本文通过不确定性视角，聚焦情境不一致，即物体出现意外或未出现在预期情境中的情况，证明此类情况增加了最先进LVLMs的识别难度。为研究此领域，我们引入了Object Recognition in Incongruous Context（ORIC）框架，通过LLM引导采样和CLIP引导采样生成不一致的物体-情境对。应用于MSCOCO，ORIC创建了ORIC-Bench和ORIC风格训练数据。评估18个LVLMs和2个开放词汇检测器发现，在不一致情境下存在显著退化和偏见。对Qwen3-VL-8B-Instruct进行视觉强化微调，使用600个ORIC样本提升了ORIC-Bench、AMBER和HallusionBench的性能。整体而言，本文表明情境不一致是不确定性的关键来源，并提供了更可靠的LVLMs的工具。数据集和代码在https://github.com/ZhaoyangLi-1/ORIC上公开。

英文摘要：

Large Vision-Language Models (LVLMs) excel at captioning, visual question answering, and robotics by combining vision and language, yet they often miss obvious objects or hallucinate nonexistent ones in atypical scenes. We examine these failures through the lens of uncertainty, focusing on contextual incongruity, where objects appear unexpectedly or fail to appear in expected contexts, and show that such cases increase recognition difficulty for state-of-the-art LVLMs. To study this regime, we introduce the Object Recognition in Incongruous Context (ORIC) framework, which constructs incongruous object-context pairs through two complementary strategies: (1) LLM-guided sampling to identify hard-to-recognize objects present in the image and (2) CLIP-guided sampling to mine plausible but absent ones. Applied to MSCOCO, ORIC creates ORIC-Bench and ORIC-style training data. Evaluating 18 LVLMs and 2 open-vocabulary detectors reveals significant degradation and bias under incongruous contexts. Visual Reinforcement Fine-Tuning of Qwen3-VL-8B-Instruct on 600 ORIC samples improves performance on ORIC-Bench, AMBER, and HallusionBench. Overall, we show that contextual incongruity is a key source of uncertainty and provide tools for more reliable LVLMs. The dataset and code are publicly available at https://github.com/ZhaoyangLi-1/ORIC.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。