Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes
作者: Sophia Sirko-Galouchenko, Monika Wysoczanska, Andrei Bursuc, Nicolas Thome, Spyros Gidaris
分类: cs.CV, cs.AI, cs.CL, cs.LG
发布日期: 2026-10-01
🔗 代码/项目: GITHUB
💡 一句话要点
提出Where-OPD以解决多模态大语言模型的自我蒸馏问题
🎯 匹配领域: 支柱二：RL算法与架构 (RL & Architecture) 支柱九：具身大模型 (Embodied Foundation Models)
关键词: 自我蒸馏 多模态大语言模型 空间引导 程序生成场景 视觉理解 无注释训练 性能提升
📋 核心要点
- 现有的自我蒸馏方法在多模态大语言模型中应用不足，且依赖于人类标注的视觉信息，限制了其普适性。
- 本文提出了一种新的自我蒸馏方法，通过空间引导的文本信息来帮助模型识别与查询相关的视觉元素，避免了对外部标注的依赖。
- 实验结果表明，该方法在计数、文档和图表理解等基准测试中均表现出显著提升，且在真实世界的感知任务中也能有效迁移。
📝 摘要（中文）
近年来，基于策略的自我蒸馏已成为提升语言模型推理能力的有效方法，通过使用冻结或EMA版本的自身进行监督。然而，该方法在多模态大语言模型（MLLMs）中的应用仍然较少。现有方法通常依赖于人类标注的视觉信息，限制了其适用范围。本文提出了一种新的自我蒸馏方法，利用文本和空间引导信息来识别与查询相关的视觉元素。通过程序生成的场景，模型能够实现无注释的后训练，并在多个基准测试中显著提升性能，尤其是在真实世界的感知任务中也表现出良好的迁移能力。
🔬 方法详解
问题定义：本文旨在解决多模态大语言模型在自我蒸馏过程中对人类标注视觉信息的依赖，限制了模型的普适性和可扩展性。
核心思路：提出了一种新的自我蒸馏方法，利用程序生成的场景和空间引导的文本信息，使模型能够在无注释的情况下进行有效的后训练。
技术框架：整体架构包括教师模型和学生模型，教师模型使用空间引导信息定位相关视觉元素，而学生模型则学习如何从图像和问题中重现教师的行为。
关键创新：最重要的创新在于引入空间引导的文本信息，使得教师模型能够更有效地整合来自多个图像区域的证据，突破了传统方法的限制。
关键设计：在模型训练中，采用了程序生成的场景，确保了对象身份和空间坐标的自动获取，设计了适应性损失函数以优化教师和学生模型之间的学习过程。
🖼️ 关键图片
📊 实验亮点
实验结果显示，使用Where-OPD方法在多个基准测试中均有显著提升，尤其是在CVBench、V*、ZoomBench等真实世界感知基准中，平均性能提升达3.23分，验证了该方法的有效性和广泛适用性。
🎯 应用场景
该研究的潜在应用领域包括智能问答系统、图像理解和数据分析等。通过提升多模态大语言模型的推理能力，该方法能够在实际应用中提供更准确的结果，推动相关领域的技术进步和应用落地。
📄 摘要（原文）
On-policy self-distillation has recently emerged as an effective approach for improving language-model reasoning by supervising students with a frozen or EMA version of themselves that receives privileged information. Its application to multimodal large language models (MLLMs), however, remains largely unexplored. Recent approaches use privileged visual information, such as image crops corresponding to a question, to improve fine-grained perception, but their gains are confined to tasks that benefit from such visual zooming and require either human-annotated grounding data or external teacher models. We introduce a different form of on-policy self-distillation for MLLMs that provides the teacher with textual, spatially grounded guidance identifying the visual elements relevant to a query. We use procedurally generated scenes with automatically available object identities and spatial coordinates, enabling scalable and annotation-free post-training. The teacher uses this spatial guidance to locate and integrate evidence from multiple relevant image regions, while the student learns to reproduce the resulting behavior from the image and question alone. Our approach consistently improves performance on counting, document and chart understanding benchmarks across multiple models. Importantly, although post-training uses only synthetic scenes, the resulting improvements transfer to real-world perception benchmarks, yielding a 3.23-point gain in average performance across CVBench, V*, ZoomBench, BLINK, HR-Bench, and MME-RealWorld. These results show that spatially grounded privileged information can induce broader perceptual capabilities through on-policy self-distillation, enabling substantial synthetic-to-real transfer beyond the task and data distribution used for post-training. Project page: https://github.com/sirkosophia/Where-OPD