V-Zero: Answer-Label-Free On-Policy Distillation with Contrastive Evidence Gating for Fine-Grained Visual Reasoning

📄 arXiv: 2606.25319v1 📥 PDF

作者: Haoxiang Sun, Zhihang Yi, Langxuan Deng, Yuhao Zhou, Peiqi Jia, Jian Zhao, Li Yuan, Jiancheng Lv, Tao Wang

分类: cs.CV

发布日期: 2026-06-24

🔗 代码/项目: GITHUB


💡 一句话要点

提出V-Zero以解决细粒度视觉推理中的标签依赖问题

🎯 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: 细粒度视觉推理 无监督学习 对比学习 多模态模型 视觉证据门控 蒸馏训练 智能监控 自动驾驶

📋 核心要点

  1. 现有方法在细粒度视觉推理中依赖于昂贵的监督和验证规则,限制了其效率和灵活性。
  2. V-Zero通过对比证据门控实现无答案标签的视觉推理,利用学生自采样的轨迹进行训练。
  3. 实验结果显示,V-Zero在多个基准上表现优异,训练速度显著提升,且泛化能力强。

📝 摘要(中文)

细粒度视觉推理要求多模态大语言模型(MLLMs)识别与任务相关的视觉证据,并将推理基于局部图像区域。现有的代理方法通常依赖于强化学习和监督微调,导致探索成本高、验证规则设计复杂或对文本监督的重度依赖。为避免外部答案标签,本文提出了一种无答案标签的视觉推理框架V-Zero,利用对比证据门控来评估学生采样的轨迹。实验表明,V-Zero在多个视觉推理基准上显著提升了细粒度视觉推理的效果,同时保持了强大的泛化能力。V-Zero的训练速度比之前的监督微调方法快5倍,比强化学习基线快10倍。

🔬 方法详解

问题定义:本文旨在解决细粒度视觉推理中对外部答案标签的依赖问题。现有方法通常需要昂贵的监督和复杂的验证规则,导致效率低下。

核心思路:V-Zero提出了一种无答案标签的框架,通过对比证据门控来评估学生自采样的轨迹,从而实现有效的视觉推理。这样的设计避免了对外部标签的依赖,提升了训练的灵活性和效率。

技术框架:V-Zero的整体架构包括问题相关区域的视觉裁剪和负视觉视图的配对,利用这些信息对学生的轨迹进行评估,并进行密集的token级蒸馏。主要模块包括对比证据门控和轨迹评估机制。

关键创新:V-Zero的核心创新在于其无答案标签的训练方式,通过对比证据门控实现了有效的token级校正,而不依赖于传统的监督学习方法。与现有方法相比,V-Zero在训练效率和推理能力上有显著提升。

关键设计:在V-Zero中,关键参数设置包括对比证据的选择和轨迹评估的策略。损失函数设计上,采用了针对token级别的蒸馏损失,确保了模型在学习过程中能够有效地进行自我校正。

🖼️ 关键图片

fig_0
fig_1
fig_2

📊 实验亮点

V-Zero在多个视觉推理基准上的实验结果显示,其性能显著优于传统方法,训练速度比监督微调方法快5倍,比强化学习基线快10倍。这一成果表明,V-Zero在细粒度视觉推理中具有强大的实用性和效率。

🎯 应用场景

V-Zero的研究成果在多个领域具有潜在应用价值,包括智能监控、自动驾驶、医疗影像分析等。通过提升细粒度视觉推理的效率和准确性,V-Zero能够为这些领域提供更为智能化的解决方案,推动相关技术的发展与应用。

📄 摘要(原文)

Fine-grained visual reasoning requires multimodal large language models (MLLMs) to identify task-relevant visual evidence and ground their reasoning in local image regions. Existing agentic methods typically rely on reinforcement learning with verifiable rewards or supervised fine-tuning on large-scale annotated reasoning traces, leading to costly exploration, hand-designed verification rules, or heavy dependence on textual supervision. A natural way to avoid such external answer labels is to learn from trajectories sampled by the student itself, which points to On-Policy Distillation (OPD). To understand what OPD can and cannot provide for visual reasoning, we revisit it as negative-free stop-gradient alignment. This perspective shows that, although OPD provides effective token-level correction, its ceiling is constrained by the absence of trajectory-level discrimination. Motivated by these observations, we propose V-Zero, an answer-label-free framework for visual reasoning with contrastive evidence gating. V-Zero uses no annotated textual answer labels; instead, during training it pairs a question-relevant regional crop with a negative visual view to evaluate student-sampled trajectories and gate dense token-level distillation. Experiments on multiple visual reasoning benchmarks show that V-Zero consistently improves fine-grained visual reasoning while preserving strong generalization. Notably, V-Zero is more than 5$\times$ faster than previous supervised fine-tuning methods and more than 10$\times$ faster than reinforcement learning baselines. Code and dataset will be released at https://github.com/eVI-group-SCU/V-Zero