跳转至

SCOPE：面向视频世界模型的评分隔离智能优化框架

文章背景与核心概要

视频世界模型正日益成为规划与具身决策的核心模拟器。然而，在推理阶段对这些模型进行优化时，往往会引发一个微妙的评估难题：提示词、采样器、验证器和选择器往往会同步演化，这使得研究人员难以准确归因性能提升，也无法完全避免测试集反馈对最终策略的隐性污染。

SCOPE（评分隔离智能优化）框架旨在解决这一问题，通过将外部控制表示为类型化状态，仅允许在有开发证据支持的情况下进行有界更新，并在进行留出评估（held-out evaluation）前冻结策略。该方法不仅提升了模型在 Physics-IQ 基准测试上的表现，还为推理时适应（inference-time adaptation）提供了一套可审计的原则性机制。

论文概览

  • 标题： SCOPE: Score-Isolated Agentic Optimization for Video World Models
  • 作者： Yuhua Jiang, Jiaming Wang, Qingbin Liu, Feifei Gao
  • 主要学科： 人工智能 (cs.AI)
  • 提交日期： 2026年8月15日
  • arXiv ID： [arXiv:2608.15043 [cs.AI]]
  • DOI： 10.48550/arXiv.2608.15043
  • 代码仓库： GitHub - YuhuaJiang2002/SCOPE

摘要

视频世界模型正越来越多地被用作规划和具身决策的模拟器，然而在推理时对其进行改进引入了一个微妙的评估问题：提示词、采样器、验证器和选择器可能会共同演化，使得难以归因性能增益，也难以防止留出反馈塑造最终策略。

Video world models are increasingly used as simulators for planning and embodied decision making, yet improving them at inference time introduces a subtle evaluation problem: prompts, samplers, verifiers, and selectors may evolve together, making it difficult to attribute gains or prevent held-out feedback from shaping the final policy.

我们引入了 SCOPE（评分隔离智能优化），这是一个用于冻结视频世界模型进行可审计推理时适应的框架。SCOPE 将外部控制表示为类型化状态，仅通过开发证据支持的有界变更来更新此状态，并在进行留出评估之前冻结最终策略。

We introduce SCOPE (Score-Isolated Agentic Optimization), a framework for auditable inference-time adaptation of frozen video world models. SCOPE represents external controls as a typed state, updates this state only through bounded changes supported by development evidence, and freezes the resulting policy before held-out evaluation.

在 Physics-IQ 基准测试上，SCOPE 较完全冻结的基准模型提升了 \(+14.24\)（95% 置信区间 \([+8.10, +21.23]\)）。受控消融实验进一步确定了来自场景规范、采样和学习选择的增益，尽管其相对于最强匹配智能体基线的优势仍有待进一步研究。跨骨干网络和前瞻性评估揭示了一个互补的结果：有用的推理时更新确实存在，但其收益并不能在不同模型和设置之间均匀迁移。总之，这些发现表明，可靠的推理时适应不仅需要更好的建议，还需要一种原则性的机制来决定哪些更新应成为部署系统的一部分。

On the Physics-IQ benchmark, SCOPE improves over the exact frozen base by \(+14.24\) (95% CI \([+8.10, +21.23]\)). Controlled ablations further identify gains from scene specification, sampling, and learned selection, while the margin over the strongest matched agentic baseline remains unresolved. Cross-backbone and prospective evaluations reveal a complementary result: useful inference-time updates exist, but their benefits do not transfer uniformly across models and settings. Together, these findings suggest that reliable inference-time adaptation requires not only better proposals, but also a principled mechanism for deciding which updates should become part of the deployed system.

关键要点与发现

  1. 评估完整性： SCOPE 通过严格冻结策略状态，将开发探索与留出测试分离开来，从而解决了推理时适应中固有的评估漏洞。

    1. Evaluation Integrity: SCOPE resolves the evaluation loophole inherent in inference-time adaptation by separating developmental exploration from held-out testing through a strictly frozen policy state.

  2. 基准性能： 在 Physics-IQ 基准测试中，SCOPE 较冻结基准模型实现了 \(+14.24\) 的统计学显著提升。

    1. Benchmark Performance: Evaluated on the Physics-IQ benchmark, SCOPE achieves a statistically significant improvement of \(+14.24\) over the frozen base model.

  3. 组件贡献： 消融研究表明，场景规范、采样修改和学习选择机制均带来了独立的性能提升。

    1. Component Contribution: Ablation studies demonstrate individual performance boosts stemming from scene specification, sampling modifications, and learned selection mechanisms.

  4. 泛化洞察： 跨骨干网络实验表明，虽然自适应更新是有利的，但其有效性高度依赖于模型本身，这强调了建立规范化部署流程的必要性。

    1. Generalization Insights: Cross-backbone experiments indicate that while adaptive updates are advantageous, their efficacy is highly model-dependent, emphasizing the need for disciplined deployment pipelines.