SCOPE: Score-Isolated Agentic Optimization for Video World Models
| Source: arXiv AI
Tags: video world models, inference-time optimization, SCOPE, Physics-IQ, video generation
SCOPE introduces auditable inference-time adaptation of frozen video world models by formalizing updates as bounded typed state transitions, achieving +14.24 Physics-IQ benchmark improvement while revealing that inference-time gains do not transfer uniformly across backbone architectures.
Details
Researchers introduce SCOPE (Score-Isolated Agentic Optimization) to address a subtle evaluation problem in video world model development: when prompts, samplers, verifiers, and selectors co-evolve during adaptation, it becomes impossible to attribute which component drove improvements or to ensure held-out feedback did not contaminate the final policy. SCOPE solves this by representing external controls as a typed state that updates only through bounded changes supported by development evidence, then freezing the resulting policy before held-out evaluation. The framework achieves +14.24 points on Physics-IQ benchmark (95% CI: +8.10 to +21.23) over the frozen baseline, with controlled ablations identifying distinct gains from scene specification, sampling strategy, and learned selection. A critical cross-architecture finding: inference-time improvements do not transfer uniformly across backbone models. Useful updates for one architecture may produce no benefit on another. This limits the assumption that inference-time optimization is a free, portable capability. The paper's methodological contribution—separating adaptation from evaluation to prevent contamination—is as important as the benchmark numbers for teams building video generation pipelines with inference-time optimization loops. Code is publicly available.