SCOPE: A New Framework for Auditable Inference-Time Adaptation of Video World Models
A recent study presents SCOPE (Score-Isolated Agentic Optimization for Video World Models), a framework aimed at enabling auditable adaptations of static video world models during inference. This research, found on arXiv (2608.15043v1), tackles a significant evaluation challenge in utilizing video world models for planning and decision-making in embodied contexts. The problem stems from the simultaneous evolution of prompts, samplers, verifiers, and selectors during inference, complicating the attribution of performance improvements and the influence of held-out feedback on the final policy. SCOPE models external controls as a typed state, modifies this state only through limited changes backed by development evidence, and solidifies the resulting policy prior to held-out evaluation. On the Physics-IQ benchmark, SCOPE shows an enhancement of +14.24 over the exact frozen base (95% CI [+8.10, +21.23]). Further controlled ablations reveal benefits from scene specification, sampling, and learned selection, while the advantage over the most competitive matched agentic baseline remains unresolved. This paper is authored by a group of researchers and has been announced as a new submission.
Key facts
- SCOPE is a framework for auditable inference-time adaptation of frozen video world models.
- It addresses the evaluation problem in video world models used for planning and embodied decision making.
- SCOPE represents external controls as a typed state and updates it through bounded changes.
- The policy is frozen before held-out evaluation to prevent feedback contamination.
- On Physics-IQ benchmark, SCOPE improves over the frozen base by +14.24 (95% CI [+8.10, +21.23]).
- Controlled ablations identify gains from scene specification, sampling, and learned selection.
- The margin over the strongest matched agentic baseline remains unresolved.
- The paper is available on arXiv with ID 2608.15043v1.
Entities
Institutions
- arXiv