Reproduce “CachedSearch: Training-Free Cached Exploration for Test-Time Search in Video Diffusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shreshth Saini, Neil Birkbeck, Yilin Wang, Balu Adsumilli, Alan C. Bovik
Researchers developed CachedSearch, a training-free method for video diffusion models that accelerates test-time search by generating candidates using aggressive caching for exploration and regenerating only the best candidate at full compute. This approach reduces wall-clock cost by 37% for an 8-candidate search while retaining 94.7% of the original reward gain, demonstrating efficiency and scalability across diverse model architectures.
Key insights
- Aggressive, training-free caching can largely preserve the relative ranking of candidate videos, with median per-prompt Spearman correlation exceeding 0.90 and top-1 agreement at 72% for an aggressive caching threshold.
- Ranking errors caused by caching are "self-limiting," primarily occurring among candidates with similar scores, which results in minimal regret in the overall delivered quality.
- CachedSearch is a training-free, model-agnostic, and composable method that acts as a plug-in multiplier for existing search acceleration techniques, extending its benefits multiplicatively.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.