CachedSearch: Training-Free Cached Exploration for Test-Time Search in Video Diffusion

What search width costs, with and without us

Every bar below is measured on Wan2.1-T2V-1.3B: a full rollout takes 68.3 s, a cached draft 34.7 s. Pick a search width to see what each strategy spends and how much of full search's reward gain survives.

What search actually buys you

One prompt per clip: an unlucky candidate on the left, the candidate CachedSearch selected on the right. Both are delivered at full compute, so search is only choosing between them. The figure is the reward gap in verifier units. Selected showcases; the aggregate is in the findings below.

Robot dancing in Times Square+2.96

A person is playing harp+3.91

A boat sailing along the Seine+3.28

A red clock+3.25

A pink clock+3.03

A book and a clock+3.03

A bench left of a truck+2.97

A couch right of a bed+2.92

A green bicycle+2.86

Measured on VBench and VBench-2.0

Every candidate generated twice from identical seeds, once at full compute and once cached, then scored by the same verifier. Configured on a pilot grid and re-measured on the full suites.

Finding 01

Rankings survive caching

Cached and full-compute orderings agree closely, which is the only property search depends on. The draft the verifier scores costs about half of the sample it stands in for.

0.905median Spearman ρ

72%top-1 agreement

2×cheaper per draft

Finding 02

When ranking breaks, it breaks cheaply

Corruption concentrates on prompts whose candidates were nearly tied to begin with, exactly where choosing wrong costs least. Most prompts lose nothing at all, so the failure mode is self-limiting rather than compounding with search width.

0median regret

+0.31corr(spread, ρ)

Finding 03

The trajectory matters, not the budget

The obvious alternative is fewer denoising steps per candidate. At a matched two-fold saving it does far worse, because a truncated rollout honestly samples a different, shorter-schedule distribution: its scores rank its own outputs, not the full-compute videos that commit will deliver. Caching perturbs the same seed-matched trajectory, which is exactly what search needs.

90.1%capture, caching

72.6%capture, truncation

6×more regret from truncation

Generality

Any published cache can be the engine

Every caching family we tested drops in unchanged and lands on one capture-versus-speedup frontier. Fidelity is set by how much compute you reuse, not by which mechanism reuses it, so a stack already running TeaCache or PAB keeps it and simply gains the explore-cheap structure.

Exploration engine	Gain capture	End-to-end cost
No caching (full best-of-8)	100.0%	621 s
PAB	99.5%	493 s
FasterCache CFG-Cache	99.6%	465 s
TeaCache	93.2%	365 s
EasyCache (our default)	90.1%	346 s

Honest edge

Keep-mode dampens motion

Keeping the cached draft instead of re-committing is cheaper still, but it dampens motion, and learned judges mildly prefer the dampened result. That is why commit is the default, and why our keep-mode claims rest on optical flow and LPIPS rather than judge scores.

Three lines in your search loop

CachedSearch wraps generation rather than replacing your model. The verifier defaults to ImageReward, the scorer behind every number here; pass your own to override it.

# pip install -r requirements.txt
from cachedsearch import cached_search

result = cached_search(pipe, prompt, n=8, tau=0.10)

result.video          # genuine full-compute sample
result.seed           # the winning seed
result.total_seconds  # about 63% of full best-of-8

Calibrated operating points

The caching threshold is the one model-specific number. Fidelity tracks architecture family rather than parameter count, so calibrate once per family and reuse it across sizes.

Cite this work

@article{saini2026cachedsearch,
  title   = {CachedSearch: Training-Free Cached Exploration for Test-Time
             Search in Video Diffusion},
  author  = {Saini, Shreshth and Birkbeck, Neil and Wang, Yilin and
             Adsumilli, Balu and Bovik, Alan C.},
  year    = {2026},
  note    = {arXiv id to be added at release}
}