Regime-Aware Temporal Granularity for Causal Video World Models

The paper

Regime-Aware Temporal Granularity for Causal Video World Models: When Does Adaptivity Pay?
Md Jisan Hossen (Department of Artificial Intelligence, NUIST, Nanjing, China). Preprint, 2026. https://doi.org/10.5281/zenodo.23693517

Causal video generators produce a rollout one temporal chunk at a time, so spending expensive per-frame compute only on hard segments looks attractive. We show that this idea rests on three preconditions — cost, signal and endpoint — and that all three are regime properties rather than universal facts. On a matched two-GPU study (Tesla T4 and A100-80GB) we (i) find the same mode pair on opposite sides of the cost break-even inequality; (ii) show that hand-crafted context statistics cannot predict coarse-unsafe segments (0.491 AUROC), that an intrinsic probe on the generator's own prefill activations reaches 0.798 on one GPU and inverts to 0.143 on the other, and that only a multi-feature model over the probe and a sampler convergence residual transfers across both (0.842 / 0.833); and (iii) derive a regime map that states exactly when adaptivity pays — only when the cheap mode is cheaper and worse.

Prior work

This programme builds directly on the earlier study:

Md Jisan Hossen. Jev-Routed Temporal Granularity with Recovery and Hysteresis for Causal Video Generation. Preprint, 2026. https://doi.org/10.5281/zenodo.23395414

That study framed adaptive granularity as resting on three preconditions — cost, signal and endpoint — and found the trade-off to be conditional. This repository audits those preconditions, then turns them into a rule.

The programme in one figure

Part 1 · JevGAG — the report

A measured audit of the three preconditions on two GPUs, with a matched design, an independent endpoint protocol and a claim-to-artifact audit.

Also established: hysteresis + forced recovery stabilise switching, and a measurable seam appears at every switch boundary. → jevgag/

Part 2 · DDAG — the method

Deadline- and Drift-Aware Granularity turns the preconditions into a regime map plus calibrated components, each validated by its own experiment.

→ ddag/ (twelve numbered experiments)

Figures

Achieved · boundary · open

Every earlier limitation is recorded as a boundary — a measured limit of the naive formulation — and each either has a resolution or defines the next experiment.

The honest headline: across two GPUs and every endpoint measured, the cheap mode was never worse — so the optimal policy is a fixed mode, and the map states exactly where a router becomes worth deploying.

Data & artifacts (public)

Every result, figure and media bundle is published on Kaggle. Nothing is private except the API-key store. Each dataset ships a README and metadata; every headline number re-derives from a stored result file.

Backbone / reproducibility assets (used by the kernels, not results): causal-rcm-ckpts (13 GB) · wan21-causal-rcm-matched (4.4 GB).

Full results

More detail than the manuscript. Confidence intervals are prompt-grouped bootstrap; AUROCs use leave-one-prompt-out cross-validation grouped by prompt.

Signals for coarse-unsafe segments (direct AUROC)

Cost (DiT seconds per segment, wall-clock, synchronised)

Quality by endpoint (lower is better)

Seam repair (T4, 32 matched segments)

Drift propagation (81 frames, 6 segments)

Propagation on T4 (one-coarse minus all-fine), 95% CIs: seg 2 +0.0313 [+0.010,+0.052] · seg 3 +0.0304 [+0.009,+0.051] · seg 4 +0.0208 [+0.000,+0.042] · seg 5 +0.0241 [+0.003,+0.045] · seg 6 +0.0212 [−0.004,+0.046] — significant at four of five tail segments. On A100 the coarse chain lies below the fine chain.

Policies (the regime map)

Jev critic and error-propagation model

Cite

@misc{hossen2026regimeaware,
  title        = {Regime-Aware Temporal Granularity for Causal Video World Models: When Does Adaptivity Pay?},
  author       = {Hossen, Md Jisan},
  year         = {2026},
  howpublished = {Preprint},
  doi          = {10.5281/zenodo.23693517},
  url          = {https://doi.org/10.5281/zenodo.23693517}
}

@misc{hossen2026jevRouted,
  title        = {Jev-Routed Temporal Granularity with Recovery and Hysteresis for Causal Video Generation},
  author       = {Hossen, Md Jisan},
  year         = {2026},
  howpublished = {Preprint},
  doi          = {10.5281/zenodo.23395414},
  url          = {https://doi.org/10.5281/zenodo.23395414}
}

@software{hossen2026jevgagddag,
  title  = {JevGAG + DDAG: Regime-Aware Adaptive Temporal Granularity for Causal Video World Models},
  author = {Hossen, Md Jisan},
  year   = {2026},
  url    = {https://github.com/j1s4nn/JevGAG}
}