Source
Assran et al. (2025), LIT-215 — ARXIV-2506.09985.
The rule
The expensive axis of an input — resolution, frames per clip, sequence length — does not have to be constant across training. Hold it low through warmup and the constant phase, then raise it during the decay.
Measured: 252K iterations at 16 frames and low resolution, then 12K cooldown iterations at raised resolution. Against the projected cost of training at full resolution throughout, that is up to an 8x speedup in GPU-days for a ViT-g.
The schedule is LIT-215’s, introduced as progressive-resolution training in the V-JEPA 2 recipe: Assran et al. ran it to pretrain a 1B-parameter video encoder on over a million hours of video, and report the 8x as a projection against full-resolution training rather than a head-to-head run to matched quality.
Why the decay phase specifically
This is what makes it a rule rather than a trick, and it is why the practice extends SOTA-140 rather than standing alone. A warmup-stable-decay schedule already concentrates the learning-rate commitment into a short final window: the constant phase explores, the decay converges. The observation here is that the decay phase is also the right place to pay for input fidelity — the model spends most of its steps learning what to look for at low cost, and the expensive representation is trained only where the weights are settling.
The dependency runs the right way: without a schedule that has a decay phase,
there is no principled place to put the resolution increase, which is why
this is stated against SOTA-140 and not against cosine.
What else that paper’s sweep says, and what it does not license
The same section attributes its gains separately, which is unusually clean:
| change | gain |
|---|---|
| data 2M → 22M videos | +1.0 |
| model 300M → 1B (ViT-g/16) | +1.5 |
| training 90K → 252K iterations | +0.8 |
| spatial resolution and clip length ↑ | to 88.2%, +4.0 cumulative |
Those are reasons to believe the resolution axis matters, not evidence for this scheduling rule — they are about the end state, and the rule is about how to get there cheaply. The 8x is the number this practice rests on.
A second simplification from the same recipe is deliberately not filed: fixed teacher EMA and weight-decay coefficients in place of ramp-up schedules, reported as having “minimal impact”. A null result about two hyperparameters in one recipe is not a recommendation.
Conditions
- Video pretraining, one model family, one lab. The 8x is ViT-g on A100s under V-JEPA 2’s recipe
- Needs a schedule with a distinct decay phase — see above
- Untested on text. The obvious analogue is sequence length, where the
record already recommends long-context extension as a separate later stage
(SOTA-151). Whether that is the same rule or a different one is open, and
is the reason this is
Proposed