Watch the complete clip.
You guess whether playback runs forward or backward. This is a perception game for intuition.
whole sequence → direction guessICLR 2026 · 2nd Workshop on World Models
Which way does
physics run?
QuestionDoes a frozen video model assign different prediction loss to a process and its time reversal?
MetricTRA compares held-out-frame loss on forward and reversed versions of the same video.
ResultV-JEPA2 responds to visible dissipation. VideoMAE v1 shows the inverse sign.
BoundaryTRA probes loss asymmetry. It does not prove that a model understands physics.
Your direction guess is not the model test.
Predict held-out frames twice.
The frozen model receives the same video in forward and reversed order. Its loss, not its label guess, defines TRA.
context frames → held-out-frame prediction lossTemporal Reversal AsymmetryTRA(v) = [ L(reverse(v)) − L(v) ] / L(v)
positive reversed order is harder
zero both directions receive similar loss
negative forward order is harder
Find the arrow before the model does.
After you answer, the page reveals the frozen model's exact loss asymmetry for this same retained clip at context 8.
ROUND 01
V-JEPA2 · context 8
forward loss
reverse loss
The displayed videos are retained examples from the original synthetic set. One clip explains a calculation; it does not represent the aggregate.
Visible energy loss peaks in the middle.
Choose only measured parameters. No value is interpolated.
Explanatory schematic · not retained footageRestitution 0.5
Earlier stateLater state
Illustrative sequence: rebound height decreases over time.
Restitution .5+0.27%
Peak visible dissipation
The ball keeps bouncing while its amplitude shrinks. Reversal makes energy gain visible across the clip.
Original continuous sweep: 220 videos. Each control point above is a measured V-JEPA2 TRA value at context 12.
The sign depends on the prediction objective.
Change context length. The zero-centered scale stays fixed, so the apparent effect cannot grow by rescaling the chart.
Dissipative videos receive greater reversal asymmetry.
Inverted at every measured context.
Near-zero control across the corrected rerun.
Three numbers describe three different things.
160 discrete-scene videos plus 220 continuous restitution or damping videos.
Forty videos per scenario, including 40 randomized domino simulations.
One displayed retained sequence per round, used to explain the loss calculation.
Corrections in context
The supported claim is narrower and cleaner.
V-JEPA2 is positive for dissipative versus low-dissipation scenes at contexts 4 to 12 in the corrected rerun.
VideoMAE v1 remains negative at every context. The corrected paper fixes the checkpoint identity.
The public MVD checkpoint lacks trained decoder weights. Hiera used a mismatched random-mask objective.
Limits
A probe of asymmetry, not proof of understanding.
- The evidence is primarily synthetic physics.
- Real human-action video contains timing confounds.
- TRA depends on the prediction objective, context length, and visible motion window.
- Very strong dissipation can stop motion early and reduce the visible signal.
- The C14 sign reversal limits the V-JEPA2 aggregate claim to contexts 4 through 12.