← Selected work KV / Research 05 Pull time backward (3D) ↓ Corrected paper ↗
ICLR 2026 · 2nd Workshop on World Models · accepted · corrected July 2026

Which way
is time running?

A frozen world model sees the same simulation forward and backward. Its change in prediction loss becomes a test for the arrow of time.

Reader’s tour ↗
ARROW-OF-TIME TESTROUND 01 / FALLING OBJECTS
YOUR SCORE0 / 0
48 FRAMES · PYBULLETDIRECTION: ???

Orbit it · 3D reversal room

Pull time backward.

A re-simulated domino run you can orbit while dragging time in both directions. Forward, the motion spends its energy and goes still. Backward, a settled pile climbs back onto its edges. Hand-authored motion for intuition. It produces no loss values.

Still image · click to load the world Open on its own page ↗

Evidence receipt

Know what each number can prove.

The interactive lab separates one exact sequence from the paper's aggregate result. It also keeps the post-publication corrections visible.

01 · EXACT CLIP

Real sequence, real loss

Six direction clips come from three checked-in 48-frame PyBullet sequences. The loss values change with scene, model, and context.

02 · AGGREGATE

380 simulations

The paper evaluates 160 discrete-scene videos and 220 continuous restitution or damping videos. A clip is an example, not the claim.

03 · CORRECTED CLAIM

Two supported models

V-JEPA2 shows the expected dissipative separation for contexts 4-12. VideoMAE v1 shows the inverse pattern.

04 · NOT AVAILABLE

Missing matched artifacts stay out

The corrected randomized-domino rerun has no matching frames here, so it appears only in the aggregate evidence. Continuous-sweep videos and per-frame loss traces are also absent.

Open aggregate evidence and correction notes
SUPPORTED

V-JEPA2

Positive dissipative separation holds across contexts 4-12 in the corrected randomized-domino rerun. The aggregate sign reverses at context 14.

SUPPORTED

VideoMAE v1

The inverse pattern survives recomputation. The corrected paper fixes the checkpoint name from VideoMAE v2 to VideoMAE v1.

WITHDRAWN

MVD

The public checkpoint had no trained decoder weights. Its result cannot support a claim about distillation.

WITHDRAWN

Hiera

The evaluator used random-mask inpainting instead of symmetric future prediction. The objective comparison is invalid.

Corrected randomized-domino rerunΔ = dissipative − low-dissipation TRA
V-JEPA2 +0.1185+0.2084+0.1708+0.1207+0.1702−0.2187
VideoMAE v1 −0.1203−0.1744−0.2375−0.2868−0.3308−0.4015

Percentage points. Exact means from the corrected 160-video discrete rerun.

Corrected dataset with 40 randomized domino seeds. V-JEPA2 separates the groups at contexts 4-12 and reverses at 14; VideoMAE v1 remains inverted.
Original-dataset TRA context sweep for V-JEPA2 EMA, V-JEPA2 without EMA, and VideoMAE v1
Original-dataset aggregate after correcting the model identity: V-JEPA2 EMA, V-JEPA2 without EMA, and VideoMAE v1. This is not the randomized-domino rerun.
Method diagram comparing forward and reversed prediction losses
The same frozen prediction pipeline receives both frame orders.

TRA measures a change in prediction loss. It does not prove that a model understands physics. The core evidence is synthetic, and the real-video results remain confounded by human action timing.