A ball can look identical and still deliver a different bounce. TailWorld, introduced in a preprint submitted on November 19, 2026, turns that puzzle into a test for anyone judging an AI simulation: keep the starting scene fixed, change a separate demonstration of how the objects behave, and check whether the forecast changes correctly.
The important word is correctly. A changed answer is not automatically a better answer. What matters is whether the forecast moves closer to the right outcome. That distinction matters to curious viewers, game creators and robotics teams deciding what a physical-AI demonstration actually proves.
Why the same-looking collision can end differently
Han Xiao and colleagues use controlled scenes containing a ball and a movable cylinder, with a flat, two-dimensional counterpart. They change the objects’ relative masses or how bouncy their collisions are while keeping their appearance and scene geometry unchanged. The scenery gives nothing away.
In the authors’ plotted example, a blue sphere approaches an orange cylinder. The alternative reference paths begin alike but separate after contact; in one, the cylinder moves much less. The collision reveals information that the objects’ looks cannot.
Before making a forecast, the model receives a separate demonstration under the same physical rule. It must then predict motion from different starting positions and speeds. Replaying the demonstration is therefore not enough: the useful information is how the objects interact, not the route they happened to take.
For the synthetic task, the prediction starts with half a second of collision-free motion and extends five seconds ahead. That opening motion is deliberately identical under the two competing rules. Without the separate interaction evidence, the model cannot distinguish those futures from the opening alone.
Does evidence outweigh a familiar pattern?
A world model predicts how a scene evolves. TailWorld asks what happens when the behavior demonstrated now conflicts with behavior that was much more common during training.
The training experiment keeps rare-rule sampling at 5% while varying the common-to-rare ratio from 1:1 to 4:1 to 16:1. Simple arithmetic puts the common rule’s share at 5%, 20% and 80%, respectively; other cases fill the remainder. With the training budget fixed, the rare rule is not becoming less available. The familiar alternative is becoming more dominant.
That separates two questions often blurred together: did the model lack evidence, or did it favor familiar behavior despite evidence? For the related question of what a training collection covers in the first place, see WHAM-RT’s training-data coverage and costs. TailWorld focuses on what a new demonstration can change after training.
The researchers’ proposed system, RIFT, infers a physical rule from observed motion and keeps that rule fixed while predicting the new scene. It draws on learned reference rules but separately controls the preference arising from their training frequency. The aim is not to erase previous learning; it is to stop popularity from overruling an informative collision.
What RIFT predicts better—and where it does not
In the authors’ main results table, RIFT has the lowest average position error among the listed methods: 0.0895 metres in 2D and 0.0989 metres in 3D. Those are distances averaged across objects and forecast steps, not percentages of perfect predictions.
It does not win every physical measure. GEPS, another compared method, has lower momentum error in 2D and lower kinetic-energy error in 3D. Getting an object’s path closer and getting its physical quantities closer are related achievements, but they are not interchangeable.
Nor is this a training-matched contest between finished products. RIFT and several alternatives were trained for these tasks, whereas the video models Cosmos-Predict2.5-2B and Wan2.2-TI2V-5B were evaluated without fine-tuning. The video models’ generated motion also has to be converted into object trajectories for scoring.
The demonstration-switching experiment is especially revealing. It holds the prediction’s starting state fixed and swaps the demonstrated rule. RIFT responds in the correct direction on the evaluated valid cases—but Cosmos and Wan also do so on most of theirs. A blanket claim that these video models ignore evidence would miss that result. Directional response is not an exact path. The scores cover valid evaluated cases, not an established success rate across every generated video.
There is another useful wrinkle: in the no-demonstration comparison, RIFT starts with worse position error than two compared methods, then improves when given a short, informative demonstration. Its advantage is specifically about using evidence, not being the strongest predictor regardless of what it sees.
How far does this extend beyond simulated collisions?
These are author-reported preprint results, not independently established robot capabilities. In the synthetic transfer tests, the rules and starting states were each encountered during training; their particular combinations were held out. That tests new pairings, not the discovery of unknown laws of nature.
The paper also reports results on real pendulum videos, but prediction uses extracted, geometrically calibrated positions. It does not establish unrestricted perception from raw video or reliable physical manipulation. No consumer availability or operating price for RIFT is established by the cited material.
An earlier, separate October study, PhysiCalWorld, gives the broader idea a playful setting: a simulated arm must bounce a ball off a table into a cup. Exploratory interaction helps reveal how the ball loses energy on contact. But its compact summary of that evidence slightly trails feeding the context directly into the model on one planning comparison, despite better visual scores. It is supporting context, not a replication of RIFT.
What to ask of the next AI physics demo
The practical implication is an evaluation proposal: ask to see the same prediction starting point after two informative demonstrations of different behavior, with each forecast compared against its correct outcome. Also ask whether different demonstrations of the same rule produce stable predictions. That helps distinguish learning a rule from merely reacting to a changed input.
For teams choosing between simulation approaches, choosing Cosmos for virtual robot trials covers the separate distinction between predicting an action’s visual consequences and changing a simulator’s imagery. TailWorld adds a narrower question about the evidence behind the motion.
Return to the ball and cylinder. When the demonstrated bounce changes but the starting scene does not, does the forecast change in the right way—and get closer to the right outcome? That is a more revealing request than asking for one convincing collision.
