← News

Models & toolsSources: arXiv · 20 Nov 2026

TailWorld benchmark shows world models favor common physics over visual evidence

Researchers behind the arXiv paper introduce TailWorld, a benchmark for visual calibration, and RIFT, a method that lets models infer rare physical rules from short video evidence.

In plain words

World models simulate how objects move and interact, but they often assume the most common physics they were trained on instead of what a short video actually shows. The paper's authors built TailWorld to measure that bias and proposed RIFT, a method that infers the governing rule from observed interactions and applies it to new starting states.

What was announced

The authors argue that training a separate world model for each environment does not scale, and that cross-environment prediction requires identifying which physical rule currently applies, something appearance alone may not reveal. Their proposed answer is visual calibration: using object motion and interactions in a short video as direct evidence to infer the rule, then predicting motion from a different initial state under that rule. The problem they identify is that uneven data availability across physical conditions creates a long-tailed training distribution, biasing models toward common rules even when the evidence supports rare ones.

TailWorld evaluates whether a demonstrated rule transfers to new initial states while independently controlling training imbalance and how informative the calibration evidence is. It scores predictions against two reference futures from the same starting state: one following the demonstrated rule and one following a common rule, which exposes bias toward common dynamics. RIFT learns how different rules affect object motion, infers the current rule from calibration interactions, and explicitly limits the influence of training frequency so informative evidence can support rare rules. In the authors' experiments on TailWorld, RIFT improved prediction accuracy and physical consistency across varying imbalances and distribution shifts, with the largest gains on rare dynamics.

What it does not show

  • The source is the paper abstract only; full experimental details, baselines and datasets are not covered, and no independent replication is reported.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.