A world foundation model is a generative simulator that lets a robot predict how a 3D scene will evolve before acting. A new framework, PAIWorld, argues the consistency problem across cameras is structural, not a data problem.
Imagine a robot that sees the world through three cameras. One sits on its head, one on its wrist, one points down at the work surface. When that robot imagines reaching for a cup, the head camera might picture the cup on the left shelf while the wrist camera pictures it on the right. The object is in two places at once, because the robot's internal simulator never learned that its cameras should agree on where things are.
That failure mode has a name in the world-model literature: cross-view drift. The same object renders in two positions from two viewpoints. Depth flips between views. Textures slip off geometry. A robot that plans against such a simulator is planning against a hall of mirrors, and the field has not had a clean structural fix for the bug.
A new research framework called PAIWorld argues that the gap is architectural. A world foundation model, in plain terms, is a generative simulator that predicts how a 3D scene will evolve over time from a robot's point of view, so the robot can imagine the outcome of an action before committing to it. PAIWorld's authors do not claim a larger model or more data. They claim a diagnosis: existing multi-view world models concatenate view tokens without explicit geometric reasoning, and that single choice produces the drift.
The diagnosis has two halves. First, the views never explicitly communicate. Today's multi-view world models just line up view tokens side by side and let attention sort it out. That works for some things, but it does not enforce that the same object is the same object across cameras. Second, the model has no 3D geometric prior, no built-in sense of where each camera is looking in space. Without that, the model is free to imagine consistent-looking but mutually inconsistent views.
PAIWorld proposes three components, together, as the fix. The first is a Geometry-Aware Cross-View Attention block, an attention module that lets the views explicitly exchange information before the model renders, instead of stitching tokens together blindly. The second is a Geometric Rotary Position Embedding, an encoding that bakes each camera's ray direction and extrinsic pose directly into the attention math, so the model knows where each view is looking. The third is Latent 3D-REPA, a distillation step that pulls 3D-aware features out of a frozen 3D foundation model and feeds them into the world model, anchoring it to real geometry. The backbone is a diffusion-transformer, the same family that has powered recent image and video generators, not a discrete-token or vector-quantized video model.
The framework is not framed as a benchmark story, but the authors do report results. PAIWorld placed first on WorldArena and second on the AgiBot-Challenge2026, two multi-view 3D consistency benchmarks for robotic manipulation, according to the arXiv preprint. The result is supporting evidence that the architectural changes move the needle on the metrics the field has agreed to measure, which is what benchmark wins are for.
The honest caveat is the closer. Consistency on a leaderboard is not the same as a deployed policy win. The paper claims downstream uses, including model-based planning, world action models, and multi-view policy post-training, but does not show any of them running in production. The structural diagnosis is real, and the components are concrete. Whether 3D-consistent world models translate into robots that manipulate the physical world better is a question the leaderboard cannot answer, and the authors do not claim it does. The watch item, stated plainly, is whether future releases move from benchmark consistency to a robotic arm picking up the cup it was supposed to pick up, every time, from any angle.