Exact equivariance, kept through training, buys zero-shot generalisation across the symmetry group

Hongbo Wang · 2026

Read Exact equivariance, kept through training, buys zero-shot generalisation across the symmetry group by Hongbo Wang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-03003.txt.

Introduction

Wang proves that exact equivariance in a latent world model buys zero-shot generalization across a symmetry group. The model pairs an equivariant encoder with an equivariant predictor. When dynamics carry a group acting on latents, the one-step prediction error is exactly invariant across the group. Fitting one slice of orientations thus fixes error on the whole orbit.

The symmetry survives real training with standard optimizers. The composed residual stayed near one part in a million after training. Error stayed flat across the group, while a larger non-equivariant baseline rose by factors of 12.7 in 2D and 17.2 in 3D. On real robot trajectories from DROID, the equivariant model stayed flat too. Wang stresses that flatness transports error unchanged; it does not lower it. The result grounds a wider certified-world-models program. A matching planner result extends the invariance to closed-loop control.

Catalog record

CREATORHongbo Wang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE476 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.03003

Browse the full library · About the Tower of Babel