When Do Conservation Laws Survive Learned Representations? Certified Horizons for Latent World Models
Hongbo Wang · 2026
Read When Do Conservation Laws Survive Learned Representations? Certified Horizons for Latent World Models 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-2606-24945.txt.
Description
Hongbo Wang asks when a conservation law stays certifiable after a model learns its latent code. A certified horizon bounds how many rollout steps provably stay on an invariant's level set. The certified object matters: not a learned latent Hamiltonian or scalar witness, since a model can conserve either while true energy drifts. Wang certifies the decoded invariant instead, obtained by decoding the latent state. Shell horizon certificates follow, their budgets split into representation, readout, and latent dynamics defects.
Tests across state, learned lift, and pixel observations give a verdict. Canonical symplectic structure gives the longest horizons among known phase coordinates, yet fails to cross a learned chart. A controlled Lipschitz aligned soft invariant persists in the nonlinear settings tested, two lift systems and pixels, with the gain growing with nonlinearity. Pixel certification is recovered on a readout sub tube. The Kepler problem exposes a boundary to the method.
Catalog record
| CREATOR | Hongbo Wang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.24945