Readable Is Not Steerable: Pre-Registered Tests of Physical Structure in Learned Fluid Representations
Abstract
Self-supervised world models are often said to organize their latents around physical quantities. We test this on 2D decaying Navier-Stokes flow in three pre-registered studies, each tested once on held-out data. In Study 1, a joint-embedding predictive architecture (JEPA) trained to predict the flow fails the pre-registered criteria for organizing its latent around fluid quantities better than an untrained network of the same architecture (not supported); a JEPA trained on non-physical linear dynamics and a static invariance model that never sees time do as well or better. In Study 2, plain autoencoders that render the field make enstrophy prominent in their slowest unsupervised latent directions (best r 0.80-0.86 vs 0.38 untrained; fresh seeds, sealed data). Adding VICReg lowers this prominence in every comparison, though not by the pre-registered margin for vorticity, while raising the autoencoders' linear decodability (partially supported). In Study 3, we build steerable directions ("dials") by routing: the decoder gets overall scale only through a small conditioning code, while its other pathway carries a degraded copy of spectral shape. Label-free rules find a loudness dial in all 8 trained models and a shape dial in 3 of 5 fresh seeds on sealed data, against a pre-registered bar of 4; neither control shows a shape dial (partially supported). Across the studies, what a latent contains and what can be controlled through it come apart. In exploration, a supervised direction reading spectral shape almost perfectly moved nothing when steered. During the build, the loudness dial had to be found by its effect on the output (closest to a pure rescaling), not by data variance; for shape, a variance rule and an effect-constrained rule each worked in 3 of 5 seeds. The program used about 70 logged hours on one consumer GPU; all pre-registrations, deviations, code and outputs are public.
Related
- From Analytic Structure to Quantum Complexity: Walsh-Pauli Representations of Continuum Operators
- Quantum field theory with scale-dependent weighted measures: nonlinear momentum representations, vacuum fluctuations, and causal structure
- Learning Stiffness Dependent Fluid Structure Dynamics from Coarse Flow Representations
- Constructing SU(N) strongly coupled dark sectors by extending the Standard Model: mixed representations, flavor structure, global symmetries and various portals
- Steerable Radiation Forces with Frequency-Detuned Acoustic Metasurfaces
BibTeX
@misc{modgil2026readable,
title = {{Readable Is Not Steerable: Pre-Registered Tests of Physical Structure in Learned Fluid Representations}},
author = {Manvendra Modgil},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.05577},
eprint = {2611.05577},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.05577},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Readable Is Not Steerable: Pre-Registered Tests of Physical Structure in Learned Fluid Representations
AU - Manvendra Modgil
PY - 2026
DA - 2026/11/03
PB - arXiv
DO - 10.48550/arxiv.2611.05577
UR - https://arxiv.org/abs/2611.05577
AB - Self-supervised world models are often said to organize their latents around physical quantities. We test this on 2D decaying Navier-Stokes flow in three pre-registered studies, each tested once on held-out data. In Study 1, a joint-embedding predictive architecture (JEPA) trained to predict the flow fails the pre-registered criteria for organizing its latent around fluid quantities better than an untrained network of the same architecture (not supported); a JEPA trained on non-physical linear dynamics and a static invariance model that never sees time do as well or better. In Study 2, plain autoencoders that render the field make enstrophy prominent in their slowest unsupervised latent directions (best r 0.80-0.86 vs 0.38 untrained; fresh seeds, sealed data). Adding VICReg lowers this prominence in every comparison, though not by the pre-registered margin for vorticity, while raising the autoencoders' linear decodability (partially supported). In Study 3, we build steerable directions ("dials") by routing: the decoder gets overall scale only through a small conditioning code, while its other pathway carries a degraded copy of spectral shape. Label-free rules find a loudness dial in all 8 trained models and a shape dial in 3 of 5 fresh seeds on sealed data, against a pre-registered bar of 4; neither control shows a shape dial (partially supported). Across the studies, what a latent contains and what can be controlled through it come apart. In exploration, a supervised direction reading spectral shape almost perfectly moved nothing when steered. During the build, the loudness dial had to be found by its effect on the output (closest to a pure rescaling), not by data variance; for shape, a variance rule and an effect-constrained rule each worked in 3 of 5 seeds. The program used about 70 logged hours on one consumer GPU; all pre-registrations, deviations, code and outputs are public.
ER -