Latent Dynamics · 2026-11-05

Readable Is Not Steerable: Pre-Registered Tests of Physical Structure in Learned Fluid Representations

Manvendra Modgil

arXiv:2611.05577PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors test whether self-supervised world models’ latents are organized around physical quantities and whether such quantities are steerable, using three pre-registered studies on 2D decaying Navier–Stokes flow. They report that what latents contain and what can be controlled come apart: a JEPA prediction objective fails prominence criteria, autoencoders can make enstrophy prominent but steering differs, and label-free “dials” yield loudness but only sometimes spectral-shape control.

Why it matters

The paper argues that reading physical structure from learned latents can be misleading without controls, separating decodability, prominence, and steerability. The authors’ pre-registered, single-open confirmation protocol on held-out fluid data provides evidence that latent representations’ “physics-like” structure is strongly shaped by objective/regularizer/decoder design and may not correspond to controllable interventions.

Method

  • Three pre-registered studies on 2D incompressible decaying Navier–Stokes flow at 64×64 resolution, evaluated once on held-out data opened a single time.
  • Measure three properties: decodability (linear ridge probe R^2), prominence (correlation between target and latent slow directions), and steerability (label-free “dials” via latent routing and decoded output changes).
  • Use controls designed to break claimed mechanisms: JEPA trained on non-physical dynamics, static model that never sees time, untrained network (same initialization), and an ablation without the key design ingredient.

Limitation

The authors state the scope is narrow and do not claim transfer to three-dimensional flow, larger models, or other data.

Abstract (from arXiv)

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 papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First announced by arXiv: 2026-11-05 · Indexed by Pub-AI: 2026-11-05 UTC

preprint

Readable Is Not Steerable: Pre-Registered Tests of Physical Structure in Learned Fluid Representations

Manvendra Modgil

First announced by arXiv: 2026-11-05 · Original record ↗

arxiv
2611.05577
Explicit source versions (1)
  • v1 · Date not supplied
  1. Preprint available

    Source date: 2026-11-05 · First detected: 2026-11-05 UTC · Historical baseline

    Source evidence (1)
    • Official evidence ↗ · RSS guid=oai:arXiv.org:2611.05577v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv

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.