Reading physics from an AI model does not mean being able to control it

A model’s internal representation can reveal a physical quantity without providing a handle for controlling it. That is the central distinction in Manvendra Modgil’s preprint, submitted to arXiv on November 3, 2026. Across three pre-registered studies of simulated fluid flow, the paper examines whether apparently meaningful internal features reflect prediction of physics, choices about training, or deliberate architectural design.

The experiments concern a narrow setting: decaying, two-dimensional Navier–Stokes flow on a periodic 64-by-64 grid, using one simulator, one initial-condition distribution and a small convolutional encoder family. The paper does not claim that its results extend to three-dimensional flows or larger models.

Three different meanings of learning a quantity

Modgil separates three properties. Decodability means a linear probe can read a quantity from the model’s internal representation, or latent. Prominence means that quantity stands out in directions found without labels. Steerability means moving along a latent direction changes the intended property in the decoded field while satisfying checks on unwanted changes and realism.

These distinctions matter because an untrained network can already expose useful field statistics. A successful readout alone therefore does not establish that training produced a representation specifically organized around physics.

Each study fixed its hypotheses and decision rules before opening its held-out test data. The paper distinguishes confirmatory results from exploratory development and post hoc analysis. This was a self-managed pre-registration protocol, not a registered report reviewed before the experiments.

Predicting real physics did not pass the test

Study 1 examined a joint-embedding predictive architecture, or JEPA, trained to predict future flow in latent space rather than reconstruct pixels.

It improved on an untrained network for three probe targets, but failed all three pre-registered criteria. In particular, it missed a mandatory test concerning the fraction of the field dominated by vorticity, as well as tests of viscosity readout and enstrophy prominence. Enstrophy measures the field’s mean squared vorticity.

The controls complicated a physical interpretation. A JEPA trained on artificial linear dynamics read several fluid quantities better than the real-flow JEPA. A static model trained for invariance under spatial transformations, without seeing time, outperformed it on all four structure targets.

Those comparisons have caveats: the artificial dynamics changed more slowly, making prediction easier, and the static model received augmentations the JEPA did not. Nevertheless, the pre-registered hypothesis was not supported—not a finding that the JEPA contained no fluid information.

Information can become less prominent but easier to read

Study 2 examined autoencoders trained to reconstruct the present field. Plain autoencoders made enstrophy prominent in their slowest unsupervised latent directions, showing that predicting the future was unnecessary for this effect.

On sealed data with fresh training seeds, the mean prominence correlations were approximately 0.86 for velocity reconstruction and 0.80 for vorticity reconstruction, compared with 0.38 for the untrained control.

Adding VICReg regularization lowered prominence in every reported plain-versus-regularized comparison, while improving the autoencoders’ linear readout of enstrophy and palinstrophy. The drop met the pre-registered margin for velocity reconstruction but not vorticity reconstruction, producing a partially supported verdict.

The mechanism remains qualified. VICReg also made autoencoder reconstruction roughly 25 times worse. The experiment cannot separate the effect of worse reconstruction from changes in latent geometry.

Building controls instead of assuming they exist

Exploratory tests made the readout–control gap concrete: a supervised direction predicted spectral shape with correlations of 0.97–0.99, yet steering along it barely changed the rendered field.

Study 3 therefore engineered information routes. A content branch received amplitude-normalized inputs with randomly distorted spectral shape. A separate, eight-dimensional code conditioned the decoder, providing the only route for overall amplitude and an alternative route for shape.

Label-free rules then sought an amplitude, or “loudness,” dial and a spectral-shape dial. Passing required more than a measurable change: steering had to be monotonic, recoverable by re-encoding, sufficiently frame-specific, and compatible with spectral and grid-artifact checks on both sealed datasets.

Loudness dials passed in all eight trained models, including three controls without spectral distortion. Shape dials passed in three of five fresh main-model seeds, below the required four. Neither control configuration produced a valid shape dial. The verdict was again partially supported.

The results also depend on processing details: decoded fields had their spatial mean removed before scoring. One passing model produced substantial mean offsets under some steering conditions; its sealed-data verdict without that correction was not computed.

Three successes in five runs leave a wide exact 95% success-rate interval of 15–95%. Reconstruction quality also suffered. The paper’s contribution is therefore a controlled demonstration of distinctions—not a reliable general-purpose control system. Readable information, prominent structure and usable intervention must be tested separately.