The Observer Effect in World Models: Invasive Adaptation Corrupts Latent Physics

TL;DR

The paper tackles whether neural world models truly encode universal physical laws or merely memorize correlations, a problem that worsens under distribution shifts. It introduces PhyIP, a non-invasive evaluation using frozen SSL representations and a time-invariant linear readout, coupled with symbolic regression to extract interpretable physical laws. The authors derive an error bound linking SSL prediction error ε and functional curvature KΦ, and demonstrate successful recovery of conserved quantities such as internal energy and the inverse-square law in fluid dynamics and orbital-like tasks, respectively; they also show that invasive adaptation can corrupt latent physics and mislead evaluations. The work argues for fixed measurement instruments in Scientific AI, showing that non-invasive probes reveal latent physics that invasive methods can erase, and outlines future directions toward subspace-preserving adaptation and more robust evaluation protocols.

Abstract

Determining whether neural models internalize physical laws as world models, rather than exploiting statistical shortcuts, remains challenging, especially under out-of-distribution (OOD) shifts. Standard evaluations often test latent capability via downstream adaptation (e.g., fine-tuning or high-capacity probes), but such interventions can change the representations being measured and thus confound what was learned during self-supervised learning (SSL). We propose a non-invasive evaluation protocol, PhyIP. We test whether physical quantities are linearly decodable from frozen representations, motivated by the linear representation hypothesis. Across fluid dynamics and orbital mechanics, we find that when SSL achieves low error, latent structure becomes linearly accessible. PhyIP recovers internal energy and Newtonian inverse-square scaling on OOD tests (e.g., $ρ> 0.90$). In contrast, adaptation-based evaluations can collapse this structure ($ρ\approx 0.05$). These findings suggest that adaptation-based evaluation can obscure latent structures and that low-capacity probes offer a more accurate evaluation of physical world models.

Figures (14)

• Figure 1: Overview of PhyIP framework.(1) Feature Extraction: We extract frozen activations, $h(t)$ from a SSL model. (2) Linear Probing: A linear probe is trained to predict a new physical quantity, $\hat{s}(t+1) = Wh(t) + b$. Its success on OOD tests indicates that $\hat{s}$ is linearly encoded. (3) Equation Discovery & Validation: SR translates the probe's function into a symbolic formula, $\hat{\Phi}_{SR}$, and tests it on OOD data to confirm its physical plausibility.
• Figure 2: Probing Latent Physical Laws. (Top) The Non-Invasive Probe successfully extracts Total Internal Energy from frozen SSL representations in TRL-2D and RSG-3D (linear alignment), but correctly identifies representational collapse in the SN-3D experiment. (Bottom) Zero-shot generalization. Discovered symbolic formulas (Mean SR) accurately predict energy dynamics on unseen simulation parameters for the successful models
• Figure 3: The Failure of Invasive Probing.Top: The orderly geometry of the SSL model (center) is destroyed by invasive fine-tuning (right), dropping correlation from $\rho=0.94 \to -0.03$. Bottom: Quantitative Impact. This geometric destruction causes the invasive probe (Green) to fail erratically on OOD task, whereas our non-invasive probe (Blue) remains robust ($\rho > 0.8$).
• Figure 4: Probe Analysis per Block Layer-wise performance of Non-Invasive (PhyIP, Blue) vs. Non-Linear (MLP, Green) probes. While early layers MLP perform better than linear, the representation spontaneously linearizes in deep layers, peaking at Block 9 ($\rho \approx 0.91$).
• Figure 5: Mechanistic Origins of Erasure.Top: Heatmap of linear decodability change ($\Delta \rho$). Fine-tuning selectively erases dynamic variables (Speed, Radius) while preserving static one (Mass). Bottom: This collapse is driven by a parameter change causing a drop in representational similarity (CKA).