Perception–Physics Paradox in Scientific Representation

• Perception–Physics Paradox is a phenomenon where perceptually plausible models or simulations misrepresent true physical states by collapsing distinct states into similar latent codes.
• Empirical evidence from vision models, virtual reality experiments, and relativistic thought experiments demonstrates significant divergence between perceived realism and physical accuracy.
• Methodologies like TC-Bench and structural isomorphism diagnostics provide actionable strategies to realign perceptual representations with underlying invariant physical laws.

Searching arXiv for the specified papers and closely related uses of the term to ground the article in cited sources. A Perception-Physics Paradox is a mismatch between what appears perceptually plausible or successful and what is physically or structurally correct. Across several distinct research literatures, the term denotes cases in which observers or models produce judgments, predictions, or narratives that seem valid within a given perceptual frame while failing to preserve the underlying physical relations that determine the actual outcome. In remote sensing and representation learning, the paradox names the possibility that a vision model may generalize perceptually while collapsing physically distinct states (Yao et al., 23 May 2026). In virtual reality, it refers to the finding that users often judge physically incorrect “movie physics” as more realistic than scale-correct dynamics after avatar resizing (Pouke et al., 2021). In relativistic thought experiments, closely related paradoxes arise when different co-moving frames tell different stories about the same event while agreeing on the invariant physical outcome, as in the rod-and-slot problem and Bell’s spaceships (Iyer et al., 2008, Lewis et al., 2017). A broader philosophical usage associates the paradox with limits of causal representation and with “off-site” phenomena that exceed an observer’s continuum (Novikov-Borodin, 2009).

1. Conceptual scope and principal formulations

The most explicit modern formulation appears in work on vision foundation models for tropical-cyclone analysis. There, the Perception–Physics Paradox is defined as a situation in which a representation “may generalize perceptually” while “collapsing” the physical degrees of freedom required for scientific reasoning (Yao et al., 23 May 2026). The central example is that two intense cyclones with minimum central pressures of $905$ hPa and $915$ hPa may have nearly identical infrared appearance and therefore be mapped to nearly the same latent vector, even though they are physically distinct. Conversely, visually salient changes can drive large prediction shifts even when the true physical change is negligible (Yao et al., 23 May 2026).

In virtual reality, the same term denotes a psychophysical mismatch under avatar rescaling. When a user’s virtual avatar is rescaled by a factor $s\neq 1$, two physics models can be rendered: “True physics (‘scale-correct’)” and “Movie physics (‘unscaled’).” The paradox is that users “strongly prefer the unscaled (movie) physics for plausibility—even though it is physically wrong at the user’s altered scale—and often judge it as ‘more realistic,’ despite having experienced objectively correct scaled dynamics” (Pouke et al., 2021).

A related but older family of paradoxes in special relativity does not generally use the same name, yet fits the same structural pattern. In the rod-and-slot problem, observers in different co-moving frames disagree about the temporal ordering or geometric appearance of impacts, while the pass-or-crash outcome is invariant and governed by a simple projection criterion (Iyer et al., 2008). In Bell’s spaceship paradox, the leading and trailing ships have sharply different optical experiences—redshift, blueshift, horizon formation, and asymptotic angular size—yet the relativistic account remains self-consistent and non-contradictory (Lewis et al., 2017).

A more expansive philosophical interpretation is given in “Physics beyond Causality,” where ordinary causal representation is treated as a closed logical system and paradoxes in quantum physics and cosmology are attributed to “off-site” phenomena lying beyond the observer’s continuum (Novikov-Borodin, 2009). This usage is ontological rather than empirical, but it likewise centers on a discrepancy between available representation and underlying physical reality.

2. Scientific alignment and structural isomorphism

The representation-learning literature frames the paradox through “scientific alignment,” introduced as an implicit objective for learning in scientific domains (Yao et al., 23 May 2026). Because alignment in full generality is difficult to formalize, the paper focuses on “structural isomorphism,” defined for a physical system $\mathcal S$ with true state $\mathbf y\in\mathbb R^m$, high-dimensional observation $\mathbf x=\phi(\mathbf y)$, and encoder $g:\mathcal X\to\mathcal Z\subset\mathbb R^k$, yielding latent $\mathbf z=g(\mathbf x)$. The encoder is structurally isomorphic to $\mathcal S$ if there exists an injective linear map $\mathbf A\in\mathbb R^{k\times m}$ such that, for all regimes $915$0 in a family $915$1,

$915$2

with the residual uniformly bounded in magnitude and Jacobian: $915$3 This definition formalizes the requirement that the latent space identify the true physical state “up to a global, distributed linear reparameterization” and, in particular, not collapse physically distinct states into a common latent code (Yao et al., 23 May 2026).

From structural isomorphism the paper derives a hierarchy of necessary conditions, presented as empirically testable probes. Static Fidelity requires the existence of a fixed linear decoder $915$4 such that

$915$5

Dynamic Coherence requires the same decoder to reconstruct time derivatives from latent differences: $915$6 Manifold Consistency requires known physical invariants $915$7 to remain approximately satisfied under reconstruction: $915$8 The resulting probes, denoted $915$9, $s\neq 1$0, and $s\neq 1$1, are described as a “falsifiable diagnostic hierarchy” for scientific alignment (Yao et al., 23 May 2026). Passing them is necessary, though not sufficient, for the latent representation to preserve physical structure.

This formulation shifts the paradox away from mere predictive performance. A model may succeed on a downstream task, including under standard out-of-distribution splits, while still failing to support the “rich inferential and causal queries scientists care about” because the latent geometry is not organized by the underlying state variables (Yao et al., 23 May 2026). The paradox therefore concerns representational adequacy, not just error minimization.

3. TC-Bench and empirical evidence in tropical-cyclone modeling

The empirical program built around this formulation is operationalized through TC-Bench, described as “the first versioned, global, reproducible benchmark for tropical cyclone satellite data” (Yao et al., 23 May 2026). The benchmark merges IBTrACS v4r01, which supplies multi-agency best-track metadata such as minimum central pressure $s\neq 1$2, maximum sustained wind $s\neq 1$3, and geolocation at 3-hour intervals from 1980–2024, with GridSat-B1 geostationary infrared brightness temperature imagery from which $s\neq 1$4 pixel crops, approximately $s\neq 1$5 km square, are extracted around each storm center (Yao et al., 23 May 2026). The automated pipeline includes quality control, missing-value imputation by fill to $s\neq 1$6 K, and agency-wise cleanups, yielding roughly $s\neq 1$7 individual storms after corrupted frames are discarded (Yao et al., 23 May 2026).

The benchmark explicitly separates “Moderate” and “Intense” regimes using $s\neq 1$8 hPa and $s\neq 1$9 hPa, respectively, in order to stress cases where visual cues saturate while physical variation remains substantial (Yao et al., 23 May 2026). This stratification is central to the paradox: the model’s visual competence can remain high precisely where its physical discrimination fails.

The study probes frozen VFMs including DINOv2/3, CLIP, SigLIP/2, MAE, VideoMAE, and V-JEPA2 by extracting global CLS-token representations $\mathcal S$0, with parallel checks on spatial-mean pooling, and fitting simple linear probes $\mathcal S$1 to predict $\mathcal S$2 (Yao et al., 23 May 2026). Evaluation uses three alignment-motivated diagnostics: Static Fidelity error $\mathcal S$3; Dynamic Coherence gap $\mathcal S$4; and Manifold Consistency violation $\mathcal S$5, which compares predicted latitude-conditioned wind–pressure differences to known physical trends (Yao et al., 23 May 2026).

The reported findings are consistent across model families and control settings. In the Moderate regime, static fidelity errors are “well below unity (mean $\mathcal S$6),” dynamic gaps remain small and stable, and manifold consistency violations stay around $\mathcal S$7 (Yao et al., 23 May 2026). Once $\mathcal S$8 falls below the $\mathcal S$9 hPa threshold, the situation changes sharply: median static errors rise by $\mathbf y\in\mathbb R^m$0–$\mathbf y\in\mathbb R^m$1, the fraction of catastrophic outliers with $\mathbf y\in\mathbb R^m$2 increases, one model’s median $\mathbf y\in\mathbb R^m$3 increases from $\mathbf y\in\mathbb R^m$4 to $\mathbf y\in\mathbb R^m$5, and error variance roughly doubles (Yao et al., 23 May 2026). Dynamic coherence “spikes for $\mathbf y\in\mathbb R^m$6 hPa,” while manifold consistency worsens from about $\mathbf y\in\mathbb R^m$7 to about $\mathbf y\in\mathbb R^m$8, implying that predicted wind–pressure separation under latitude stratification is more than halved relative to the physical gap (Yao et al., 23 May 2026).

A final geometric analysis on DINOv3 features shows a more explicit latent collapse in the Intense regime. The first principal component’s sensitivity to $\mathbf y\in\mathbb R^m$9 shrinks to near zero; effective dimensionality, measured by the participation ratio of the covariance spectrum, falls by roughly $\mathbf x=\phi(\mathbf y)$0; and average pairwise distance among features in a single pressure bin shrinks by a similar factor (Yao et al., 23 May 2026). The paper interprets these diagnostics as evidence that visually extreme but physically distinct storms are mapped into a lower-rank latent subspace, eliminating variation needed for scientific inference.

4. Virtual reality, rescaling, and plausibility

The virtual-reality literature studies a perceptual counterpart of the paradox in interactive simulation. Here the key condition is abnormal scale: the user is either shrunk by a factor of $\mathbf x=\phi(\mathbf y)$1 or enlarged by a factor of $\mathbf x=\phi(\mathbf y)$2, while interacting with objects under either scale-correct or unscaled dynamics (Pouke et al., 2021). The physical basis is dynamic similarity. Under scaling factor $\mathbf x=\phi(\mathbf y)$3, the paper gives

$\mathbf x=\phi(\mathbf y)$4

It further notes that if gravity is scaled, then the characteristic time satisfies

$\mathbf x=\phi(\mathbf y)$5

whereas if gravity is held constant in a physics engine, one observes

$\mathbf x=\phi(\mathbf y)$6

Corresponding velocity relations are also given, including $\mathbf x=\phi(\mathbf y)$7 when $\mathbf x=\phi(\mathbf y)$8 is constant (Pouke et al., 2021).

Two simulation conditions are contrasted. In the True Physics condition, rigid-body dynamics obey the equations of motion at the altered scale, implemented through UE4’s “WorldToMeters” so that gravity, mass, inertia, and length units rescale automatically (Pouke et al., 2021). In the Movie Physics condition, the scene is visually rescaled but physics units remain at default human-scale values, so objects “fell and flew exactly as they would if the user were unscaled ($\mathbf x=\phi(\mathbf y)$9)” (Pouke et al., 2021).

The experiments are within-subjects and counterbalanced. Study 1, the scaled-down case, began with $g:\mathcal X\to\mathcal Z\subset\mathbb R^k$0 and analyzed $g:\mathcal X\to\mathcal Z\subset\mathbb R^k$1 after two exclusions; participants, aged $g:\mathcal X\to\mathcal Z\subset\mathbb R^k$2–$g:\mathcal X\to\mathcal Z\subset\mathbb R^k$3 with $g:\mathcal X\to\mathcal Z\subset\mathbb R^k$4 and $g:\mathcal X\to\mathcal Z\subset\mathbb R^k$5, dropped and threw five soda-can pull-tabs of approximately $g:\mathcal X\to\mathcal Z\subset\mathbb R^k$6 cm by $g:\mathcal X\to\mathcal Z\subset\mathbb R^k$7 cm and mass $g:\mathcal X\to\mathcal Z\subset\mathbb R^k$8 g (Pouke et al., 2021). With users shrunk to one tenth normal height, visually perceived drop heights were about $g:\mathcal X\to\mathcal Z\subset\mathbb R^k$9–$\mathbf z=g(\mathbf x)$0 cm, implying a true-physics fall time of about $\mathbf z=g(\mathbf x)$1 s versus a movie-physics fall time of about $\mathbf z=g(\mathbf x)$2 s (Pouke et al., 2021). Study 2, the scaled-up case, began with $\mathbf z=g(\mathbf x)$3 and analyzed $\mathbf z=g(\mathbf x)$4 after three exclusions; participants, aged $\mathbf z=g(\mathbf x)$5–$\mathbf z=g(\mathbf x)$6 with $\mathbf z=g(\mathbf x)$7 and $\mathbf z=g(\mathbf x)$8, dropped and threw pine logs of $\mathbf z=g(\mathbf x)$9 m by $\mathcal S$0 m and mass approximately $\mathcal S$1 kg (Pouke et al., 2021). At $\mathcal S$2 enlargement, true-physics fall time from about $\mathcal S$3 m was about $\mathcal S$4 s, while movie physics yielded about $\mathcal S$5 s (Pouke et al., 2021).

The forced-choice results exhibit the paradox directly. In the scaled-down study, $\mathcal S$6 chose movie physics as more realistic, with binomial two-tailed $\mathcal S$7, and $\mathcal S$8 chose it as matching expectations, with $\mathcal S$9 (Pouke et al., 2021). In the scaled-up study, $\mathbf A\in\mathbb R^{k\times m}$0 chose movie physics as more realistic, with $\mathbf A\in\mathbb R^{k\times m}$1, while $\mathbf A\in\mathbb R^{k\times m}$2 chose it as matching expectations, with $\mathbf A\in\mathbb R^{k\times m}$3, not significantly different from chance (Pouke et al., 2021). Likert-scale analyses further show that ratings of fall speed, throw distance, bounciness, and gravity were generally closer to the “correct” point for movie physics, although enlarged-scale interaction displayed some asymmetry: the true-physics condition more strongly supported the sense of being “giant” and made the logs feel heavier (Pouke et al., 2021).

The paper’s explanation is that human expectations are calibrated to “normal-scale” Newtonian physics. Very rapid falls and short throw arcs at miniature scale, or slow-motion trajectories at giant scale, feel implausible even when they are physically appropriate (Pouke et al., 2021). This produces a systematic dissociation between physical correctness and judged realism. A plausible implication is that, in this context, perceptual realism is anchored to habitual sensorimotor priors rather than to scale-invariant physical law.

5. Relativistic analogues: invariant outcomes and frame-dependent narratives

Special-relativistic thought experiments provide a distinct but structurally related class of perception–physics discrepancies. In “Differing perceptions on the landing of the rod into the slot,” the rod-and-slot problem is analyzed for a rod of proper length $\mathbf A\in\mathbb R^{k\times m}$4 making acute angle $\mathbf A\in\mathbb R^{k\times m}$5 with the line of motion $\mathbf A\in\mathbb R^{k\times m}$6 in its rest frame, and a slot of proper length $\mathbf A\in\mathbb R^{k\times m}$7 making acute angle $\mathbf A\in\mathbb R^{k\times m}$8 with the same line in its rest frame (Iyer et al., 2008). In the Galilean limit, the criterion for just passing is derived from the law of sines: $\mathbf A\in\mathbb R^{k\times m}$9 The rod passes if and only if

$915$00

Under Lorentz transformation, lengths parallel to motion contract and angles skew, but the quantity “length $915$01 sin angle to motion” is invariant, since in the slot frame $915$02 and $915$03, so that

$915$04

The same numerical test therefore governs both Galilean and relativistic kinematics (Iyer et al., 2008).

What differs between frames is the explanation. In the slot frame, when the rod does not pass, the leading edge hits the front edge of the slot at event $915$05, and the trailing edge hits the back edge at a later event $915$06 (Iyer et al., 2008). In the rod frame, Lorentz transformation makes the two collisions simultaneous, $915$07, while the slot appears contracted and more steeply tilted (Iyer et al., 2008). The paper’s conclusion is that the pass-or-crash event as a “wholesome event” is unaffected by relativistic kinematics, and that the paradox lies only in differing frame-dependent accounts of why the event occurs (Iyer et al., 2008).

Bell’s spaceship paradox exhibits a similar structure. Two ships begin at $915$08 and $915$09, accelerate simultaneously in an initial inertial frame, and each experiences the same constant proper acceleration $915$10 (Lewis et al., 2017). Their worldlines are

$915$11

$915$12

Observed photon shifts satisfy

$915$13

and the resulting frequency-shift formulae differ sharply for photons exchanged before and after acceleration onset (Lewis et al., 2017). The leading ship eventually loses sight of the trailing ship behind a Rindler horizon, with horizon time

$915$14

while radar-ranging attempts ultimately fail as the round-trip proper time diverges near the horizon (Lewis et al., 2017). Yet the trailing ship’s observed redshift, proper separation, and angular size asymptote to their original rest-frame values: $915$15 Thus the ships do not share the same perceptual story, but there is no inconsistency in the causal or geometric structure (Lewis et al., 2017).

These relativistic cases are not presented under the same terminology as the machine-learning or VR literature. Nonetheless, they exemplify the same general pattern: frame-dependent appearances and descriptions can diverge without changing the invariant physical outcome.

6. Philosophical extension beyond causality

A broader usage appears in “Physics beyond Causality,” where the observer’s causal representation is treated as a closed group or monoid of perceived events $915$16, with closure, associativity, identity, and, when time reversibility holds, inverses (Novikov-Borodin, 2009). In relativity, the continuum $915$17 is acted on by inertial-frame transformations $915$18, Galilean in Newtonian mechanics and Lorentz in special relativity, with

$915$19

The paper argues that causality and geometry form a logically closed system, so that no cause–effect chain can “jump” outside $915$20 (Novikov-Borodin, 2009).

Within that framework, “off-site” phenomena are defined by another continuum $915$21, together with mutual action regions $915$22 and $915$23 connected by a smooth mapping $915$24 that does not extend globally (Novikov-Borodin, 2009). Phenomena outside these action regions are off-site relative to the observer. The paper uses this construction to reinterpret wave–particle duality, entanglement, dark matter, and dark energy. For example, the initiated fields of off-site matter are required to vanish outside $915$25, and standing-wave solutions of

$915$26

with edge-zero conditions lead to quantization conditions $915$27 (Novikov-Borodin, 2009). Entanglement is treated as correlation between off-site subsystems sharing a joint action region. Dark sectors are represented through an additional term $915$28 in the Einstein equations

$915$29

with negative pressure in a “spacelike” region identified as the hallmark of dark energy (Novikov-Borodin, 2009).

The same paper introduces a hierarchy of “levels of cognition” indexed by infinite cardinalities,

$915$30

with $915$31 for discrete events, $915$32 for continuum spacetime, $915$33 for the function space on $915$34, and higher levels for increasingly off-site continua (Novikov-Borodin, 2009). The stated consequence is that relativity, framed on a manifold $915$35, occupies a different cognitive level from quantum mechanics, framed in Hilbert spaces of functions on $915$36, and therefore their strict unification within one closed causal system is “logically doubtful” (Novikov-Borodin, 2009).

This interpretation is substantially more speculative than the benchmarked ML and VR formulations. A plausible implication is that it should be read as a philosophical generalization of representation limits rather than as an experimentally established theory of the same status as the other cited works.

7. Recurring themes, misconceptions, and significance

Across these literatures, the recurring structure is not simply “perception versus reality” in a colloquial sense. The deeper issue is a mismatch between an operative representation and the invariants that govern physical or scientific validity. In the scientific-alignment setting, predictive success on standard benchmarks can be a poor proxy for whether a latent representation preserves the structure required for causal, dynamical, or counterfactual reasoning (Yao et al., 23 May 2026). In VR, subjective plausibility can favor physically wrong dynamics because the observer’s prior is keyed to familiar human-scale motion rather than to scaling laws (Pouke et al., 2021). In relativity, different observers can describe collisions, distances, and optical appearance differently while remaining in full agreement about invariant events and causal consistency (Iyer et al., 2008, Lewis et al., 2017).

A common misconception is that such paradoxes imply contradiction in the physical theory itself. The rod-and-slot and Bell’s-spaceship analyses explicitly reject that interpretation: the paradox arises from differing descriptions, not from disagreement about the actual event (Iyer et al., 2008, Lewis et al., 2017). A second misconception is that robust out-of-distribution performance guarantees scientific usefulness. The TC-Bench results are presented precisely to show that standard perceptual or OOD benchmarks may fail to detect latent collapse in intense physical regimes (Yao et al., 23 May 2026). A third misconception is that perceived realism is identical to physical realism. The VR studies show the opposite under extreme scale manipulations, where “movie physics” is frequently judged more realistic than true physics (Pouke et al., 2021).

The significance of the paradox, in its modern empirical form, is methodological. It motivates evaluation criteria that target structural fidelity rather than surface accuracy alone. The proposed remedies in the TC-Bench work include “alignment-aware pretraining objectives,” “contrastive or dictionary-learning methods that target distribution shifts along known physical axes,” “integration of sparse in-situ measurements,” and “domain-specific architectural inductive biases” such as equivariance to rotations or radial flows (Yao et al., 23 May 2026). In VR, the findings motivate design choices that may deliberately sacrifice physical scaling to maximize plausibility, particularly when users are shrunk; for enlarged-scale experiences, the paper suggests that a mixed approach may be appropriate depending on task demands (Pouke et al., 2021).

Taken together, these studies suggest that the Perception–Physics Paradox is best understood as a family resemblance concept spanning several domains. In each case, perceptual adequacy, phenomenological plausibility, or frame-local explanation can diverge from the structures that determine physical truth. The central research problem is therefore not merely to improve prediction or appearance, but to identify and preserve the invariants on which explanation, intervention, and scientific trust depend.