Realism: Theory, Evidence, and Applications

• Realism is a family of positions that claims scientific theories truthfully describe a mind-independent reality by combining metaphysical, semantic, and epistemic theses.
• It addresses challenges like underdetermination in quantum mechanics by comparing ontological frameworks and using operational tools such as irreality measures and context-dependent evaluations.
• Applied across disciplines, realism informs methodologies in generative modeling, quantum measurement, and media authentication through quantitative metrics and rigorous hypothesis testing.

Realism denotes a family of positions about the relation between theory, observation, and a mind-independent world. In philosophy of science, it is the claim that successful theories are answerable to what exists independently of observers; in quantum foundations, it concerns whether states, properties, correlations, or contexts describe what is real; in machine learning and computer vision, it is often operationalized as whether an observation is a plausible sample from a target data-generating process or a physically unstaged scene (Arroyo et al., 2020, Theis, 2024). Across these uses, the central questions are stable: what counts as an element of reality, how objectivity is secured, and whether realism can be measured rather than merely asserted.

1. Conceptual range and basic distinctions

In its standard philosophical formulation, scientific realism holds that the best scientific theories are approximately true and that one is justified in committing to the existence of the unobservable entities they posit (Arroyo et al., 2020). The same literature separates ontological questions—what exists according to a theory—from metaphysical questions—what those entities and laws are like. A related decomposition treats realism as the conjunction of three elements: a metaphysical thesis that the world exists mind-independently, a semantic thesis that theoretical claims have truth values and should be taken at face value, and an epistemic thesis that science can identify those truth values (Allzén, 19 Nov 2025).

This broad meaning diverges from several nearby notions. In generative modeling, realism is not identical to fidelity, which measures similarity to a specific target $x^\*$, nor to weak typicality, which can hold for many unrealistic sequences, nor to heuristic notions of plausibility or naturalness. Instead, realism is formalized as whether the null hypothesis $x \sim P$ explains an observation better than alternative computable processes $Q$ (Theis, 2024). A similar shift from metaphysical to operational language appears in applied media systems, where “realism” or “credibility” can mean consistency with multisensory evidence of a physical, unstaged scene rather than mere provenance metadata (Radharapu et al., 2024).

The term also supports domain-specific research programs. In political realism, political power is modeled as a stock-and-flow quantity over a network of agents, with dynamics given by

$$\mathbf{s}(t+1) = (\mathbf{T}\circ \mathbf{M})\,\mathbf{s}(t),$$

where tactics distribute constructive or destructive power across the network (Poulshock, 2019). In cognitive neuroscience, “radical realism” names a pragmatist realism constrained by mind-independent “unamendability,” and is explicitly opposed to substance ontology, neurocentric essentialism, and self-vindicating laboratory styles (Hinrichs et al., 2024). These variations do not erase the common core; they indicate that realism is a structured family of doctrines rather than a single thesis.

2. Scientific realism, evidence, and underdetermination

A recurrent difficulty for realism is underdetermination. In non-relativistic quantum mechanics, the empirical success of the formalism does not by itself select a unique ontology, because Bohmian mechanics, GRW/objective-collapse theories, and Everettian relative-state theories can all recover the standard Born-rule statistics in accessible regimes (Arroyo et al., 2020). The measurement problem sharpens this difficulty: the completeness of the wave function, universal linear Schrödinger evolution, and unique definite outcomes cannot all be retained together in ordinary superposition scenarios. Realist commitment therefore becomes conditional on an interpretation, and a further metaphysical choice is still required about laws, individuality, structure, or the status of the wave function.

One proposed response is methodological rather than doctrinal. The meta-Popperian method fixes a theory’s ontology and dynamics, articulates candidate metaphysical profiles, and progressively eliminates those incompatible with the theory’s formal constraints—such as Bell nonlocality, collapse stochasticity, or branching structure—without pretending that physics yields a unique metaphysics (Arroyo et al., 2020). This preserves realism in a tempered form: physics narrows metaphysical options, but does not uniquely determine them.

A more critical line argues that much of the modern realism debate no longer concerns “reality-as-it-is” at all. On this view, the distinction between an empirically adequate theory $\mathrm{EmpT}$ and an ontological interpretation $\mathrm{Int}_{ont}$ is decisive: the scientific enterprise can proceed on the basis of formalism plus empirical rules, whereas ontological narratives are not scientifically adjudicated once empirical adequacy is secured (Arroyo et al., 2022). A more radical anti-representational stance appears in “reality without realism,” which affirms the existence of quantum reality while denying that standard quantum mechanics represents, or even permits the conception of, the underlying quantum objects and processes (Plotnitsky et al., 2015).

The force of underdetermination is not confined to quantum theory. In dark-matter debates, causal-descriptive semantics is argued to require canonical empirical confirmation rich enough to fix determinate intrinsic properties. In the current low-evidence regime, coarse descriptors such as non-baryonic, electromagnetically neutral, gravitationally interactive, and collisionless do not uniquely secure reference, and may both overgenerate and undergenerate the extension of “dark matter” (Allzén, 19 Nov 2025). This suggests that realist commitment can be stronger in domains of canonical empirical confirmation than in frontier regimes where evidence is sparse and semantically thin.

3. Quantum-mechanical realisms

Quantum realism has been reformulated in several mutually incompatible ways. One influential proposal is contextual realism. On this view, standard quantum mechanics is realist within its domain of application, but only after abandoning the classical conception of reality as separable, noncontextual, and composed of self-subsistent objects with definite intrinsic properties at all times. What can be known and predicated as real is conditioned by a measurement context: a maximal set of commuting observables, a concrete apparatus–system coupling, and a Heisenberg cut that sufficiently suppresses entanglement to make outcomes well defined. Once such a context is fixed, objective probabilities are given by the Born rule,

$P(a_i)=\operatorname{Tr}(\rho P_{a_i}),$

or, for POVMs, $P(i)=\operatorname{Tr}(\rho E_i)$, and the resulting claims are intersubjectively reproducible and invariant across competent observers (Karakostas, 2012).

A different realist program is Density Matrix Realism. It holds that the universal quantum state is objective and may be impure, represented by a density matrix $\rho$ rather than necessarily by a pure wave function $\psi$. Closed-system dynamics is then

$x \sim P$0

with measurement probabilities again given by trace rules such as $x \sim P$1 or $x \sim P$2. The thesis is formulated across Bohmian, GRW, and Everettian frameworks, and in its closed-system universal form is presented as empirically equivalent to Wave Function Realism when the same primitive ontologies and measurement rules are used (Chen, 2024). Its distinctive attraction lies in the possibility of a law-like initial state, for example the Initial Projection Hypothesis $x \sim P$3.

Taking decoherence “at face value” yields yet another option: strong perspectival realism. Here the only globally objective item is a universal quantum state evolving unitarily, while determinate classical-looking facts arise only relative to a perspective defined by an observer vantage, a subsystem/environment split, and a coarse-graining. In cosmological settings, where there is no external environment and factorization becomes arbitrary, this view is pushed toward mereological anti-realism: there are no in-principle physical joints at which nature must be carved (Vassallo et al., 2021). Objectivity is correspondingly weakened to inter-perspective agreement sustained by redundant environmental records.

Potentiality realism offers a more explicitly indeterministic ontology. It treats propensities—intrinsic, objective tendencies for individual events to occur under specified causal conditions—as elements of reality alongside actual properties. In quantum contexts, the relevant quantities take the Born-rule form

$x \sim P$4

but are not identified with Kolmogorov probabilities at the single-case level. The framework argues, via a Humphreys-style result, that single-case causal propensities cannot obey the full axiom set that yields Bayes’ rule while still encoding nontrivial causal influence, even though they continue to support statistics and the law of large numbers (Santo et al., 2023).

These realist options also illuminate the status of the wave function. Realism about quantum mechanics need not require realism about $x \sim P$5 as a material field. One can instead be realist about particles, flashes, or mass density in spacetime, while treating $x \sim P$6 instrumentally, nomologically, or dispositionally. The metaphysics of $x \sim P$7 remains underdetermined by the physics, even when one adopts a realist stance toward the theory as a whole (Dorato et al., 2014).

4. Operational and quantitative realism in quantum theory

A notable development in recent foundations work is the attempt to quantify realism operationally. In the Bilobran–Angelo framework, a property $x \sim P$8 is real in a quantum state $x \sim P$9 when an unrevealed projective measurement of $Q$0 leaves the state invariant: $Q$1 Irreality is then the entropy increase produced by this dephasing,

$Q$2

which vanishes exactly when $Q$3 is real in $Q$4 (Fucci et al., 2024).

This program has been extended to continuous variables by operationally discretizing position and momentum. In that setting, unrevealed measurements of discretized $Q$5 and $Q$6 yield irreality measures that obey an uncertainty relation. For Gaussian states,

$Q$7

and for minimum-uncertainty states,

$Q$8

The conclusion is that quantum mechanics forbids simultaneous classical realism for conjugate observables. Applied to the Caldirola–Kanai Hamiltonian, the same framework finds that position and velocity can both acquire elements of reality asymptotically, yielding a quantum counterpart of rest (1904.02490).

The operationalization can be generalized beyond quantum theory by using generalized probabilistic theories. There the criterion is entirely probabilistic: a property $Q$9 is real for a state $$\mathbf{s}(t+1) = (\mathbf{T}\circ \mathbf{M})\,\mathbf{s}(t),$$0 if every later measurement $$\mathbf{s}(t+1) = (\mathbf{T}\circ \mathbf{M})\,\mathbf{s}(t),$$1 has unchanged outcome statistics after an unrevealed measurement of $$\mathbf{s}(t+1) = (\mathbf{T}\circ \mathbf{M})\,\mathbf{s}(t),$$2,

$$\mathbf{s}(t+1) = (\mathbf{T}\circ \mathbf{M})\,\mathbf{s}(t),$$3

Two theory-independent quantifiers follow from this: a robustness measure,

$$\mathbf{s}(t+1) = (\mathbf{T}\circ \mathbf{M})\,\mathbf{s}(t),$$4

and a KL-based divergence of realism,

$$\mathbf{s}(t+1) = (\mathbf{T}\circ \mathbf{M})\,\mathbf{s}(t),$$5

This recasts realism as a property of a state–observable pair rather than as an all-or-nothing doctrine about an entire theory (Fucci et al., 2024).

5. Realism as plausibility in generative modeling and media authentication

In machine learning, realism is increasingly treated as a hypothesis-testing problem rather than as an informal perceptual label. An observation $$\mathbf{s}(t+1) = (\mathbf{T}\circ \mathbf{M})\,\mathbf{s}(t),$$6 is realistic relative to a target process $$\mathbf{s}(t+1) = (\mathbf{T}\circ \mathbf{M})\,\mathbf{s}(t),$$7 if it appears to have come about in the particular way modeled by $$\mathbf{s}(t+1) = (\mathbf{T}\circ \mathbf{M})\,\mathbf{s}(t),$$8, equivalently if it is a plausible sample from $$\mathbf{s}(t+1) = (\mathbf{T}\circ \mathbf{M})\,\mathbf{s}(t),$$9. The central objection to naive likelihood is that high density need not imply realism: the most probable sequence under a nearly fair coin, or a norm-constrained sample under an isotropic Gaussian, can still be implausible or structurally unrealistic (Theis, 2024).

The proposed remedy is the universal critic. Using Kolmogorov complexity $\mathrm{EmpT}$0 and the Solomonoff universal distribution $\mathrm{EmpT}$1, the single-sample critic is

$\mathrm{EmpT}$2

Large $\mathrm{EmpT}$3 means that a simple computable alternative explains $\mathrm{EmpT}$4 better than $\mathrm{EmpT}$5. For batches,

$\mathrm{EmpT}$6

and the additive complexity penalty vanishes as $\mathrm{EmpT}$7, recovering $\mathrm{EmpT}$8. The universal critic is presented as an idealized realism measure: unlike adversarial critics, it does not depend on the specific generator under evaluation and does not require adversarial training (Theis, 2024).

Because exact $\mathrm{EmpT}$9 and $\mathrm{Int}_{ont}$0 are uncomputable, the same work emphasizes practical proxies: compression-based approximations $\mathrm{Int}_{ont}$1, learned mixtures of artifact models, and score-based approximations related to classifier-free guidance in diffusion models (Theis, 2024). This suggests that realism evaluation is fundamentally comparative: it asks whether $\mathrm{Int}_{ont}$2 explains the data better than available alternatives, not merely whether $\mathrm{Int}_{ont}$3 assigns high likelihood.

A more concrete system proposal is RealSeal, which operationalizes realism at the point of capture. Here realism, or credibility, is the degree to which multisensory evidence captured at recording time is consistent with a physical, unstaged scene unfolding in three-dimensional space and time. The pipeline is Sensing $\mathrm{Int}_{ont}$4 Scoring $\mathrm{Int}_{ont}$5 Signing within a secure OS/firmware environment. Per-dimension scores are computed for 3D spatial, auditory, temporal motion, and thermal consistency, then aggregated into an overall realism score $\mathrm{Int}_{ont}$6; the bundle is hashed and signed, for example by

$\mathrm{Int}_{ont}$7

to make the realism score tamper-evident (Radharapu et al., 2024). The paper is explicit that this is a blue-sky proposal: it does not report datasets, latency, or accuracy metrics.

6. Applied realism metrics in vision and image generation

Application-specific realism metrics have become increasingly explicit. For text-to-image generation, REAL evaluates realism along three axes: fine-grained visual attributes, unusual visual relationships, and visual style. The attribute score is normalized by visible parts,

$\mathrm{Int}_{ont}$8

where $\mathrm{Int}_{ont}$9 counts visible attributes and $P(a_i)=\operatorname{Tr}(\rho P_{a_i}),$0 counts visible-and-correct ones. The style score $P(a_i)=\operatorname{Tr}(\rho P_{a_i}),$1 is a CLIP-based probability of the image being photographic rather than illustrative, and the benchmark average is

$P(a_i)=\operatorname{Tr}(\rho P_{a_i}),$2

The framework reports a Spearman’s rho score of up to $P(a_i)=\operatorname{Tr}(\rho P_{a_i}),$3 in alignment with human judgement, and shows that high-scoring images improve F1 scores of image classification by up to $P(a_i)=\operatorname{Tr}(\rho P_{a_i}),$4, while low-scoring ones degrade that by up to $P(a_i)=\operatorname{Tr}(\rho P_{a_i}),$5 (Li et al., 15 Feb 2025).

A distinct line of work measures realism by exploiting contradictions generated by large vision-LLMs. In the RealityCheck method, an LVLM extracts $P(a_i)=\operatorname{Tr}(\rho P_{a_i}),$6 atomic facts from an image, a DeBERTaV3 NLI model scores entailment and contradiction for all ordered pairs, and a contradiction-dominant weighted score is aggregated by minimum, absolute maximum, or $P(a_i)=\operatorname{Tr}(\rho P_{a_i}),$7-means clustering. The best zero-shot result on the WHOOPS! benchmark is $P(a_i)=\operatorname{Tr}(\rho P_{a_i}),$8 accuracy using the clustering aggregation with $P(a_i)=\operatorname{Tr}(\rho P_{a_i}),$9, $P(i)=\operatorname{Tr}(\rho E_i)$0, and $P(i)=\operatorname{Tr}(\rho E_i)$1 (Rykov et al., 20 Mar 2025). Here realism is semantic coherence with common-sense world knowledge rather than photorealistic texture alone.

For generated LiDAR point clouds, realism is defined through local neighborhoods that resemble the learned distribution of real-world LiDAR rather than synthetic or miscellaneous point clouds. A PointNet++-style feature extractor and classifier produce local $P(i)=\operatorname{Tr}(\rho E_i)$2 probabilities, and a dataset-level realism score is

$P(i)=\operatorname{Tr}(\rho E_i)$3

Adversarial heads suppress dataset-specific cues so that the embedding emphasizes LiDAR-specific local statistics. The reported result is not a universal correlation coefficient, but the paper confirms that the metric provides an indication for downstream segmentation performance (Triess et al., 2022).

Realism can also be perceptual rather than synthetic. The realism hypothesis for line drawings states that, for basic drawing styles, the human visual system interprets a line drawing as if it were a realistic image of a $P(i)=\operatorname{Tr}(\rho E_i)$4D scene. The paper instantiates this with headlight-plus-Lambertian shading,

$P(i)=\operatorname{Tr}(\rho E_i)$5

and argues that artists trace occluding contours and valleys of the shading field. Suggestive contours correspond to local minima of shading or, in the DeCarlo formulation, to view-dependent radial-curvature zero crossings. This explains why line drawings support recognition and shape perception despite not occurring in the natural world: they preserve the subset of image structure that a realistic rendering would make most diagnostic (Hertzmann, 2020).

Realism therefore appears in current technical practice as a family of evaluative functions tied to specific explanatory aims. In some settings it is semantic consistency with a theory-independent world; in others it is invariance under unrevealed measurement, contextually objective quantum description, or the ability of a generated artifact to withstand physically and statistically informed scrutiny. The common thread is not a single ontology, but a demand that claims, states, or samples answer to structures that are not exhausted by immediate appearance.