MetaLight: Multifunctional Optical & RL Systems

• MetaLight is a multifaceted concept that designates engineered intermediate layers—optical metasurfaces or meta-learned initializations—between input and task-specific outputs.
• In nanophotonics, MetaLight involves metasurfaces that sculpt dispersion, polarization, and strong light–matter coupling to achieve programmable optics and efficient flat optical devices.
• In transportation, MetaLight refers to a value-based meta-reinforcement learning approach that rapidly adapts traffic signal control under changing traffic distributions.

MetaLight is a term used in the recent literature in more than one technically distinct sense. In nanophotonics and metaoptics, it denotes a metasurface- or metamaterial-centered approach in which subwavelength optical structure is the primary mechanism for creating, controlling, or distributing electromagnetic fields, including strong light–matter coupling, fiber-tip wavefront conversion, programmable optical computation, and manufacturable flat optics (Chen et al., 2020, Li et al., 2023, Tanuwijaya et al., 2023, Kim et al., 2022). In transportation machine learning, MetaLight denotes a value-based meta-reinforcement learning method for traffic signal control that is designed to adapt rapidly to shifted traffic distributions from a learned initialization (Taschin et al., 18 Sep 2025). The shared feature across these usages is not a single device or algorithmic formalism, but the placement of a structured intermediate layer—optical or algorithmic—between a simple input and a task-specific output.

1. Terminological scope and conceptual uses

The photonics usage of MetaLight is centered on meta-optical control. In this sense, the metasurface is not treated as a passive boundary condition alone, but as the element that encodes mode profile, dispersion, polarization response, radiation leakage, or wavefront transformation. This framing appears explicitly in work on monolayer exciton-polaritons integrated with a silicon nitride metasurface, where the metasurface is described as the central tool for sculpting light–matter interaction at the nanoscale (Chen et al., 2020). A related usage appears in fiber-integrated structured-light generation, where a polymeric metasurface fabricated on a polarization-maintaining single-mode fiber end-face converts the emitted fiber mode directly into vectorial or vortex fields on the hybrid-order Poincaré sphere (Li et al., 2023). In programmable quantum optics, MetaLight refers to a metasurface-based optical processor in which a fixed nanostructured device, together with programmable illumination, implements selected unitary transformations for quantum information processing (Tanuwijaya et al., 2023).

A broader but compatible usage appears in manufacturable metaoptics. The scalable visible metalens work does not define MetaLight as a single named platform, but it is presented as aligned with a broader metaoptics vision in which flat optics move from laboratory demonstration toward industrial manufacturing, especially for compact display systems such as virtual reality (Kim et al., 2022). Related work on ultralight thermal metasurfaces similarly fits a MetaLight-style interpretation by combining optical functionality with exceptionally low areal mass in planar emissive structures (Naqavi et al., 2017).

The transportation usage is narrower and algorithmic. There, MetaLight is a specific value-based meta-reinforcement learning method for single-intersection traffic signal control, built on FRAP, FRAP++, DQN-style value learning with Bellman loss, and MAML-like meta-learning (Taschin et al., 18 Sep 2025). A common misconception is to treat these photonic and transportation usages as belonging to a single research lineage. The available literature instead indicates homonymy: the same label is applied to unrelated optical and reinforcement-learning systems.

2. MetaLight as metasurface-engineered light–matter interaction

A canonical photonic realization is the metasurface-integrated monolayer exciton-polariton platform based on a silicon nitride metasurface coupled to a tungsten diselenide monolayer (Chen et al., 2020). The physical structure is a SiN metasurface on a SiO$_2$/Si substrate, consisting of SiN meta-atoms arranged in a square lattice of circular holes. The reported design uses a 130 nm-thick SiN slab, a 459 nm lattice period, and 120 nm hole diameter. A hBN-encapsulated WSe$_2$ monolayer is transferred directly on top of the metasurface, enabling evanescent coupling to the near field of the guided mode resonance. The SiN lattice supports optical Bloch modes inside the light cone, so they couple to free space while remaining strongly confined in the slab (Chen et al., 2020).

The bare metasurface already supports dispersive guided resonances with Fano lineshapes in reflection. Two modes, M1 and M2, exhibit distinct dispersions, one roughly linear and one parabolic, and their anti-crossing at finite in-plane momentum is a photonic-photonic coupling effect visible in both simulation and measured angle-resolved reflection. The reflection spectrum is fit using a Fano form,

$R(\omega) = R_{\mathrm{Fano} + R_{\mathrm{FP} + R_{\mathrm{bg},$

with the resonant term described by

$R_{\mathrm{Fano} = I \left(1 - \frac{q}{x}\right),$

and

$x = \frac{E - E_0}{\Delta \omega}.$

For the bare metasurface, the extracted resonance energy is 1.726 eV and the $Q$ factor is about 143 (Chen et al., 2020).

After transfer of the hBN-capped WSe$_2$ monolayer, energy-momentum spectroscopy at 22 K reveals a clear anti-crossing between the cavity-like guided mode resonance and the exciton resonance near 715 nm, in both reflection and photoluminescence. The polariton branches are modeled with the coupled-oscillator matrix

$$\begin{pmatrix} E_x + i\gamma_x & g \ g & E_c + i\gamma_c \end{pmatrix},$$

with eigenvalues

$$E_p = \frac{E_c + E_x + i(\gamma_c + \gamma_x)}{2} \pm \sqrt{ g^2 + \left[\frac{E_c - E_x + i(\gamma_c - \gamma_x)}{2}\right]^2 }.$$

The fit yields $E_x = 1.728$ eV, $_2$0 meV, $_2$1 meV, and $_2$2 meV, corresponding to a measured Rabi splitting of 18 meV. The strong-coupling criteria are reported as

$_2$3

The measured 18 meV splitting agrees with the observed anti-crossing and follows the trend of numerical simulations, although it is smaller than a theoretical estimate of about 66 meV; this reduction is attributed to an unintended air gap caused by wrinkles or transfer residues (Chen et al., 2020).

The same platform also shows that the metasurface controls emission geometry, not only coupling strength. Diffraction by the nano-patterned SiN surface funnels lower-polariton emission into specific far-field directions. In the back focal plane, the emission is concentrated along the $_2$4 axis near $_2$5, with a measured far-field divergence of about $_2$6 from a Lorentzian fit with FWHM about $_2$7. Subwavelength engineering of thickness, duty cycle, periodicity, and lattice symmetry tunes surface field strength, cavity leakage, dispersion curvature, and far-field pattern. The normalized mode energy obeys

$_2$8

and the surface field exhibits an optimum around 100 nm thickness. By changing periodicity, the photonic dispersion can be reshaped from linear or parabolic to W-shaped; changing the lattice from square to hexagonal alters the emission from a simple diffractive peak pattern to a star-like distribution (Chen et al., 2020).

This usage of MetaLight therefore denotes a compact, planar route to exciton-polaritons in which meta-atom geometry directly sets coupling strength, mode lifetime, momentum-space dispersion, and directionality. A plausible implication is that the term here signifies a design philosophy in which the optical environment is deliberately sculpted to create hybrid quasiparticles rather than merely route classical beams.

3. Fiber-tip light shaping and programmable optical computation

In the metafiber platform, MetaLight corresponds to fiber-integrated light shaping using a polymeric metasurface printed directly on the end-face of a polarization-maintaining single-mode fiber (Li et al., 2023). The motivation is to move structured-light generation from free-space optical trains onto the fiber tip. The key innovation is a 3D laser-nanoprinted polymeric metasurface composed of anisotropic 3D nanopillar meta-atoms fabricated by two-photon polymerization. A hollow tower of about 550 $_2$9m height is first printed on the fiber end-face so the output expands to fill a metasurface aperture about 100 $R(\omega) = R_{\mathrm{Fano} + R_{\mathrm{FP} + R_{\mathrm{bg},$0m in diameter (Li et al., 2023).

The underlying state space is described on the hybrid-order Poincaré sphere. The paper writes the general state as

$R(\omega) = R_{\mathrm{Fano} + R_{\mathrm{FP} + R_{\mathrm{bg},$1

with the full field carrying different OAM charges $R(\omega) = R_{\mathrm{Fano} + R_{\mathrm{FP} + R_{\mathrm{bg},$2 and $R(\omega) = R_{\mathrm{Fano} + R_{\mathrm{FP} + R_{\mathrm{bg},$3 in the right- and left-circular components. Each anisotropic nanopillar behaves as a local waveplate, with the output under incident x-linear polarization given by

$R(\omega) = R_{\mathrm{Fano} + R_{\mathrm{FP} + R_{\mathrm{bg},$4

where $R(\omega) = R_{\mathrm{Fano} + R_{\mathrm{FP} + R_{\mathrm{bg},$5 are transmission amplitudes, $R(\omega) = R_{\mathrm{Fano} + R_{\mathrm{FP} + R_{\mathrm{bg},$6 is the phase difference between x- and y-polarized modes, $R(\omega) = R_{\mathrm{Fano} + R_{\mathrm{FP} + R_{\mathrm{bg},$7 is the in-plane rotation angle, and $R(\omega) = R_{\mathrm{Fano} + R_{\mathrm{FP} + R_{\mathrm{bg},$8 is the propagation phase. This provides simultaneous local control of polarization via $R(\omega) = R_{\mathrm{Fano} + R_{\mathrm{FP} + R_{\mathrm{bg},$9 and $R_{\mathrm{Fano} = I \left(1 - \frac{q}{x}\right),$0, and wavefront phase via $R_{\mathrm{Fano} = I \left(1 - \frac{q}{x}\right),$1 (Li et al., 2023).

The platform uses a 3D meta-atom library parameterized by length $R_{\mathrm{Fano} = I \left(1 - \frac{q}{x}\right),$2, width $R_{\mathrm{Fano} = I \left(1 - \frac{q}{x}\right),$3, height $R_{\mathrm{Fano} = I \left(1 - \frac{q}{x}\right),$4, and rotation angle $R_{\mathrm{Fano} = I \left(1 - \frac{q}{x}\right),$5, with a commercial polymer of refractive index about 1.5 at 1550 nm. This height degree of freedom expands the design space beyond that of planar metasurfaces and enables independent, complete, and precise control of phase and polarization. Demonstrated outputs include radial and azimuthal cylindrical vector beams, a right-handed circularly polarized vortex beam with topological charge $R_{\mathrm{Fano} = I \left(1 - \frac{q}{x}\right),$6, a left-handed circularly polarized vortex beam with topological charge $R_{\mathrm{Fano} = I \left(1 - \frac{q}{x}\right),$7, and an arbitrary vector state with spatially varying elliptical polarization (Li et al., 2023).

A different computational use of MetaLight appears in the programmable quantum metasurface platform (Tanuwijaya et al., 2023). There, the device is a geometric-phase metalens array containing nine metalenses on one metasurface. Each metalens is designed to map an input spatial mode into several output directions with prescribed complex amplitudes so that the collection realizes a unitary operator $R_{\mathrm{Fano} = I \left(1 - \frac{q}{x}\right),$8. For an $R_{\mathrm{Fano} = I \left(1 - \frac{q}{x}\right),$9-qubit computation, the desired transformation is

$x = \frac{E - E_0}{\Delta \omega}.$0

and in the reported experiment $x = \frac{E - E_0}{\Delta \omega}.$1, so a $x = \frac{E - E_0}{\Delta \omega}.$2 unitary is implemented. The transmission of the $x = \frac{E - E_0}{\Delta \omega}.$3-th metalens is

$x = \frac{E - E_0}{\Delta \omega}.$4

Programmability is achieved not by reconfiguring the nanostructure but by using a spatial light modulator to prepare input amplitudes and phases and to select which subset of metalenses is illuminated (Tanuwijaya et al., 2023).

The demonstrated algorithms are Grover’s search and the quantum Fourier transform. In the Grover case, the four basis inputs are encoded as

$x = \frac{E - E_0}{\Delta \omega}.$5

while for the QFT the input states are

$x = \frac{E - E_0}{\Delta \omega}.$6

The output field is the coherent sum

$x = \frac{E - E_0}{\Delta \omega}.$7

or, equivalently,

$x = \frac{E - E_0}{\Delta \omega}.$8

Using a single-photon camera in the Fourier plane, the work reports root-mean-square errors of 0.148 for Grover’s search and 0.216 for QFT across all 16 matrix elements (Tanuwijaya et al., 2023).

Taken together, these two systems illustrate a characteristic MetaLight pattern in optics: a simple input mode, whether a PM-SMF output or a classical/single-photon illumination pattern, is transformed by a fixed meta-structured surface into a target field or linear operator in a single optical stage.

4. Manufacturing, thermal metasurfaces, and evaluation frameworks

A major strand of the MetaLight-related literature concerns manufacturability and systems integration. The visible metalens manufacturing paper demonstrates a scalable route based on a 12-inch master stamp written by a 193 nm ArF immersion scanner and replicated by wafer-scale nanoimprint lithography (Kim et al., 2022). The scanner exposed 669 dies of 1 cm metalenses continuously on a 12-inch Si wafer, achieving width critical dimension of 75 nm and gap critical dimension of 40 nm. Once a 12-inch master stamp is imprinted, hundreds of centimeter-scale metalenses can be fabricated, and imprint replication takes only about 15 minutes (Kim et al., 2022).

The printed polymer nanostructures are then thinly coated with high-index TiO$x = \frac{E - E_0}{\Delta \omega}.$9 by ALD, forming an atomic layer-polymer hybrid metasurface. This coating raises conversion efficiency from about 10% for the uncoated printed metasurface to as high as 90% after coating. The reported printable meta-atom has period 450 nm, height 900 nm, TiO$Q$0 thickness 23 nm, length 380 nm, and width 70 nm. Simulated conversion efficiencies are 68.2% at 450 nm, 89.6% at 532 nm, and 79.7% at 635 nm; experimental beam-steering measurements give 60.9% at 450 nm, 77.8% at 532 nm, and 64.8% at 635 nm (Kim et al., 2022).

The fabricated metalens is 1 cm in diameter, designed for 532 nm with $Q$1 cm and $Q$2. The ideal phase profile is reported as

$Q$3

and the local rotation rule is

$Q$4

Measured diffraction efficiencies are 20.9% at 450 nm, 34.4% at 532 nm, and 23.0% at 635 nm, with Strehl ratios of 0.81, 0.90, and 0.86, respectively. The same work demonstrates a near-eye transmissive VR prototype showing virtual images in RGB colors (Kim et al., 2022).

A distinct but related direction is ultralight thermally emissive metasurfaces (Naqavi et al., 2017). These structures are based on Salisbury screen and Jaumann absorber concepts and use Cr, CP1 polyimide, and SiO$Q$5. They comprise one to three pairs of alternating nanometer-scale metallic and dielectric layers and achieve broadband infrared emission over 2 to 35 $Q$6m with effective 300 K hemispherical emissivities of 0.7 to 0.9. Reported areal masses are 3.3 g/m$Q$7 for the Salisbury screen and less than 10 g/m$Q$8 for all structures, with a free-standing membrane at about 3 g/m$Q$9. The best designs reach about 0.84–0.85 hemispherical emissivity and are described as mechanically flexible, low outgassing, and resistant to UV radiation and atomic oxygen (Naqavi et al., 2017).

For applications in light collection rather than imaging, characterization itself becomes part of the MetaLight ecosystem. A detector-oriented methodology has been proposed for a rectangular fused-silica metalens designed for 632 nm, with 1200 nm tall nanoposts, 500 nm unit cell, 1.2 mm width, 13.7 mm length, 7 mm focal length, and NA 0.75 (Contreras et al., 2023). The design is hybrid: phase-map below about $_2$0 and supercell metagrating above. Efficiency into diffraction order $_2$1 is defined as

$_2$2

The measurements show first-order efficiency above about 65% up to around $_2$3 in the metagrating region, and good agreement between experiment and simulation (Contreras et al., 2023). This suggests that within MetaLight-style flat optics, performance metrics depend strongly on application domain: imaging quality, collection efficiency, angular acceptance, and manufacturability are not interchangeable criteria.

5. Imaging correction and atomic or metamaterial extensions

The limitations of metalens imaging have motivated computational correction frameworks. MetaFormer, an aberration-correcting Vision Transformer method for metalens-captured images, models aberrated imaging as

$_2$4

and uses a non-blind pipeline based on multiple Wiener deconvolutions and stage-aware transformer attention (Lee et al., 2024). Its Multiple Adaptive Filters Guidance module uses

$_2$5

and

$_2$6

Its Spatial and Transposed self-Attention Fusion module combines encoder and decoder attention streams through

$_2$7

$_2$8

On synthetically aberrated images, the reported performance improves from 28.92 dB PSNR for Restormer to 32.16 dB for MetaFormer, with SSIM increasing from 0.7719 to 0.8159 and LPIPS decreasing from 0.3039 to 0.2810 (Lee et al., 2024). This does not redefine MetaLight as an algorithm, but it shows how metaoptics increasingly includes joint optical-computational stacks.

At a more fundamental level, atomic-emitter arrays have been proposed as a route to a metalens formed by structured arrays of identical emitters (Andreoli et al., 2024). The optical phase target is

$_2$9

while the single-layer cooperative linewidth is

$$\begin{pmatrix} E_x + i\gamma_x & g \ g & E_c + i\gamma_c \end{pmatrix},$$0

The analysis shows that one layer obeys the bound

$$\begin{pmatrix} E_x + i\gamma_x & g \ g & E_c + i\gamma_c \end{pmatrix},$$1

so arbitrary phase with high transmission is not available. Two layers remove that constraint in the lossless limit, and three layers provide robustness to losses. For an illustrative design with $$\begin{pmatrix} E_x + i\gamma_x & g \ g & E_c + i\gamma_c \end{pmatrix},$$2, $$\begin{pmatrix} E_x + i\gamma_x & g \ g & E_c + i\gamma_c \end{pmatrix},$$3, and $$\begin{pmatrix} E_x + i\gamma_x & g \ g & E_c + i\gamma_c \end{pmatrix},$$4 atoms, the reported efficiency is $$\begin{pmatrix} E_x + i\gamma_x & g \ g & E_c + i\gamma_c \end{pmatrix},$$5 for $$\begin{pmatrix} E_x + i\gamma_x & g \ g & E_c + i\gamma_c \end{pmatrix},$$6, with $$\begin{pmatrix} E_x + i\gamma_x & g \ g & E_c + i\gamma_c \end{pmatrix},$$7 (Andreoli et al., 2024). A plausible implication is that MetaLight-style flat optics is broadening from nanofabricated dielectric elements toward cooperative many-body optical materials.

Another extension appears in metamaterial waveguides with three-level $$\begin{pmatrix} E_x + i\gamma_x & g \ g & E_c + i\gamma_c \end{pmatrix},$$8 atoms in a dielectric core (Lavoie et al., 2013). Here the central quantity is slow light enabled by electromagnetically induced transparency, with susceptibility

$$\begin{pmatrix} E_x + i\gamma_x & g \ g & E_c + i\gamma_c \end{pmatrix},$$9

and group velocity

$$E_p = \frac{E_c + E_x + i(\gamma_c + \gamma_x)}{2} \pm \sqrt{ g^2 + \left[\frac{E_c - E_x + i(\gamma_c - \gamma_x)}{2}\right]^2 }.$$0

For the same amount of slowing, metamaterial cladding yields lower attenuation than metal cladding, with about 20% lower attenuation when $$E_p = \frac{E_c + E_x + i(\gamma_c + \gamma_x)}{2} \pm \sqrt{ g^2 + \left[\frac{E_c - E_x + i(\gamma_c - \gamma_x)}{2}\right]^2 }.$$1 and about 40% when $$E_p = \frac{E_c + E_x + i(\gamma_c + \gamma_x)}{2} \pm \sqrt{ g^2 + \left[\frac{E_c - E_x + i(\gamma_c - \gamma_x)}{2}\right]^2 }.$$2 (Lavoie et al., 2013). This work is not named MetaLight, but it belongs to the same broader tendency of engineered media being used to reshape optical propagation rather than merely transmit it.

6. MetaLight in traffic signal control and its evaluation under distribution shift

In transportation research, MetaLight is a value-based meta-reinforcement learning method for traffic signal control (Taschin et al., 18 Sep 2025). The setting is single-intersection control with 8 possible movements. The state $$E_p = \frac{E_c + E_x + i(\gamma_c + \gamma_x)}{2} \pm \sqrt{ g^2 + \left[\frac{E_c - E_x + i(\gamma_c - \gamma_x)}{2}\right]^2 }.$$3 is a traffic observation such as queue lengths on lanes, the action $$E_p = \frac{E_c + E_x + i(\gamma_c + \gamma_x)}{2} \pm \sqrt{ g^2 + \left[\frac{E_c - E_x + i(\gamma_c - \gamma_x)}{2}\right]^2 }.$$4 is the chosen signal phase, and the reward is tied to queue length or pressure-style objectives. The method is built on FRAP, FRAP++, DQN-style value learning, and MAML-like meta-learning, with the goal of learning a base model that can be rapidly adapted to a new traffic scenario using only a small amount of new data (Taschin et al., 18 Sep 2025).

The reinforcement-learning objective is written in terms of the return

$$E_p = \frac{E_c + E_x + i(\gamma_c + \gamma_x)}{2} \pm \sqrt{ g^2 + \left[\frac{E_c - E_x + i(\gamma_c - \gamma_x)}{2}\right]^2 }.$$5

the optimal action-value function

$$E_p = \frac{E_c + E_x + i(\gamma_c + \gamma_x)}{2} \pm \sqrt{ g^2 + \left[\frac{E_c - E_x + i(\gamma_c - \gamma_x)}{2}\right]^2 }.$$6

and the Bellman squared loss

$$E_p = \frac{E_c + E_x + i(\gamma_c + \gamma_x)}{2} \pm \sqrt{ g^2 + \left[\frac{E_c - E_x + i(\gamma_c - \gamma_x)}{2}\right]^2 }.$$7

The MAML-style objective is stated as

$$E_p = \frac{E_c + E_x + i(\gamma_c + \gamma_x)}{2} \pm \sqrt{ g^2 + \left[\frac{E_c - E_x + i(\gamma_c - \gamma_x)}{2}\right]^2 }.$$8

with inner-loop adaptation

$$E_p = \frac{E_c + E_x + i(\gamma_c + \gamma_x)}{2} \pm \sqrt{ g^2 + \left[\frac{E_c - E_x + i(\gamma_c - \gamma_x)}{2}\right]^2 }.$$9

and global meta-update

$E_x = 1.728$0

FRAP++ modifies FRAP by replacing the sum of lane demand for a phase with the mean of lane demand, reducing sensitivity to lane-count differences across intersections (Taschin et al., 18 Sep 2025).

The principal evaluation concern is distribution shift. Mismatch between train and test traffic distributions is quantified using

$E_x = 1.728$1

and

$E_x = 1.728$2

In synthetic experiments in CityFlow, MetaLight is compared with FRAP++ with adaptation and FRAP++ with no adaptation. The reported average travel times for MetaLight are 97 s, 96 s, 101 s, 116 s, and 153 s over five scenarios, corresponding to degradations of +22%, +12%, +5%, +5%, and +4% relative to the strongest baseline figures listed in the paper. FRAP++ with no adaptation reaches as high as 245 s in one scenario, corresponding to +66% (Taschin et al., 18 Sep 2025).

Real-world-derived evaluation uses Utah Department of Transportation ATSPM data from the intersection of US-89 and W 1200 N in Orem, Utah, with AM, Midday, and PM peak distributions. In one experiment, MetaLight reports 75 s, 68 s, and 96 s; in another, 61 s, 63 s, and 90 s. The same study finds that MetaLight can perform well in some shifted settings but does not consistently outperform simpler baselines, and that in the most difficult PM-peak shift RL with no adaptation sometimes beats MetaLight (Taschin et al., 18 Sep 2025).

An ablation on the number of meta-gradient steps tests 1, 2, 3, 5, and 10 steps. Performance improves initially, but after around 3–4 steps it worsens again. Reported timing is about 2.5 h for base-model training, about 2 min for adaptation, and about 2 h for FRAP++ training from scratch (Taschin et al., 18 Sep 2025). The stated mechanistic explanation is that Q-learning relies on bootstrapping and backward reward propagation, while MetaLight adaptation uses only a small number of gradient steps. If the new scenario’s optimal Q-function is far from the training one, a few updates may be insufficient (Taschin et al., 18 Sep 2025).

This transportation usage gives MetaLight a sharply different meaning from the optical one. Here it is not a metasurface or wavefront-shaping platform, but a fast-adaptation RL initialization for distributionally shifted control.

7. Extensions, controversies, and research significance

The most direct algorithmic extension of transportation MetaLight is ModelLight, a model-based meta-reinforcement learning framework for traffic signal control (Huang et al., 2021). ModelLight combines MetaLight-style meta-learning with an ensemble of learned intersection dynamics models $E_x = 1.728$3, an LSTM-based predictor of $E_x = 1.728$4 from $E_x = 1.728$5, and FRAP++ as the value network. The model-learning objective is

$E_x = 1.728$6

while the meta-learning objective is

$E_x = 1.728$7

Reported gains over the best baseline are 5.71% on Task 1, 17.94% on Task 2, and 9.69% on Task 3, and ModelLight-10r is reported to match or exceed MetaLight-100r on all tasks (Huang et al., 2021). This suggests that within the traffic-control literature, MetaLight became a reference point for rapid adaptation, but not the endpoint of the line of work.

In photonics, a different form of controversy concerns what metasurfaces should be expected to do. One recurring issue is the distinction between passive wavefront shaping and genuinely active or hybrid light–matter functionality. The exciton-polariton system explicitly advances beyond passive optics by using the metasurface to engineer strong coupling and polariton emission directionality (Chen et al., 2020). The quantum metasurface processor likewise moves beyond generation and tomography toward algorithm execution with both classical and single-photon light (Tanuwijaya et al., 2023). The detector-oriented metalens characterization paper, by contrast, argues that standard imaging metrics do not capture what matters for concentration onto small photodetectors, shifting attention from image fidelity to order-resolved collection efficiency versus position and incidence angle (Contreras et al., 2023). These works collectively indicate that MetaLight-related photonics is not a single application class but a design paradigm spanning strong coupling, light shaping, computation, imaging, and detection.

A final interpretive point follows from the full set of usages. In optics, MetaLight typically denotes a platform in which geometry-programmed subwavelength structure serves as the active intermediary between input field and output functionality. In transportation RL, MetaLight denotes a learned initialization that serves as the intermediary between a new traffic scenario and a task-adapted controller. This suggests a family resemblance at the level of systems architecture—an engineered intermediate layer that enables rapid transformation—while the underlying mathematics, physics, and application domains remain entirely different.