LightBeam: Beam Shaping & Neuroprosthetic Decoding

• LightBeam is a multifaceted concept that includes source-level OLED beam shaping, programmable metasurface steering for infrared links, and efficient CTC decoding for speech neuroprostheses.
• It leverages microcavity design, polarization control, and electrical modulation to dynamically adjust emission patterns, supporting adaptive lighting and high-speed communications.
• The neuroprosthetic variant integrates a delayed LLM fusion in a non-WFST CTC decoder, significantly reducing memory requirements while maintaining real-time performance.

LightBeam is a name applied in the cited literature to several distinct technical constructs. In optics, it denotes real-time, source-level beam-shaping with a two-unit OLED microcavity that regulates emission between strongly forward and strongly sideward modes without secondary optical elements or mechanical adjustment (Fries et al., 2017), and it also denotes a passively field-programmable metasurface architecture for a 20 Gbps beam-steered infrared wireless link (Huang et al., 2020). In speech neuroprosthetics, LightBeam names a non-WFST CTC decoder that integrates an LLM into beam search via delayed fusion while reducing memory from the scale of a large WFST graph to approximately 10 GB (Feghhi et al., 14 Mar 2026). A plausible common theme is direct control at the primary source, interface, or decoding layer rather than reliance on bulky secondary subsystems.

1. Nomenclature and research scope

Within beam-control research, the optical uses of LightBeam sit inside a broader structured-light landscape that includes engineered 3D waveguide arrays for free-space projection (Crespi et al., 2016), multi-layer liquid-crystal cells for shifting, steering, and expanding beams (Mur et al., 2022), dielectric-particle photonic hooks (Yue et al., 2017), extended Frozen Waves for simultaneous longitudinal and transverse shaping (Zamboni-Rached, 2021), integrated steerable vortex lasers based on photonic-crystal bound states in the continuum (Bahari et al., 2017), and light springs with tunable orbital group velocity (Vaz et al., 21 Feb 2026). These works span distinct physical regimes—microcavities, metasurfaces, birefringent media, mesoscale dielectric focusing, non-diffracting beam synthesis, topological photonic resonances, and spatiotemporal OAM-frequency coupling—but all address the controlled production of nontrivial angular, spatial, phase, or temporal beam structure.

The term therefore does not refer to a single canonical device class. In the optical literature represented here, LightBeam can indicate either a source-level emitter whose far-field distribution is changed by selecting among stacked emissive sub-units, or a remote passive metasurface whose output angle is determined by centrally controlled wavelength and polarization. Separately, in neuroprosthetics, it denotes a decoding algorithm rather than an optical system. This terminological multiplicity is important because the design variables, figures of merit, and constraints differ sharply across these domains.

2. Source-level optical LightBeam in stacked OLED microcavities

The LightBeam concept summarized from "Real-time beam-shaping without additional optical elements" employs two vertically stacked bottom-emitting OLED sub-units, with total thickness of approximately $800\,\mathrm{nm}$, on a glass/ITO substrate, separated by a $10\,\mathrm{nm}$ Au/Ag wetting-layer metal electrode that serves as the common electrode in AC operation (Fries et al., 2017). Each sub-unit follows a p-i-n design with electrically doped transport layers for low resistance. The hole injection/transport layer is $4\,\mathrm{wt}\,\%$ F6-TCNNQ doped into Spiro-TTB; the emissive layer is $10\,\mathrm{wt}\,\%$ phosphorescent Ir-complex doped in NPB, using $\mathrm{Ir(piq)_3}$ for forward-mode and $\mathrm{Ir(MDQ)_2(acac)}$ for side-mode; the electron transport/injection layer is Cs-doped BPhen; and blocking layers are chosen from BPhen or BAlq$_2$, and Spiro-TAD or NPB, to confine carriers. The bottom sub-unit, OLED 1, places its EML at an optical field maximum to favor on-axis emission, whereas the top sub-unit, OLED 2, places its EML at a field minimum at $0^\circ$, causing its emission to peak off-axis near $50^\circ$ to $80^\circ$.

The beam-shaping mechanism is the microcavity itself. The full two-unit stack forms a second-order planar microcavity in which field maxima and minima appear across the thickness, and the EML position relative to those extrema determines the angular emission profile. Mirror reflectivities set the cavity finesse, and as the observation angle $10\,\mathrm{nm}$0 increases, the cavity resonance blueshifts approximately as

$10\,\mathrm{nm}$1

Placing an EML at a field node for $10\,\mathrm{nm}$2 therefore suppresses on-axis electroluminescence and allows strong off-axis emission only at larger $10\,\mathrm{nm}$3. By selecting two sub-units with complementary cavity lengths and EML placements, the device emits a tight forward beam when OLED 1 is addressed, a ring-shaped side beam when OLED 2 is addressed, and intermediate patterns by time-multiplexing the two.

The spectral radiant intensity per unit area follows

$10\,\mathrm{nm}$4

where only $10\,\mathrm{nm}$5 varies with $10\,\mathrm{nm}$6. The external quantum efficiency is

$10\,\mathrm{nm}$7

Measured EQE maxima are $10\,\mathrm{nm}$8 for OLED 1 and $10\,\mathrm{nm}$9 for OLED 2, and over the full beam-shape tuning range the device remains between approximately $4\,\mathrm{wt}\,\%$0 and $4\,\mathrm{wt}\,\%$1. For angular irradiance mapping, the reported conversion between the spherical-angle scan and the brightness distribution on a flat projection screen is

$4\,\mathrm{wt}\,\%$2

Experimentally, the $4\,\mathrm{wt}\,\%$3–$4\,\mathrm{wt}\,\%$4–$4\,\mathrm{wt}\,\%$5 curves are steep despite the thick cavity, which is attributed to the highly conductive doped layers. Angle-resolved emission shows that OLED 1 peaks at $4\,\mathrm{wt}\,\%$6 and falls by $4\,\mathrm{wt}\,\%$7 by $4\,\mathrm{wt}\,\%$8, while OLED 2 peaks at $4\,\mathrm{wt}\,\%$9 with a $10\,\mathrm{wt}\,\%$0 enhancement relative to its $10\,\mathrm{wt}\,\%$1 value. The forward unit has $10\,\mathrm{wt}\,\%$2 at $10\,\mathrm{wt}\,\%$3 with FWHM of approximately $10\,\mathrm{wt}\,\%$4; the side unit peaks at $10\,\mathrm{wt}\,\%$5 near $10\,\mathrm{wt}\,\%$6, leaving a $10\,\mathrm{wt}\,\%$7 shift even after choosing a blue-shifted emitter to compensate cavity dispersion.

Real-time tuning is obtained through AC/DC driving with pulse-width modulation that addresses only one sub-unit at a time. A square-wave generator and high-voltage amplifier switch the polarity so that positive polarity addresses OLED 1 and negative polarity addresses OLED 2. The duty cycle determines the time-averaged mixture of forward and side emission, and sweeping the duty ratio in approximately $10\,\mathrm{wt}\,\%$8 steps continuously morphs the beam from a tight spot to a uniform glow to a hollow ring. Switching speed is set by OLED turn-on dynamics, approximately $10\,\mathrm{wt}\,\%$9 for phosphorescent devices, which permits beam-shape modulation from the kHz to MHz range.

System integration with additional optics remains possible even though the primary beam shaping is source-level. Adding a glass half-sphere suppresses total internal reflection and reduces beam divergence; in the reported measurements, the forward-mode FWHM narrows from $\mathrm{Ir(piq)_3}$0 to $\mathrm{Ir(piq)_3}$1, and the side-mode contrast between center and ring rises from $\mathrm{Ir(piq)_3}$2 to $\mathrm{Ir(piq)_3}$3. A prism produces asymmetric beam patterns, such as different narrowing along $\mathrm{Ir(piq)_3}$4 and $\mathrm{Ir(piq)_3}$5, without moving parts.

3. Passively field-programmable metasurface LightBeam for infrared wireless links

In "A 20-Gbps Beam-steered Infrared Wireless Link Enabled by a Passively Field-programmable Metasurface", LightBeam denotes a beam-steering system built around a passive metasurface whose function is selected remotely through centralized control of wavelength and polarization (Huang et al., 2020). Each meta-atom sits on a Si substrate and consists of a $\mathrm{Ir(piq)_3}$6 Au ground plane, a $\mathrm{Ir(piq)_3}$7 SiO$\mathrm{Ir(piq)_3}$8 spacer, and a top Au nanopattern. Type 1 meta-atoms, used for six of the eight phase steps, comprise two identical rectangular Au patches in parallel; type 2, used for the $\mathrm{Ir(piq)_3}$9 phase, is a single rectangular patch. Along the $\mathrm{Ir(MDQ)_2(acac)}$0-direction, a super-cell contains eight pixels engineered to span the discrete phase sequence $\mathrm{Ir(MDQ)_2(acac)}$1.

Under $\mathrm{Ir(MDQ)_2(acac)}$2-polarized incidence, the super-cell produces a monotonic phase ramp of $\mathrm{Ir(MDQ)_2(acac)}$3 over a spatial period $\mathrm{Ir(MDQ)_2(acac)}$4, leading to anomalous reflection governed by

$\mathrm{Ir(MDQ)_2(acac)}$5

Under $\mathrm{Ir(MDQ)_2(acac)}$6-polarized incidence, all eight pixels have essentially the same reflection phase, within $\mathrm{Ir(MDQ)_2(acac)}$7, so the device behaves as a conventional mirror. The more general phase-gradient relation is

$\mathrm{Ir(MDQ)_2(acac)}$8

or equivalently

$\mathrm{Ir(MDQ)_2(acac)}$9

The defining point is that no local electrical addressing is required; a liquid-crystal polarization controller placed before the metasurface selects the polarization state, and thus whether a given beam is steered anomalously or reflected specularly.

The system architecture distributes steering across wavelength, free-space geometry, and polarization. An AWGR in the communication control center accepts discrete wavelengths $_2$0 and routes each to a distinct fiber in a $_2$1 $_2$2 array. After emerging from the fibers, the beams pass through a half-lens, focus onto the metasurface, and reflect into free space. Tuning the wavelength selects which column of the fiber array is illuminated and therefore provides discrete steering along the $_2$3-axis, while switching polarization via the liquid-crystal controller provides an additional steering degree along the $_2$4-axis.

The reported steering range is set by both anomalous and normal branches. Maximum anomalous deflection is associated with incidence up to the critical angle $_2$5, where $_2$6, giving $_2$7. In practice, beams are launched at $_2$8 from $_2$9 to $0^\circ$0, and the anomalous branch covers approximately $0^\circ$1 of output angle, exemplified as $0^\circ$2 to $0^\circ$3. Normal reflection covers roughly $0^\circ$4 in $0^\circ$5. In the numerical example with $0^\circ$6, $0^\circ$7, $0^\circ$8, $0^\circ$9, and $50^\circ$0, the angular step in $50^\circ$1 is approximately $50^\circ$2, while the step in $50^\circ$3 is non-uniform but averages about $50^\circ$4 per column.

The proof-of-concept communication experiment used a tunable laser over $50^\circ$5 to $50^\circ$6, $50^\circ$7 optical power, a $50^\circ$8 PAM-4 modulator chain, a passive metasurface polarization beam-splitter chip at the remote access point, and a $50^\circ$9 free-space link collected into SMF. The achieved data rate was $80^\circ$0. At the $80^\circ$1-FEC threshold of $80^\circ$2, the penalty relative to back-to-back operation was approximately $80^\circ$3, attributed mainly to EDFA noise, and the eye diagrams showed clean four-level openings at threshold.

Its efficiency figures are explicitly angular and polarization dependent. For anomalous reflection under $80^\circ$4-polarization, efficiency exceeds $80^\circ$5 for $80^\circ$6, exceeds $80^\circ$7 for $80^\circ$8, and remains above $80^\circ$9 out to $10\,\mathrm{nm}$00. For normal reflection under $10\,\mathrm{nm}$01-polarization, efficiency exceeds $10\,\mathrm{nm}$02 for most incidence angles up to $10\,\mathrm{nm}$03, with notches of $10\,\mathrm{nm}$04 to $10\,\mathrm{nm}$05 near $10\,\mathrm{nm}$06 and $10\,\mathrm{nm}$07. Polarization isolation is $10\,\mathrm{nm}$08 across $10\,\mathrm{nm}$09 to $10\,\mathrm{nm}$10. Main losses arise from Au/SiO$10\,\mathrm{nm}$11 absorption, angular dispersion in TM mode at some angles, and the discrete $10\,\mathrm{nm}$12 phase steps, which introduce minor side-lobes.

4. Placement within structured-light research

The optical meanings of LightBeam can be situated relative to several beam-engineering strategies that control different parts of the field-generation chain.

Platform	Primary control variable	Reported beam effect
3D waveguide arrays (Crespi et al., 2016)	$10\,\mathrm{nm}$13, $10\,\mathrm{nm}$14, $10\,\mathrm{nm}$15	OAM, Hermite-like lobes, multi-singularity arrays, Bessel-like beams
Multi-layer liquid-crystal cells (Mur et al., 2022)	$10\,\mathrm{nm}$16, $10\,\mathrm{nm}$17, $10\,\mathrm{nm}$18, $10\,\mathrm{nm}$19	Shift, steering, expansion
Photonic hook (Yue et al., 2017)	$10\,\mathrm{nm}$20, $10\,\mathrm{nm}$21, $10\,\mathrm{nm}$22	Curved high-intensity focus
Extended Frozen Waves (Zamboni-Rached, 2021)	$10\,\mathrm{nm}$23	Independent longitudinal and transverse shaping
BIC vortex lasers (Bahari et al., 2017)	$10\,\mathrm{nm}$24, hole radius $10\,\mathrm{nm}$25	Integrated OAM generation and steering
Light springs (Vaz et al., 21 Feb 2026)	$10\,\mathrm{nm}$26	Helical wavepacket, tunable orbital group velocity

In 3D waveguide arrays, the field is synthesized by coherent superposition of Gaussian single-mode outputs,

$10\,\mathrm{nm}$27

and in the far field by the Fourier-transform relation between the single-mode envelope and the phasor sum over waveguide positions. Regular polygon placement with the cyclic phase ramp $10\,\mathrm{nm}$28 produces OAM beams whose on-axis behavior scales as $10\,\mathrm{nm}$29, while other arrangements yield two-lobe beams, multi-singularity patterns, Bessel-like non-diffracting beams, or Fresnel-lens focusing (Crespi et al., 2016). This route differs from OLED LightBeam and metasurface LightBeam because the beam is formed by interferometric synthesis at the emitting facet rather than by cavity asymmetry or polarization-selective reflection.

The multi-layer liquid-crystal device combines three function-specific tunable birefringent layers: a shifter using walk-off, a deflector using a transverse phase gradient, and an expander using a radially escaped $10\,\mathrm{nm}$30-defect profile. The Jones-matrix representation of a sub-cell is

$10\,\mathrm{nm}$31

and the steering relation is expressed through $10\,\mathrm{nm}$32 and the resulting angle $10\,\mathrm{nm}$33. Reported values include a maximum lateral shift $10\,\mathrm{nm}$34 for $10\,\mathrm{nm}$35 and $10\,\mathrm{nm}$36, steering up to approximately $10\,\mathrm{nm}$37 for $10\,\mathrm{nm}$38 and up to approximately $10\,\mathrm{nm}$39 for $10\,\mathrm{nm}$40, and beam-waist expansion from $10\,\mathrm{nm}$41 to approximately $10\,\mathrm{nm}$42 (Mur et al., 2022).

The photonic hook is a distinct near-field phenomenon generated by an asymmetric dielectric particle. With local thickness

$10\,\mathrm{nm}$43

the exit-face phase

$10\,\mathrm{nm}$44

becomes asymmetric, and interference in the near field produces a curved high-intensity focus that bends toward the thicker side of the trapezoid. The reported maximum bending angle is approximately $10\,\mathrm{nm}$45 at $10\,\mathrm{nm}$46, and the fused-silica/air example yields FWHM below $10\,\mathrm{nm}$47 (Yue et al., 2017). By contrast, the OLED LightBeam and metasurface LightBeam operate in far-field shaping and steering regimes.

Extended Frozen Waves start from the scalar Helmholtz equation and a superposition of co-propagating Bessel beams,

$10\,\mathrm{nm}$48

with $10\,\mathrm{nm}$49 and coefficients

$10\,\mathrm{nm}$50

The key extension is the decomposition $10\,\mathrm{nm}$51, where $10\,\mathrm{nm}$52 controls the longitudinal intensity profile and $10\,\mathrm{nm}$53 controls the local transverse radius via $10\,\mathrm{nm}$54 (Zamboni-Rached, 2021). This offers an exact analytic route to beams whose spot radius or ring radius varies with $10\,\mathrm{nm}$55, which is conceptually different from the discrete beam-state interpolation of the OLED device or the angular redirection of the metasurface.

Integrated BIC vortex lasers use accidental bound states in the continuum in an InGaAsP photonic-crystal slab so that the high-$10\,\mathrm{nm}$56 singularity appears at nonzero in-plane wavevector $10\,\mathrm{nm}$57, with emission angle satisfying

$10\,\mathrm{nm}$58

Varying the hole radius $10\,\mathrm{nm}$59 shifts $10\,\mathrm{nm}$60 and hence steers the emitted OAM beam. The reported device produces $10\,\mathrm{nm}$61 and $10\,\mathrm{nm}$62 vortex modes, threshold pump of approximately $10\,\mathrm{nm}$63, output power up to approximately $10\,\mathrm{nm}$64, FWHM divergence of approximately $10\,\mathrm{nm}$65, and OAM purity above $10\,\mathrm{nm}$66 (Bahari et al., 2017).

Light springs extend structured-light control into the spatiotemporal domain. Their defining relation is the linear OAM-frequency mapping

$10\,\mathrm{nm}$67

which leads to a helical wavepacket and an orbital group velocity

$10\,\mathrm{nm}$68

Experiments tuned $10\,\mathrm{nm}$69 from approximately $10\,\mathrm{nm}$70 to approximately $10\,\mathrm{nm}$71 by varying $10\,\mathrm{nm}$72, and particle-in-cell simulations showed that, for $10\,\mathrm{nm}$73, interaction with a thin overdense plasma produces superradiant terahertz emission (Vaz et al., 21 Feb 2026). The paper explicitly states that $10\,\mathrm{nm}$74 does not transport energy or information superluminally.

5. Applications, integration pathways, and technical limitations

The optical LightBeam variants address different application layers. The OLED implementation targets adaptive lighting, including automotive headlamps that switch between spot and wide-angle road illumination, dynamic architectural or stage lighting, projection systems, structured illumination for microscopy or sensing, and display systems with per-pixel beam-shape control that could enable high-contrast or privacy modes (Fries et al., 2017). The metasurface implementation targets infrared optical wireless links with centralized control, simple and cheap remote-side devices, and scalability to multiple beams in an $10\,\mathrm{nm}$75 addressing scheme (Huang et al., 2020). Related liquid-crystal, BIC-laser, Frozen-Wave, photonic-hook, and light-spring systems extend the application envelope toward AR/VR optics, LiDAR, optical tweezers, biological sensing, microscopy, optical trapping, atom guidance, lithography, ultrafast spectroscopy, and THz-source engineering (Mur et al., 2022, Bahari et al., 2017, Zamboni-Rached, 2021, Yue et al., 2017, Vaz et al., 21 Feb 2026).

System integration proceeds differently in each class. The OLED device can be combined with a glass half-sphere or a prism while keeping the tuning origin at the light source itself rather than in mechanically adjustable secondary optics (Fries et al., 2017). The metasurface system combines AWGR-based wavelength routing, a half-lens, a liquid-crystal polarization controller, and a passive polarization-selective metasurface in a remote access point (Huang et al., 2020). The liquid-crystal approach relies on ultra-thin polymer walls, patterned surface anchoring, and electrically addressed ITO or conductive polymer electrodes (Mur et al., 2022). Integrated BIC vortex lasers collapse generation and steering into a single photonic-crystal emitter (Bahari et al., 2017). A plausible implication is that LightBeam-like architectures are most naturally compared not by beam shape alone, but by where the control variable is physically located: emitter stack, reflective interface, waveguide network, anisotropic medium, photonic-crystal band structure, or spatiotemporal spectral synthesis.

The limitations are equally heterogeneous. In the OLED case, a color shift of approximately $10\,\mathrm{nm}$76 remains between modes because of cavity dispersion; extension to green and blue requires tighter thickness tolerances, and maintaining high-contrast beam shapes on a target plane imposes the étendue-related condition $10\,\mathrm{nm}$77 (Fries et al., 2017). In the metasurface case, discrete $10\,\mathrm{nm}$78 phase steps create quantization lobes, steering resolution is set by AWGR port count and polarization states, fabrication by e-beam lithography is slow and small-area, and large wavelength tunability is limited by AWGR channel spacing and metasurface dispersion (Huang et al., 2020). In the liquid-crystal device, typical $10\,\mathrm{nm}$79 nematic cells respond in $10\,\mathrm{nm}$80 to $10\,\mathrm{nm}$81, motivating polymer-stabilized or blue-phase materials for kHz operation (Mur et al., 2022). Extended Frozen Waves are constrained by finite aperture, truncation-induced sidelobes, the scalar approximation, and the condition $10\,\mathrm{nm}$82 with $10\,\mathrm{nm}$83 for the standard construction (Zamboni-Rached, 2021). Light springs add a subtler conceptual limitation: even when $10\,\mathrm{nm}$84, the superluminal quantity is a synthetic transverse motion rather than superluminal transport of information or energy (Vaz et al., 21 Feb 2026).

6. LightBeam as a CTC decoder for speech neuroprostheses

Outside optics, "LightBeam: An Accurate and Memory-Efficient CTC Decoder for Speech Neuroprostheses" uses the name for a GPU-accelerated, non-WFST CTC decoder intended for real-time speech neuroprostheses (Feghhi et al., 14 Mar 2026). Its input is CTC encoder logits over phoneme, blank, and space tokens from neural recordings such as intracranial ECoG. The decoder replaces the WFST plus 5-gram graph used in prior leading published work with a vectorized lexicon and delayed LLM fusion. The high-level scoring combines scaled acoustic log-probabilities, shallow fusion with a small N-gram LM for homophone disambiguation, and delayed first-pass integration of a fine-tuned LLM, specifically Llama 3.2 1B.

The CTC training objective is stated as

$10\,\mathrm{nm}$85

and the hypothesis score during decoding is

$10\,\mathrm{nm}$86

At every $10\,\mathrm{nm}$87 frames, active orthographic beams are batched through the LLM and the score is updated by

$10\,\mathrm{nm}$88

Beam search uses top-$10\,\mathrm{nm}$89 pruning, a relative threshold $10\,\mathrm{nm}$90, and bonuses $10\,\mathrm{nm}$91 and $10\,\mathrm{nm}$92 for token extension and word insertion. The decoder is implemented in Python, uses PyTorch 2.x and torchaudio’s CUCTCDecoder, and stores the lexicon as a dense transition table $10\,\mathrm{nm}$93.

Its principal distinction is memory efficiency. The reported WFST-based decoder requires peak RAM of approximately $10\,\mathrm{nm}$94, whereas LightBeam requires approximately $10\,\mathrm{nm}$95 RAM and $10\,\mathrm{nm}$96 to $10\,\mathrm{nm}$97 VRAM. On Brain-to-Text ’24 and ’25 with the baseline GRU encoder, the reported LightBeam results are $10\,\mathrm{nm}$98 WER on B2T ’24, $10\,\mathrm{nm}$99 WER on B2T ’25 public, and $4\,\mathrm{wt}\,\%$00 WER on B2T ’25 private, with RAM of $4\,\mathrm{wt}\,\%$01 and $4\,\mathrm{wt}\,\%$02 and RTF of $4\,\mathrm{wt}\,\%$03 and $4\,\mathrm{wt}\,\%$04. The WFST re-implementation reports $4\,\mathrm{wt}\,\%$05, $4\,\mathrm{wt}\,\%$06, and $4\,\mathrm{wt}\,\%$07 WER with RAM of $4\,\mathrm{wt}\,\%$08 and $4\,\mathrm{wt}\,\%$09. With a causal, time-masked Transformer encoder, LightBeam reaches $4\,\mathrm{wt}\,\%$10, $4\,\mathrm{wt}\,\%$11, and $4\,\mathrm{wt}\,\%$12 WER, versus $4\,\mathrm{wt}\,\%$13, $4\,\mathrm{wt}\,\%$14, and $4\,\mathrm{wt}\,\%$15 for WFST. Ablations show significant degradation when delayed LLM rescoring or next-word fine-tuning is removed.

This non-optical use of the name is technically unrelated to beam shaping in photonics, but it preserves the central role of beam search as the object being controlled. In that sense, LightBeam in neuroprosthetics is not a light-emission technology at all; it is an algorithmic decoder whose novelty lies in replacing a large graph-based search structure with a memory-efficient lexicon and LLM-assisted delayed fusion (Feghhi et al., 14 Mar 2026).