Dual-Efficiency: Coupled Performance Gains

• Dual-Efficiency is a recurring concept defined as the simultaneous optimization of two efficiency metrics, applied in ranging from electromagnetics to socio-technical systems.
• Research demonstrates its impact in dual-feed antennas, dual-beam EUV sources, and agentic AI by coordinating dual mechanisms to yield performance gains not attainable with single-channel approaches.
• Approaches in dual-efficiency balance multiple trade-offs and constraints in hardware design, optimization frameworks, and fairness in decision-making, emphasizing specialized dual structures over simple aggregation.

Dual-Efficiency is a field-dependent term used in recent research to denote the simultaneous improvement of two coupled efficiency dimensions, or the use of a dual structure whose interaction yields an efficiency gain not available to a single-channel design. In recent literature, the term has been used for dual-feed radiation-efficiency enhancement in open-earbud antennas (Liu et al., 20 Oct 2025), simultaneous dual-beam improvement of extreme ultraviolet conversion efficiency (Nagahama et al., 6 Jun 2026), and explicit step-level and trajectory-level efficiency for LLM agents (Chen et al., 19 Nov 2025). This suggests that Dual-Efficiency is not a single standardized theory, but a recurring technical motif spanning electromagnetics, optics, optimization, autonomous systems, and socio-technical analysis.

1. Field-specific meanings

Across the cited literature, Dual-Efficiency appears in several distinct but related senses. In some papers it denotes a physical dual-excitation mechanism that raises device efficiency; in others it denotes simultaneous optimization across two efficiency axes; and in still others it denotes a measurement or decision framework that evaluates multiple efficiency-relevant properties at once.

Context	Dual structure	Efficiency target
Earbud antenna	Dual-feed excitation with controlled phase difference	Radiation and total efficiency
EUV source	Simultaneous dual-beam irradiation	Conversion efficiency
LLM agents	Step-level and trajectory-level efficiency	Tokens per step and number of steps
Self-driving	Small reactive model plus larger informative model	Performance and near-real-time inference
Allocation and fairness	Simultaneous efficiency across dimensions or constraints	Welfare, error control, equal opportunity
Radiography	Joint treatment of artifact, resolution, and noise	Objective dual-energy image quality

The most explicit definitional use occurs in DEPO, where dual-efficiency is defined as “(i) step-level efficiency, which minimizes tokens per step, and (ii) trajectory-level efficiency, which minimizes the number of steps to complete a task” (Chen et al., 19 Nov 2025). A broader socio-technical variant appears in work on AI’s “dual paradox of efficiency, greater resource consumption, and displacement of traditional labor,” where efficiency gains are analyzed together with rebound energy demand and hybrid labor effects (Akpan et al., 9 Apr 2025). In formal allocation theory, the same family of ideas appears as simultaneous efficiency across multiple dimensions, with utilitarian social welfare and egalitarian social welfare treated separately but under one multidimensional framework (Kawase et al., 19 Jun 2026).

2. Dual excitation and dual mechanisms in physical systems

In electromagnetics and photonics, Dual-Efficiency often refers to a concrete hardware architecture in which two coordinated excitations reshape fields, currents, or energy deposition. The open-earbud antenna study provides a particularly direct example. It replaces a conventional single-feed configuration with a dual-feed excitation using two probes and a controlled phase difference, approximately $82^\circ$ in the realized antenna and around $90^\circ$ in the conceptual analysis. The device occupies approximately $25\times 15\times 3.6~\text{mm}^3$, achieves a measured $-6$ dB impedance bandwidth of 134 MHz in free space from 2373 to 2507 MHz and 140 MHz on head from 2380 to 2520 MHz, and reports measured average total efficiencies of $-8.5$ dB in free space and $-9.5$ dB on head. The paper attributes the gain primarily to radiation-efficiency improvement by current shaping: the dual-feed, controlled-phase architecture enlarges the “effective radiation area,” producing more uniform and stronger surface current over a larger patch region, with idealized radiation-efficiency gains of about $2$–$2.5$ dB and a realized improvement exceeding $1$ dB over single-feed operation (Liu et al., 20 Oct 2025). The same paper also stresses that the method is not mainly a bandwidth trick; the capacitor and lumped network mainly improve matching, whereas the dual feed drives the core efficiency enhancement.

A second physical instantiation appears in extreme ultraviolet source engineering. There, a 2090-nm, 20-ns Ho:YAG pulse with total on-target energy fixed at 40 mJ was either delivered as a single beam or split into two simultaneous 20 mJ beams while preserving the same optimal peak intensity of $1.6\times 10^{11}\,\mathrm{W/cm^2}$. The in-band EUV conversion efficiency rose from $90^\circ$0 to $90^\circ$1, a $90^\circ$2 relative enhancement, while the measured EUV source size remained about $90^\circ$3–$90^\circ$4 in both cases and the energetic-ion spectra were similar. The paper interprets the gain as more favorable spatial-temporal heating geometry and effective emitting surface, rather than simply a larger or hotter plasma (Nagahama et al., 6 Jun 2026). Here the dual structure is not a pair of independent sources in competition; it is a coordinated repartition of fixed total energy that preserves per-beam optimum intensity while changing plasma formation geometry.

A related but conceptually broader case arises in dielectric laser-driven accelerators, where “dual-efficiency” is realized by combining two distinct efficiency-improving mechanisms rather than two equal excitations. One mechanism, a Bragg reflector, enhances the accelerating field in the channel; the other, pulse-front tilt, extends effective interaction length. For a 100-period structure with $90^\circ$5, the paper reports energy gains of $90^\circ$6 keV for a bare dual grating with normal illumination, $90^\circ$7 keV with a Bragg reflector, $90^\circ$8 keV with pulse-front tilt, and $90^\circ$9 keV when both are combined. It states that the combined case increases energy gain by more than $25\times 15\times 3.6~\text{mm}^3$0 relative to using the Bragg reflector with a normal laser and by about $25\times 15\times 3.6~\text{mm}^3$1 relative to using standard structures with a pulse-front-tilted laser (Wei et al., 2017). In this usage, Dual-Efficiency means joint improvement of local field utilization and interaction-length utilization.

3. Dual-frequency, dual-polarization, dual-channel, and dual-window formulations

A second major usage of Dual-Efficiency is structural rather than purely energetic: efficiency becomes a function of how two coupled channels, modes, polarizations, or representations are coordinated. In nonlinear fiber optics, dispersive-wave generation from a dual-frequency beat signal is governed not only by phase matching but by the detuning-dependent higher-order soliton dynamics of individual beat cycles. With two monochromatic CW pumps at angular frequencies $25\times 15\times 3.6~\text{mm}^3$2 and $25\times 15\times 3.6~\text{mm}^3$3, the beat detuning is $25\times 15\times 3.6~\text{mm}^3$4, and the paper derives the effective soliton order

$25\times 15\times 3.6~\text{mm}^3$5

It shows that efficiency is lowest near $25\times 15\times 3.6~\text{mm}^3$6, recovers when $25\times 15\times 3.6~\text{mm}^3$7 and higher-order soliton compression and fission efficiently seed the dispersive wave, and then collapses again when the compression distance exceeds the 100 m fiber length. For the reported experiment, the high-efficiency region is approximately $25\times 15\times 3.6~\text{mm}^3$8, corresponding to $25\times 15\times 3.6~\text{mm}^3$9 (Webb et al., 2016). Here “dual” refers to a two-pump drive whose efficiency is set by the dynamics of the beat cycles that the two pumps create.

In metasurface image processing, dual-polarization operation is treated as an efficiency enabler rather than as a mere robustness feature. The reported nonlocal silicon metasurface performs isotropic Laplacian edge detection in transmission, with numerical aperture above 0.35 and a spectral bandwidth of 35 nm around 1500 nm. Its two engineered modes are designed so that both $-6$0- and $-6$1-polarized channels exhibit approximately the same Laplacian-like co-polarized transfer function while cross-polarized terms remain negligible. The paper introduces image-level throughput metrics,

$-6$2

and reports measured efficiencies $-6$3 and $-6$4 over 1452–1485 nm for narrowband almost-unpolarized illumination, with $-6$5 and $-6$6 for broadband input over a 35 nm spectrum (Cotrufo et al., 2022). The practical point is explicit in the paper: dual-polarization symmetry preserves useful edge information for both orthogonal incident polarizations rather than only one, thereby increasing effective throughput.

HybridGS applies a related logic to 3D Gaussian Splatting compression by introducing a “dual-channel sparse representation” that supervises primitive position and feature bit depth before standardized point-cloud encoding. The method generates compact explicit integer-valued 3DGS data, then uses GPCC with lossless octree geometry coding and RAHT attribute coding to produce standard output bitstreams. It explicitly models the per-primitive bit count as

$-6$7

and reports encoding/decoding times such as $-6$8 s and $-6$9 s for its high-rate settings, versus tens of seconds for several baselines. The paper states that HybridGS provides comparable reconstruction performance with evidently higher encoding and decoding speed, while also noting that its optimal compression efficiency is lower than end-to-end generation compression methods using RD loss (Yang et al., 3 May 2025).

A mathematically narrower but related usage appears in Gabor frame expansions, where the efficiency of reconstruction depends on the choice of dual window. Using $-8.5$0, $-8.5$1, and $-8.5$2 as generators, the paper evaluates canonical, symmetric, asymmetric, recursive, and perturbation-based duals by Average Mean Squared Error across Blocks, Bumps, Heavisine, Doppler, and Quadchirp. It concludes that exponential B-splines yield the lowest AMSE among the three generators, and that the symmetric dual and its perturbation-based variant consistently achieve the best reconstruction accuracy (Raghoothaman et al., 3 Nov 2025). In this setting, “dual-efficiency” means practical effectiveness of a dual window for exact Gabor reconstruction.

4. Dual-efficiency in agentic AI and autonomous systems

In machine learning systems, Dual-Efficiency is often formulated as simultaneous optimization of quality and inference economy, or of two granularities of search cost. DEPO provides the cleanest formal definition. It models agent-environment interaction as a POMDP and adds an efficiency bonus on desirable trajectories,

$-8.5$3

where $-8.5$4 is average tokens per step and $-8.5$5 is total steps. This directly encodes step-level and trajectory-level efficiency in one preference objective. The reported gains are up to $-8.5$6 lower token usage, up to $-8.5$7 fewer steps, and up to a $-8.5$8 improvement in performance on WebShop and BabyAI, with additional generalization to GSM8K, MATH, and SimulEq (Chen et al., 19 Nov 2025).

ETA addresses a related tension in self-driving. Its “dual approach” combines a small model for rapid reactive decisions with a larger model for slower but more informative analyses. The key shift is temporal: expensive computations for the current frame are moved to previous time steps, and a forecasting module propagates informative large-model features to the present while a small model processes the current frame in real time. On Bench2Drive CARLA Leaderboard-v2, ETA Async reports a driving score of $-8.5$9 at $-9.5$0 ms latency, compared with $-9.5$1 at $-9.5$2 ms for the base large-model system, and the abstract states that ETA advances state-of-the-art performance by $-9.5$3 while maintaining a near-real-time inference speed at $-9.5$4 ms (Hamdan et al., 9 Jun 2025). The “dual” element is a coupled large/small-model architecture rather than a single monolithic controller.

HyperEyes makes the efficiency objective itself two-grained. It argues that multimodal search agents should “search wider rather than longer,” and trains a parallel multimodal agent with a macro-level trajectory reward and micro-level token correction. At the macro level, TRACE uses the number of tool-call rounds $-9.5$5 and total tool invocations $-9.5$6, together with a reference that is monotonically tightened during training; at the micro level, On-Policy Distillation applies token-level KL supervision from an external teacher only on failed rollouts. The paper reports that HyperEyes-30B surpasses the strongest comparable open-source agent by $-9.5$7 in accuracy with $-9.5$8 fewer tool-call rounds on average across six benchmarks (Li et al., 8 May 2026).

A broader systems interpretation appears in the AI “dual paradox” paper, which frames AI efficiency as a tension between operational gains and system-level energy/resource expansion, and between automation efficiency and the persistent need for human labor and hybrid roles. Its proposition framework states: $-9.5$9

$2$0

$2$1

The paper reports, among other figures, data-center consumption of 240 to 340 TWh in 2022, approximately $2$2 to $2$3 of total electricity consumption, and states that only $2$4 of labor roles across major sectors have been impacted by AI (Akpan et al., 9 Apr 2025). This usage extends Dual-Efficiency from model design to socio-technical system analysis.

5. Simultaneous efficiency as a formal optimization problem

A more formal and abstract usage of Dual-Efficiency treats it as simultaneous satisfaction of multiple efficiency criteria under one allocation or decision rule. In multidimensional allocation of indivisible items, the MDEA model assigns each item value in multiple dimensions and studies simultaneous efficiency under utilitarian social welfare and egalitarian social welfare. Exact simultaneous efficiency may fail to exist even for $2$5, and for ESW even deciding whether two dimensions can be optimized simultaneously is NP-hard with binary valuations. For approximate simultaneous efficiency in every dimension, however, the paper identifies a tight threshold of order $2$6. Specializing to dual-efficiency, $2$7, this yields a worst-case factor of $2$8: a $2$9-sUmax$2.5$0 allocation and a $2.5$1-sEmax$2.5$2 allocation always exist and are computable in polynomial time, while any asymptotically better dependence on $2.5$3 is impossible in general (Kawase et al., 19 Jun 2026).

In fair classification, the NP-EO framework combines economic efficiency and fairness by minimizing type II error subject to a Neyman–Pearson cap on type I error and an equal-opportunity disparity bound,

$2.5$4

The resulting oracle classifier has a group-specific likelihood-ratio-threshold form, and the paper also gives finite-sample algorithms with high-probability population-level guarantees (Fan et al., 2023). In this setting, Dual-Efficiency is not a weighted average of fairness and utility; it is a constrained design that treats both as hard requirements.

A closely related optimization-over-efficient-set interpretation appears in multiobjective integer linear fractional programming. There the underlying $2.5$5 defines an efficient set $2.5$6, and two additional linear-fractional utility functions $2.5$7 are optimized over that efficient set: $2.5$8 The exact branch-and-bound plus cutting-plane method is designed to find solutions efficient for both the utility pair and the original multiobjective system “without going through all the efficient solutions of the two problems” (Chaiblaine et al., 2020). This is a literal dual-efficiency problem over an already Pareto-filtered domain.

6. Evaluation metrics, trade-offs, and limitations

Not all uses of Dual-Efficiency are constructive; some are metrological. In tissue-subtraction radiography, Dual-Energy Subtraction Efficiency (DSE) is an objective figure of merit defined as

$2.5$9

where $1$0 measures residual unwanted-tissue artifact power, $1$1 is spatial-resolution transfer, and $1$2 is noise transfer. The metric was proposed because dual-energy image quality cannot be judged by noise alone, resolution alone, or subtraction level alone; it must capture cancellation quality, retained resolution, and quantum-noise amplification together (Maurino et al., 2020). This is a pure measurement-theoretic version of Dual-Efficiency.

The literature also repeatedly stresses trade-offs. The heralded-multiplexed SPDC source improves the fidelity–success-probability trade-off of dual-rail polarization-entangled photon-pair generation, but the paper identifies a hard implementation threshold: there is “a threshold of $1$3 dB of loss per switch, beyond which multiplexing hurts the Fidelity versus Success Probability trade-off” (Dhara et al., 2021). The earbud antenna paper emphasizes that dual-feed enhancement comes at the price of a power divider, a phase shifter, a matching network, insertion loss, tuning effort, and sensitivity to phase accuracy (Liu et al., 20 Oct 2025). HybridGS explicitly states that its optimal compression efficiency is lower than end-to-end generation compression methods using RD loss, even though it is much faster and more standard-friendly (Yang et al., 3 May 2025). The AI dual-paradox analysis likewise argues that efficiency gains can increase total electricity use, cooling demand, and infrastructure expansion rather than reducing them (Akpan et al., 9 Apr 2025).

A common misconception, addressed differently across domains, is that a dual structure helps merely by “adding more.” The antenna study states that the effect is not a mysterious matching trick and not mainly a bandwidth trick; it is a current-distribution and radiation-aperture effect (Liu et al., 20 Oct 2025). The EUV study argues that the $1$4 gain is not explained simply by a larger source size, since both configurations produced a $1$5–$1$6 source (Nagahama et al., 6 Jun 2026). In the allocation and fairness papers, simultaneous efficiency is likewise not achieved by scalar aggregation alone; it requires explicit multidimensional or constrained formulations (Kawase et al., 19 Jun 2026, Fan et al., 2023).

This suggests that Dual-Efficiency functions less as a single doctrine than as a recurring design heuristic: when a system is limited by one-axis optimization, introducing a carefully coordinated second channel, second objective, or second granularity can expose an efficiency frontier that a single-feed, single-beam, single-metric, or single-scale formulation cannot reach.