Less In More Out Device: Efficient Design

• Less In More Out Device is a design principle that removes redundant inputs to achieve higher task-level utility across domains like edge retrieval and federated learning.
• It employs structured suppression to optimize system architectures, resulting in improved efficiency, stability, and reduced energy consumption.
• Applications span always-on systems, optical spaceplates, and soft robotic hearts, where controlled input reduction yields higher effective output performance.

“Less In More Out Device” denotes, across several otherwise unrelated research programs, a design pattern in which deliberately reducing stored data, local workload, interface width, physical path length, incentive intensity, or actuator count yields higher end utility than a naïve “keep everything, compute everything, or enlarge everything” strategy. In always-on cross-modal retrieval, an edge camera stores far fewer semantically novel frames and only compact embeddings, yet improves Hit@5 (Abdelwahab, 31 Mar 2026). In federated learning, battery-constrained clients use only a fraction of local data per round and remain active for more rounds, improving stability and final accuracy (Pereira et al., 2024). In optics, a thin multilayer stack behaves like a much longer region of free-space propagation (Hogan et al., 25 Jun 2025). In soft robotics, a ventricular chamber with fewer pouches delivers higher stroke volume, albeit with higher stress (Lauber et al., 9 Sep 2025). Taken together, these works suggest that “less in” is not simple minimization; it is structured suppression of redundancy, excess coupling, or inefficient intermediates.

1. Definitional scope and semantic range

Across the cited works, the phrase has both a literal and an analogical use. In some cases it names a concrete device, as with the fabric-based soft ventricular chamber called the “Less In More Out” device (Lauber et al., 9 Sep 2025). In others it functions as a systems principle: store fewer frames but retrieve more correct events; spend less device energy but obtain more global learning; assign less responsibility to an on-device SLM but obtain a more reliable feature; offer less reward in D2D offloading and obtain more usable capacity once epidemic effects are considered (Abdelwahab, 31 Mar 2026, Pereira et al., 2024, Oliveira, 27 Apr 2026, Xu et al., 2016).

Domain	“Less in”	“More out”
Always-on edge retrieval	Fewer frames; embeddings only	Higher text-to-video retrieval
Federated learning	Less local data and energy per round	More rounds, more active devices, higher accuracy
On-device SLM integration	Narrower model responsibility	More robust production behavior
D2D offloading incentives	Less reward	More usable capacity and higher operator utility
DL accelerator hardware	Fewer ineffectual term-term products; narrower memory interface	More throughput and energy efficiency
Integrated photonics and spaceplates	Shorter device or thinner element	Larger mode transformation or longer effective propagation
Soft robotic heart	Fewer pouches	Higher stroke volume

This suggests a unifying description: a Less In More Out device is one whose architecture is explicitly organized to remove redundancy, crowding, or physically wasteful path length before downstream utility is measured. The “more out” side is therefore typically not raw output volume but a task-level metric: Hit@5, global accuracy, effective participation, usable offloading capacity, expansion ratio, compression ratio, stroke volume, or efficiency.

2. Information, learning, and incentive systems

In always-on edge retrieval, the concept is instantiated as a streaming retrieval architecture for AR glasses, body cams, dashcams, or robots recording continuously at approximately $5$ FPS, where redundant frames both increase storage and crowd correct results out of top-$k$ search (Abdelwahab, 31 Mar 2026). Each incoming frame $f_i$ is mapped to a unit-norm embedding $\mathbf{v}_i = \phi(f_i)$, and the device retains the frame only if

$$s_i = \max_{\mathbf{m}\in\mathcal{M}} \mathbf{v}_i^\top \mathbf{m} < \tau .$$

With $\varepsilon = 1-\tau$, the retained keyframes form a maximal $\varepsilon$-net: discarded frames are covered, and retained frames remain separated. This single-pass filter outperforms offline alternatives such as k-means, farthest-point, uniform, and random across eight vision-LLMs on AEA and EPIC-KITCHENS. Combined with a LiT-style cross-modal adapter and SigLIP 2 re-ranking, the system reaches $45.6\%$ Hit@5 on held-out data using an $8$M on-device encoder at an estimated $2.7$ mW; on AEA at $k$0 FPS and $k$1, the index is reduced by approximately $k$2, and each kept frame stores about $k$3 KB instead of about $k$4 KB for a raw image. The key geometric claim is that redundancy is not only a storage problem but a retrieval problem: near-duplicate clusters induce cluster size bias and crowd out semantically relevant events.

In federated learning, “less in” appears as adaptive data usage under battery constraints. LeanFed assigns each device $k$5 a per-round data fraction

$k$6

with $k$7 the battery budget, $k$8 the local epochs, $k$9 the expected energy cost per round on a full dataset, and $f_i$0 the expected participation rate (Pereira et al., 2024). Rather than training on all of $f_i$1 in every round, a device uses only $f_i$2 samples and updates its battery estimate by $f_i$3. The resulting effect is system-level: more devices remain active deeper into training, dropout is mitigated, and convergence is more stable than naïve FedAvg under heterogeneity. The reported evaluations cover CIFAR-10 and CIFAR-100 with $f_i$4 and $f_i$5 devices, $f_i$6 and $f_i$7 rounds, and participation rates from $f_i$8 to $f_i$9. For CIFAR-100, $\mathbf{v}_i = \phi(f_i)$0, and $\mathbf{v}_i = \phi(f_i)$1, LeanFed reaches about $\mathbf{v}_i = \phi(f_i)$2 test accuracy, while the best FedAvg variant is around $\mathbf{v}_i = \phi(f_i)$3–$\mathbf{v}_i = \phi(f_i)$4. Here the device does less local work per round, but the federation learns more over the full horizon.

The same logic appears in practitioner-oriented on-device SLM integration. In the Palabrita Android word-guessing game, the original architecture required the on-device model to generate an entire puzzle as JSON: word, category, difficulty, and five progressive hints. After a five-day development sprint comprising $\mathbf{v}_i = \phi(f_i)$5 commits, the production architecture instead used curated word lists and deterministic app logic for word selection, difficulty, and progression, leaving the SLM responsible only for three short hints and supplying a deterministic fallback if it failed (Oliveira, 27 Apr 2026). The observed failure categories were output format violations, constraint violations, context quality degradation, latency incompatibility, and model selection instability. The converged rule was explicit: “the most reliable on-device LLM feature is one where the LLM does the least.” In this case, “less in” means less schema complexity and less model responsibility; “more out” means a feature that remains playable, offline, and robust.

In D2D computation offloading, the principle is inverted into incentive design. The operator pays reward $\mathbf{v}_i = \phi(f_i)$6 per task, but higher participation increases infection exposure in an SIS epidemic model with effective infection rate $\mathbf{v}_i = \phi(f_i)$7 (Xu et al., 2016). The operator’s utility under infection is

$\mathbf{v}_i = \phi(f_i)$8

where only healthy devices contribute to effective participation. The paper shows that once the attack-free participation would exceed the epidemic threshold, persistent infection emerges, effective participation is pinned by the epidemic equilibrium, and further increases in $\mathbf{v}_i = \phi(f_i)$9 only raise payments. The resulting structural condition is that the security-aware reward design is equivalent to an attack-free problem with the cap

$$s_i = \max_{\mathbf{m}\in\mathcal{M}} \mathbf{v}_i^\top \mathbf{m} < \tau .$$0

Thus less reward can produce more usable offloading capacity and higher operator utility.

3. Arithmetic and accelerator realizations

In digital hardware, the principle appears as removal of ineffectual micro-operations. Laconic deep learning computing decomposes multiplication to the term level using Booth-encoded signed powers of two, then computes only products of non-zero terms, not dense fixed-precision MACs (Sharify et al., 2018). If

$$s_i = \max_{\mathbf{m}\in\mathcal{M}} \mathbf{v}_i^\top \mathbf{m} < \tau .$$1

then

$$s_i = \max_{\mathbf{m}\in\mathcal{M}} \mathbf{v}_i^\top \mathbf{m} < \tau .$$2

The device therefore skips term pairs that never contribute, because zero terms are never generated. The paper reports that, for typical image-classification CNNs, decomposing into bit-level products and skipping ineffectual ones can theoretically reduce work by about two orders of magnitude; for a sparse version of AlexNet, the ideal work reduction exceeds $$s_i = \max_{\mathbf{m}\in\mathcal{M}} \mathbf{v}_i^\top \mathbf{m} < \tau .$$3.

The practical implementation organizes processing around one-offset decoding, exponent addition, one-hot bucket accumulation, and a concatenation-enhanced adder tree. This design choice yields a Less In More Out effect on both bandwidth and computation. A Laconic configuration with a $$s_i = \max_{\mathbf{m}\in\mathcal{M}} \mathbf{v}_i^\top \mathbf{m} < \tau .$$4-wire weight memory interface outperforms a conventional accelerator with a $$s_i = \max_{\mathbf{m}\in\mathcal{M}} \mathbf{v}_i^\top \mathbf{m} < \tau .$$5K-wire interface by $$s_i = \max_{\mathbf{m}\in\mathcal{M}} \mathbf{v}_i^\top \mathbf{m} < \tau .$$6 on average while being $$s_i = \max_{\mathbf{m}\in\mathcal{M}} \mathbf{v}_i^\top \mathbf{m} < \tau .$$7 more energy efficient on average; a $$s_i = \max_{\mathbf{m}\in\mathcal{M}} \mathbf{v}_i^\top \mathbf{m} < \tau .$$8K-wire Laconic configuration outperforms the same baseline by $$s_i = \max_{\mathbf{m}\in\mathcal{M}} \mathbf{v}_i^\top \mathbf{m} < \tau .$$9 and is $\varepsilon = 1-\tau$0 more energy efficient. The architecture does not modify the model and does not affect accuracy. The “less in” side is thus narrower interface width and fewer effectual arithmetic events; the “more out” side is exact inference at higher throughput and better energy efficiency.

This suggests that, in hardware terms, the concept depends less on reducing nominal operation count than on exposing latent sparsity at the finest level where ineffectual work becomes visible. In Laconic that level is the term-term product rather than the scalar MAC.

4. Photonic and optical embodiments

In integrated photonics, the principle takes a geometrical rather than informational form. An ultra-compact spot size converter connects the fundamental TE mode of a $\varepsilon = 1-\tau$1 nm silicon waveguide to the fundamental mode of a $\varepsilon = 1-\tau$2 waveguide, achieving a $\varepsilon = 1-\tau$3 mode expansion over a total device length of $\varepsilon = 1-\tau$4 (Sánchez-Sánchez et al., 11 Jan 2025). The structure comprises a $\varepsilon = 1-\tau$5 transition plus a $\varepsilon = 1-\tau$6 matrix of $\varepsilon = 1-\tau$7 metamaterial cells, each modeled during optimization as a homogeneous anisotropic medium with tensor $\varepsilon = 1-\tau$8. Optimization proceeds in two stages—first 2D FDTD with homogeneous anisotropic cells, then 3D FDTD with explicit subwavelength gratings—using CMA-ES. The experimentally demonstrated device shows less than $\varepsilon = 1-\tau$9 dB insertion loss over at least $\varepsilon$0 nm bandwidth, with about $\varepsilon$1 dB simulated average loss across $\varepsilon$2–$\varepsilon$3, and a reported figure of merit

$\varepsilon$4

Here “less in” is a tightly confined mode and an ultra-short interaction length; “more out” is a much larger mode with low loss and broad bandwidth.

Optical spaceplates push the same logic into free-space propagation itself. A spaceplate of thickness $\varepsilon$5 is engineered so that its transmission phase approximates free-space propagation over an effective distance $\varepsilon$6, with compression ratio

$\varepsilon$7

The ideal spaceplate phase is

$\varepsilon$8

and the lateral beam shift obeys

$\varepsilon$9

The multilayer implementations use alternating SiO$45.6\%$0 and a-Si:H layers on fused silica, designed either as Fabry–Perot cavity stacks or by gradient descent (Hogan et al., 25 Jun 2025). The strongest reported experimental result is $45.6\%$1, which the paper describes as $45.6\%$2 times higher than any previous device; for the GD2 device the physical thickness is $45.6\%$3, angular range is about $45.6\%$4, bandwidth is $45.6\%$5 nm, and transmission at peak is about $45.6\%$6. A separate imaging test shows a focal-plane advance of $45.6\%$7 mm relative to glass alone, produced by a device of thickness $45.6\%$8. In this optical setting, “more out” means more effective propagation in less physical space.

Taken together, these photonic examples suggest that the phrase can refer either to expansion—small mode in, large mode out—or to compression—little space occupied, large optical path realized. The common ingredient is engineered nonlocality or anisotropy that reshapes the field more efficiently than adiabatic free propagation.

5. Soft robotic and biomedical embodiment

The most literal Less In More Out device in the cited material is a fluidically actuated fabric-based ventricle intended as the core module of a soft total artificial heart (Lauber et al., 9 Sep 2025). The ventricular chamber is a sealed fabric sack containing liquid and surrounded by multiple flat fabric pouches that inflate with air and compress the ventricle from the outside. In the computational idealization, the geometry is symmetric with an even number of pouches $45.6\%$9, and the structure is modeled as a geometrically nonlinear but materially linear shell with St. Venant–Kirchhoff elasticity: $8$0 The material parameters are $8$1 GPa, $8$2, $8$3 GPa, and $8$4 GPa.

The ventricular and pouch volumes are computed from surface integrals, and stroke volume is defined as

$8$5

The key performance variable is pouch number. Devices with fewer pouches deliver higher stroke volumes, larger wall motion per pouch, and higher mechanical efficiency, but also exhibit up to $8$6 higher peak von Mises stresses. For a representative physiological stroke volume of $8$7 mL, the reported mechanical efficiency

$8$8

ranges approximately $8$9–$2.7$0 across designs and afterloads. The “more out” effect is also visible in the volume ratio $2.7$1, which exceeds $2.7$2 once $2.7$3: the ventricle ejects more liquid than the net air volume that has entered the pouches because of global shape change and buckling rather than simple piston displacement.

The model further resolves strain directions, stress concentrations, and fatigue-critical regions. Fatigue analysis based on a strain-life approach identifies heat-sealed seams and buckling regions as durability-limiting features. This is therefore a canonical Less In More Out trade-off rather than a free gain: fewer pouches raise stroke volume but intensify structural loading and reduce life margins.

6. Recurrent mechanisms, trade-offs, and common misconceptions

A common misconception is that these systems advocate indiscriminate minimization. The cited works do not support that interpretation. In edge retrieval, lowering the novelty threshold too aggressively risks missing short events, and EPIC-KITCHENS requires higher $2.7$4 values, at least $2.7$5, to retain enough frames (Abdelwahab, 31 Mar 2026). In on-device SLM deployment, the winning strategy is not “remove the model,” but place the model only on the soft creative subproblem and surround it with deterministic parsing, validation, and fallbacks (Oliveira, 27 Apr 2026). In D2D offloading, less reward helps only relative to an infection-aware optimum; below that, the usual incentive logic still applies (Xu et al., 2016).

A second misconception is that the benefit always comes from raw compression. In the strongest cases, the gain is geometric or population-level. The $2.7$6-net filter improves retrieval because it removes cluster size bias in nearest-neighbor search (Abdelwahab, 31 Mar 2026). LeanFed improves learning because later-round device diversity matters more than per-round local exhaustiveness (Pereira et al., 2024). The D2D result emerges from epidemic externalities rather than local device efficiency (Xu et al., 2016). Optical spaceplates work because the angle-dependent transmission phase approximates a longer propagation operator, not because thickness is simply reduced (Hogan et al., 25 Jun 2025).

A third misconception is that “less in” implies lower capability. The evidence is more conditional. In photonics, a $2.7$7 spot size converter attains $2.7$8 mode expansion with sub-dB loss (Sánchez-Sánchez et al., 11 Jan 2025). In hardware, a $2.7$9-wire Laconic interface outperforms a $k$00K-wire dense baseline (Sharify et al., 2018). In retrieval, storing fewer frames can increase Hit@5 (Abdelwahab, 31 Mar 2026). Yet the compensating mechanisms are essential: a cross-modal adapter and cloud re-ranker in retrieval, defensive parsing and deterministic fallbacks in mobile SLMs, or anisotropic metamaterials and resonant multilayers in optics.

The core limitation is that every instantiation introduces a new constraint surface. Higher optical compression ratios narrow bandwidth and angular range (Hogan et al., 25 Jun 2025). Fewer pouches in a soft ventricle raise peak stress and fatigue risk (Lauber et al., 9 Sep 2025). Greater reduction of local FL workload may slow local convergence even as it improves system behavior (Pereira et al., 2024). In the most general sense, these studies suggest that a Less In More Out device is successful only when “less in” is applied to a quantity that is redundant, crowding, destabilizing, or otherwise misaligned with the objective metric. Where the reduced quantity is genuinely load-bearing, the same principle fails.

The resulting design doctrine is precise rather than aphoristic: remove near-duplicates, not informative exemplars; reduce per-round workload, not participation horizon; shrink output schema, not user-visible guarantees; cap incentives at the epidemic threshold; eliminate ineffectual term products, not exact arithmetic; compress space by engineering phase, not by sacrificing transfer quality; reduce pouch count only within allowable stress and fatigue limits. Under those conditions, “less in” becomes a rigorous systems strategy for obtaining “more out.”