DoTCoM brings high-performance vision transformers to mobile devices via co-bias

DoTCoM: The Lightweight Vision Transformer Built to Bring High-Accuracy AI to Every Smartphone

Vision transformers have become the reigning champions of large-scale artificial intelligence, matching or beating convolutional networks on nearly every major benchmark in computer vision. But inside the device sitting in your pocket, that revolution has largely stalled. Transformers are hungry — for parameters, for memory, for every millijoule of battery life — and the mobile silicon they must run on is unforgiving. A team of South Korean researchers now reports a credible way through the deadlock. In a study published on 29 August 2026 in the journal Multimedia Tools and Applications, YoungWook Kwon, WanSoo Kim, SuMin Oh and HyunJin Kim of Dankook University, with WanSoo Kim also affiliated with the R&D team of LIG Defense & Aerospace, describe an architecture called DoTCoM — short for Distribution optimized Transformer with Co-bias for Mobile environment. On the standard ImageNet-1K classification benchmark, their smallest variant reached 74.0 percent Top-1 accuracy with just 1.7 million parameters, while the largest hit 81.6 percent with 11.1 million — figures the authors describe as significant improvements over existing lightweight vision transformer models. Just as important for everyday users, the models were profiled not only in theory but on real mobile hardware, a Samsung mobile CPU, where latency and memory pressure expose weaknesses that benchmark tables tend to hide.

The story begins with a fundamental tension in deep learning. Ever since the landmark 2020 paper “An Image is Worth 16×16 Words” introduced the Vision Transformer, or ViT, researchers have marveled at how self-attention lets a network weigh relationships between any two patches of an image, however far apart they sit. Convolutional neural networks, by contrast, see the world through small sliding windows, stacking layers so that each neuron’s receptive field — the slice of the image it can influence — grows only gradually with depth. That locality is an inductive bias: a built-in assumption about images that allows CNNs to learn efficiently even from modest amounts of data. Transformers carry no such prior, so they must discover spatial structure from scratch, which makes them data-hungry and computationally expensive. And because attention compares every token with every other token, its cost balloons as image resolution rises — precisely the condition demanded by tasks such as segmentation. On a data-center GPU with thousands of parallel cores, that cost is tolerable; on a phone, where memory is measured in gigabytes and inference budgets in milliseconds, it is not. The obvious remedy — hybrid designs that fuse convolutional stems with transformer blocks — has spawned a crowded field of lightweight contenders, including MobileViT, EdgeViTs, FastViT, EfficientViT, SHViT and RepViT.

Yet hybrids carry a hidden cost that has frustrated the field for years: convolution and attention do not speak the same statistical language. When a convolutional layer processes an image, it produces feature maps whose values follow one distribution, shaped by local weighted sums, shared filters and batch normalization. A transformer’s attention mechanism, which aggregates information globally across tokens and typically relies on layer normalization, produces features with a different distribution altogether. The DoTCoM team argues that this discrepancy creates an additional optimization challenge during joint training. In effect, gradient updates are pulled in two directions at once, as the convolutional pathway and the attention pathway settle into feature spaces with mismatched variances. Training becomes less stable, accuracy suffers, and designers often respond by adding more machinery — more parameters, more operations — which erodes the very efficiency the hybrid was meant to deliver. The consequence is a stubborn trade-off between performance and latency in mobile environments: make the model smarter and it gets slower; make it faster and it gets less accurate. DoTCoM’s central claim is that this trade-off can be attacked directly, at the level of the statistics of the features themselves.

The first of the team’s two key inventions is structural. Called the Quarter-Inverted Bottleneck, or QIB, it fuses two classic efficiency tricks from the mobile era. One ingredient is the inverted bottleneck, popularized by MobileNetV2: instead of compressing information through a narrow middle layer, as classical residual networks do, an inverted bottleneck first expands the channel count, applies cheap depthwise convolution in that expanded space, and then projects back down. It is remarkably parameter-efficient, but its receptive field remains limited, because depthwise convolutions inspect only small neighborhoods of the input. A model that cannot see enough context tends to compensate with deeper stacks, and on a phone, depth means latency. The QIB compensates by pairing the inverted bottleneck with a Quarter Groups, or QG, convolutional block, which — as its name suggests — divides channels into quarter-sized groups so that convolutional processing spreads context across branches without multiplying cost. According to the authors, this combination remedies the insufficient receptive fields of the inverted bottleneck alone while striking an optimal trade-off between parameters and performance. In plain terms, the block lets the network see far enough to build rich visual features without spending the parameter budget that would normally buy that field of view.

The second invention is subtler and arguably more consequential. Co-optimization Bias, or Co-Bias, is DoTCoM’s answer to the distribution mismatch. It inserts learnable per-channel bias parameters into the network — small trainable offset values, one per channel — that are tuned during training to reduce the variance discrepancies between convolutional and transformer feature distributions. Rather than waiting for the two branches to converge on their own, Co-Bias actively nudges each channel’s statistics toward a shared regime, aligning the feature distributions so that the convolutional and attention pathways can be optimized together instead of in tension. The reported effects are twofold: training becomes more stable, and final accuracy improves. The elegance of the idea lies in its cost. Bias parameters are among the cheapest objects in deep learning — a single floating-point value per channel — yet they sit precisely where the two statistical worlds collide. At a time when mobile AI designs often fight inefficiency with brute-force measures such as structural reparameterization or aggressive pruning, DoTCoM’s proposal is almost surgical: fix the mismatch where it lives, inside the numbers themselves.

The benchmark numbers suggest the approach works. On ImageNet-1K, the canonical million-image classification test, DoTCoM ships in four sizes. The flagship DoTCoM-L reached 81.6 percent Top-1 accuracy with 11.1 million parameters; DoTCoM-B scored 79.8 percent with 5.8 million; the compact DoTCoM-S managed 77.5 percent with just 2.8 million; and the featherweight DoTCoM-T posted 74.0 percent with a mere 1.7 million parameters. That smallest model is tiny enough to be a rounding error beside the billion-parameter giants dominating headlines, yet accurate enough for genuine recognition work. A few million parameters is exactly the territory in which mobile vision models live or die, and every fraction of a percentage point of Top-1 accuracy at that scale is hard-won. The authors state that DoTCoM achieved these results while remaining lightweight and outperforming existing lightweight ViT models, the crowded class of architectures built to bring transformer power to phones. Crucially, the evaluation did not rest solely on FLOPs — floating-point operation counts that can flatter a design in the laboratory. The team also measured the models on an actual Samsung mobile CPU, where memory bandwidth, cache behavior and scheduling quirks can humble even elegant mathematics.

Classification, however, is only the entrance exam. The true test of a backbone is whether its learned features transfer to the denser, more demanding tasks that real applications require: locating objects and outlining their shapes. The team plugged DoTCoM into standard pipelines for both. On MS-COCO, the sprawling object-detection benchmark filled with cluttered everyday scenes, DoTCoM achieved a mean average precision of 28.9, a solid figure at mobile-friendly scale, where every detector must balance accuracy against the strict latency budgets of on-device inference. On PASCAL VOC 2012, the semantic segmentation benchmark that asks a network to assign a class to every pixel, DoTCoM reached a mean intersection over union of 83.2, meaning its predicted regions overlapped the ground-truth labels to a high degree. Detection and segmentation also demand higher-resolution feature maps, which is where attention’s cost spiral usually begins — making a mobile model’s success there a sterner test than classification alone. The authors summarize the pattern plainly: the results indicate that DoTCoM transfers consistently to standard dense-prediction downstream tasks at mobile-deployable cost. That consistency matters, because a backbone that excels only at classification is a laboratory curiosity, while one that holds its accuracy across detection and segmentation on phone-class hardware is a candidate for production systems.

Why does this matter beyond the leaderboard? Because the future of AI on phones is being decided now, and the terms of that debate are set by architectures like this one. Running vision models directly on a device rather than shipping pixels to a cloud server slashes latency, protects privacy and keeps features alive when connectivity fails — imperatives that have driven handset makers and chip designers to invest heavily in mobile-first neural networks. Hybrid CNN-transformer backbones are the workhorses of that effort, underpinning everything from computational photography to real-time augmented reality, and every improvement in the accuracy-per-parameter curve ripples outward into battery life and the features a phone can support. The work was funded by a grant from South Korea’s Institute of Information & Communications Technology Planning & Evaluation under the ICT Challenge and Advanced Network of HRD program, financed by the Ministry of Science and ICT, and the author list reflects a collaboration between Dankook University’s Department of Electronics and Electrical Engineering and an industry R&D team. The paper itself took the long road: received in October 2024, revised in July 2026, accepted that August, and published on 29 August 2026 as volume 85, article 724 of the journal, with YoungWook Kwon and WanSoo Kim credited as equal first authors.

None of this settles the contest for mobile vision. DoTCoM’s numbers were earned against a specific roster of rivals, on specific hardware, under specific training conditions, and independent replication is the customary next step for any architecture claiming a new efficiency frontier. The field’s pace is ferocious: the paper’s reference list alone reads like a genealogy of half a decade of efficient design, running from MobileNet and ShuffleNet through EfficientNet, MobileViT, EdgeNeXt, FastViT, MobileOne and RepViT, each generation shaving milliseconds and megabytes from the last. What DoTCoM adds to that lineage is a reminder that some of the thorniest problems in hybrid AI are statistical rather than architectural — that friction between two ways of seeing an image can be smoothed not by stacking layers but by aligning distributions. If the approach survives scrutiny and adoption, the practical payoff could be transformer-grade image understanding running quietly inside ordinary handsets: cameras that grasp scenes more deeply and assistants that see more accurately, all without a single byte leaving the device. For now, a compact Korean-built model has made the case that the road to smarter phones may run through better statistics, not simply bigger networks.