UniScale: Unified Scaling Across Domains

• UniScale is a unified design framework that integrates scale as a learnable variable, turning decoupled optimization problems into a single formulation.
• It is applied in diverse domains such as LLM inference, 3D reconstruction for robotic perception, and industrial search ranking to jointly optimize performance and resource efficiency.
• The framework’s unified approach streamlines model routing, geometric prior injection, and data-model co-design, enabling improved empirical outcomes and cost management.

UniScale is a name used in recent arXiv literature for several frameworks that elevate scale from a background hyperparameter to an explicit design variable. In current usage, it denotes at least three distinct 2026 systems: an online framework for jointly optimizing model routing and test-time scaling in large-language-model inference, a unified scale-aware multi-view 3D reconstruction model for robotic perception, and a co-design framework that scales both data and architecture for industrial search ranking. Related work uses the term more informally for unified multivariate scaling-law formalisms and scale-ordered generative processes, but these are separate research programs rather than a single shared method (Huang et al., 29 May 2026, Mahdavian et al., 26 Feb 2026, Yu et al., 25 Mar 2026, Caballero et al., 25 May 2026, Chen et al., 25 May 2026).

1. Scope and principal usages

Three papers use UniScale as the formal system name, each in a different subfield.

Work	Domain	Core unification
"UniScale: Adaptive Unified Inference Scaling via Online Joint Optimization of Model Routing and Test-Time Scaling"	LLM inference	Joint action space over model choice and TTS
"UniScale: Unified Scale-Aware 3D Reconstruction for Multi-View Understanding via Prior Injection for Robotic Perception"	Robotic vision	Single feed-forward model with optional geometric priors and scene metric scale
"UniScale: Synergistic Entire Space Data and Model Scaling for Search Ranking"	Industrial search ranking	Co-design of entire-space data scaling and heterogeneous ranking architecture

The three systems are not variants of one codebase or one theory. Their commonality lies in a repeated design pattern: each replaces a previously decoupled optimization problem with a unified formulation in which scale becomes part of the learned state, the model interface, or the training system (Huang et al., 29 May 2026, Mahdavian et al., 26 Feb 2026, Yu et al., 25 Mar 2026).

2. UniScale in adaptive LLM inference

In LLM serving, UniScale is built on Unified Inference Scaling (UIS), a formulation that merges model routing and test-time scaling into one decision space. Each action is a joint configuration

$a=(M,QP,CP,BS),$

where $M$ is the base model, $QP$ is question parallelism, $CP$ is candidate parallelism, and $BS$ is beam size. The framework casts adaptive UIS as a contextual multi-armed bandit and learns policies with LinUCB. The reward is a weighted combination of correctness, verifier score, and normalized cost,

$$r_t = w_1 \cdot \mathrm{Correct}(a_t) + w_2 \cdot \mathrm{Score}(a_t) + w_3 \cdot (1-\tilde{C}_{\mathrm{UIS}}(a_t)),$$

and the action selection rule is

$$a_t=\arg\max_{a\in\mathcal A}\left(\mathbf{x}_{t,a}^\top \hat{\boldsymbol{\theta}}_t+\alpha\sqrt{\mathbf{x}_{t,a}^\top \mathbf{A}_t^{-1}\mathbf{x}_{t,a}}\right).$$

The cost model is hardware-aware and uses equivalent FLOPs,

$$\mathrm{eFLOPs}=C_{\mathrm{comp}}+C_{\mathrm{mem}}\cdot I.$$

The reported action space includes Qwen3-0.6B, 1.7B, 4B, 8B, 14B, and 32B, combined with verifier-guided search choices over $QP$, $CP$, and $M$0 (Huang et al., 29 May 2026).

The empirical claim is that joint optimization across this UIS space is stronger than routing-only or TTS-only control. In the full UIS scenario, UniScale obtains reward $M$1 in the Cost-Sensitive setting, compared with the next-best baseline k-NN at $M$2, while using cost $M$3 Tera-eFLOPs and achieving accuracy $M$4. In the Quality-Priority setting it obtains reward $M$5, compared with Thompson Sampling at $M$6, and reaches the highest non-oracle accuracy, $M$7, at cost $M$8. The framework also includes path-aware early exiting; averaged across UIS configurations, this reduces total cost by $M$9 while changing accuracy from $QP$0 to $QP$1. The main limitations stated in the paper are dependence on verifier precision, action-space discretization, and a linear contextual reward model (Huang et al., 29 May 2026).

3. UniScale in robotic multi-view 3D reconstruction

In robotic perception, UniScale is a feed-forward, transformer-based multi-view 3D reconstruction framework that jointly predicts camera intrinsics and extrinsics, scale-invariant depth maps, scale-invariant point maps, and a scene-level metric scale. The model interface is

$QP$2

with optional priors $QP$3 and $QP$4. The camera output is parameterized as

$QP$5

Its main architectural addition over VGGT is a dedicated scale head,

$QP$6

implemented through token fusion and a positive per-frame prediction,

$QP$7

The design is explicitly semantically informed: pose priors are injected into camera tokens and the scale head, while intrinsics are converted into an origin-free ray image and injected into patch tokens and the scale head. The model is initialized from DINOv2 and VGGT and does not require training from scratch (Mahdavian et al., 26 Feb 2026).

The reported benchmark results emphasize image-only metric prediction and prior-aware improvement. On Robust-MVD in the image-only metric setting, UniScale records KITTI rel $QP$8, $QP$9 $CP$0, versus MapAnything rel $CP$1, $CP$2 $CP$3, and ScanNet rel $CP$4, $CP$5 $CP$6, versus MapAnything rel $CP$7, $CP$8 $CP$9. In the aligned setting with images plus intrinsics, it reports KITTI $BS$0 rel / $BS$1 $BS$2 versus MapAnything $BS$3 / $BS$4, and ScanNet $BS$5 / $BS$6 versus $BS$7 / $BS$8. The dense-$BS$9-view evaluation is described as showing especially strong depth quality. The paper simultaneously notes that it does not provide runtime, FLOPs, or memory benchmarks, and that the system remains a large transformer-based model whose scale prediction is a single scene-level scalar (Mahdavian et al., 26 Feb 2026).

4. UniScale in industrial search ranking

For e-commerce search, UniScale is a joint data-model scaling framework deployed on Taobao Search. Its ranking problem is standard binary feedback prediction with

$$r_t = w_1 \cdot \mathrm{Correct}(a_t) + w_2 \cdot \mathrm{Score}(a_t) + w_3 \cdot (1-\tilde{C}_{\mathrm{UIS}}(a_t)),$$0

and target $$r_t = w_1 \cdot \mathrm{Correct}(a_t) + w_2 \cdot \mathrm{Score}(a_t) + w_3 \cdot (1-\tilde{C}_{\mathrm{UIS}}(a_t)),$$1. The framework combines ES$$r_t = w_1 \cdot \mathrm{Correct}(a_t) + w_2 \cdot \mathrm{Score}(a_t) + w_3 \cdot (1-\tilde{C}_{\mathrm{UIS}}(a_t)),$$2, the Entire-Space Sample System, with HHSFT, the Heterogeneous Hierarchical Sample Fusion Transformer. ES$$r_t = w_1 \cdot \mathrm{Correct}(a_t) + w_2 \cdot \mathrm{Score}(a_t) + w_3 \cdot (1-\tilde{C}_{\mathrm{UIS}}(a_t)),$$3 expands training beyond exposed search samples through intra-domain unexposed candidate expansion, hierarchical label attribution, and cross-domain sample searchification. HHSFT is then used to model the heterogeneous distribution induced by this enlarged sample space (Yu et al., 25 Mar 2026).

The architecture has two stages. In Heterogeneous Hierarchical Feature Interaction, token-specific QKV projections are used:

$$r_t = w_1 \cdot \mathrm{Correct}(a_t) + w_2 \cdot \mathrm{Score}(a_t) + w_3 \cdot (1-\tilde{C}_{\mathrm{UIS}}(a_t)),$$4

followed by token-specific feed-forward updates,

$$r_t = w_1 \cdot \mathrm{Correct}(a_t) + w_2 \cdot \mathrm{Score}(a_t) + w_3 \cdot (1-\tilde{C}_{\mathrm{UIS}}(a_t)),$$5

A subsequent Global Feature Attention stage uses composite projections on the concatenation of token outputs. The Entire Space User Interest Fusion stage then applies Domain-Routed Expert Fusion,

$$r_t = w_1 \cdot \mathrm{Correct}(a_t) + w_2 \cdot \mathrm{Score}(a_t) + w_3 \cdot (1-\tilde{C}_{\mathrm{UIS}}(a_t)),$$6

and Domain-Aware Personalized Gated Attention, beginning with

$$r_t = w_1 \cdot \mathrm{Correct}(a_t) + w_2 \cdot \mathrm{Score}(a_t) + w_3 \cdot (1-\tilde{C}_{\mathrm{UIS}}(a_t)),$$7

then

$$r_t = w_1 \cdot \mathrm{Correct}(a_t) + w_2 \cdot \mathrm{Score}(a_t) + w_3 \cdot (1-\tilde{C}_{\mathrm{UIS}}(a_t)),$$8

and a personalized gate

$$r_t = w_1 \cdot \mathrm{Correct}(a_t) + w_2 \cdot \mathrm{Score}(a_t) + w_3 \cdot (1-\tilde{C}_{\mathrm{UIS}}(a_t)),$$9

Training uses multi-domain binary cross-entropy over domain-specific outputs (Yu et al., 25 Mar 2026).

The data-scaling effect of ES$$a_t=\arg\max_{a\in\mathcal A}\left(\mathbf{x}_{t,a}^\top \hat{\boldsymbol{\theta}}_t+\alpha\sqrt{\mathbf{x}_{t,a}^\top \mathbf{A}_t^{-1}\mathbf{x}_{t,a}}\right).$$0 is quantified directly. Relative to the search-exposed-only baseline, unexposed expansion increases samples to $$a_t=\arg\max_{a\in\mathcal A}\left(\mathbf{x}_{t,a}^\top \hat{\boldsymbol{\theta}}_t+\alpha\sqrt{\mathbf{x}_{t,a}^\top \mathbf{A}_t^{-1}\mathbf{x}_{t,a}}\right).$$1 while keeping requests at $$a_t=\arg\max_{a\in\mathcal A}\left(\mathbf{x}_{t,a}^\top \hat{\boldsymbol{\theta}}_t+\alpha\sqrt{\mathbf{x}_{t,a}^\top \mathbf{A}_t^{-1}\mathbf{x}_{t,a}}\right).$$2; hierarchical label attribution keeps requests and samples at $$a_t=\arg\max_{a\in\mathcal A}\left(\mathbf{x}_{t,a}^\top \hat{\boldsymbol{\theta}}_t+\alpha\sqrt{\mathbf{x}_{t,a}^\top \mathbf{A}_t^{-1}\mathbf{x}_{t,a}}\right).$$3 and $$a_t=\arg\max_{a\in\mathcal A}\left(\mathbf{x}_{t,a}^\top \hat{\boldsymbol{\theta}}_t+\alpha\sqrt{\mathbf{x}_{t,a}^\top \mathbf{A}_t^{-1}\mathbf{x}_{t,a}}\right).$$4 respectively but doubles click positives to $$a_t=\arg\max_{a\in\mathcal A}\left(\mathbf{x}_{t,a}^\top \hat{\boldsymbol{\theta}}_t+\alpha\sqrt{\mathbf{x}_{t,a}^\top \mathbf{A}_t^{-1}\mathbf{x}_{t,a}}\right).$$5; adding cross-domain searchification yields $$a_t=\arg\max_{a\in\mathcal A}\left(\mathbf{x}_{t,a}^\top \hat{\boldsymbol{\theta}}_t+\alpha\sqrt{\mathbf{x}_{t,a}^\top \mathbf{A}_t^{-1}\mathbf{x}_{t,a}}\right).$$6 requests, $$a_t=\arg\max_{a\in\mathcal A}\left(\mathbf{x}_{t,a}^\top \hat{\boldsymbol{\theta}}_t+\alpha\sqrt{\mathbf{x}_{t,a}^\top \mathbf{A}_t^{-1}\mathbf{x}_{t,a}}\right).$$7 samples, and $$a_t=\arg\max_{a\in\mathcal A}\left(\mathbf{x}_{t,a}^\top \hat{\boldsymbol{\theta}}_t+\alpha\sqrt{\mathbf{x}_{t,a}^\top \mathbf{A}_t^{-1}\mathbf{x}_{t,a}}\right).$$8 click positives. Offline, HHSFT alone reaches $$a_t=\arg\max_{a\in\mathcal A}\left(\mathbf{x}_{t,a}^\top \hat{\boldsymbol{\theta}}_t+\alpha\sqrt{\mathbf{x}_{t,a}^\top \mathbf{A}_t^{-1}\mathbf{x}_{t,a}}\right).$$9 AUC and $$\mathrm{eFLOPs}=C_{\mathrm{comp}}+C_{\mathrm{mem}}\cdot I.$$0 GAUC over the DLRM-MLP base, and HHSFT+ES$$\mathrm{eFLOPs}=C_{\mathrm{comp}}+C_{\mathrm{mem}}\cdot I.$$1 reaches $$\mathrm{eFLOPs}=C_{\mathrm{comp}}+C_{\mathrm{mem}}\cdot I.$$2 AUC and $$\mathrm{eFLOPs}=C_{\mathrm{comp}}+C_{\mathrm{mem}}\cdot I.$$3 GAUC. The synergy becomes stronger at larger model size: the gap between HHSFT+Search and HHSFT+ES$$\mathrm{eFLOPs}=C_{\mathrm{comp}}+C_{\mathrm{mem}}\cdot I.$$4 grows from $$\mathrm{eFLOPs}=C_{\mathrm{comp}}+C_{\mathrm{mem}}\cdot I.$$5 $$\mathrm{eFLOPs}=C_{\mathrm{comp}}+C_{\mathrm{mem}}\cdot I.$$6AUC at $$\mathrm{eFLOPs}=C_{\mathrm{comp}}+C_{\mathrm{mem}}\cdot I.$$7 to $$\mathrm{eFLOPs}=C_{\mathrm{comp}}+C_{\mathrm{mem}}\cdot I.$$8 at $$\mathrm{eFLOPs}=C_{\mathrm{comp}}+C_{\mathrm{mem}}\cdot I.$$9. The ablations also show that naive heterogeneous-data mixing is harmful: with ES$QP$0, a plain HHFI backbone without ESUIF produces $QP$1 AUC, $QP$2 GAUC, and $QP$3 HR@5, while adding DREF and DAPGA recovers performance to $QP$4, $QP$5, and $QP$6. In a 10-day online A/B test on 5% traffic, the system yields $QP$7 purchase and $QP$8 GMV, while the engineering optimizations are reported to reduce GPU inference costs by approximately $QP$9 and training overhead by approximately $CP$0 (Yu et al., 25 Mar 2026).

5. Related unified-scaling frameworks discussed alongside UniScale

Although not formally named UniScale, several nearby works use the term as an informal shorthand for broader unification strategies. The paper "Unified Neural Scaling Laws" explicitly states that there is no separate formal model named UniScale in that work; the actual term is UNSL, a multivariate functional form for performance as model parameters, training dataset size, training steps, inference steps, compute, and hyperparameters vary simultaneously. Its reported extrapolation win rates are $CP$1 on downstream image classification and $CP$2 on language tasks (Caballero et al., 25 May 2026).

"Everything at Every Scale: Scale-Invariant Diffusion with Continuous Super-Resolution" introduces SKILD, a DCT/k-space diffusion model in which generation and continuous super-resolution are treated as the same reverse process started from different timesteps. The paper is explicit that the same learned reverse dynamics performs both tasks with “no task-specific architecture, no conditioning branch, no classifier-free guidance, no retraining per scale factor.” Its unconditional CIFAR-10 result is FID $CP$3 and Inception Score $CP$4, and the same checkpoint supports $CP$5–$CP$6 ImageNet super-resolution (Chen et al., 25 May 2026).

In graph learning, "Cross-Domain Graph Data Scaling: A Showcase with Diffusion Models" presents UniAug, described as a universal graph structure augmentor built from a self-conditioned discrete diffusion model pre-trained on thousands of graphs across domains. The paper presents it as the first demonstration of a data-scaling graph structure augmentor on graphs across domains (2406.01899).

For model-family growth, "$CP$7pscaling small models: Principled warm starts and hyperparameter transfer" studies width upscaling with exact finite-width equivalence and extends $CP$8Transfer to upscaled training, again making scale transition part of the formal method rather than an ad hoc heuristic (Ma et al., 11 Feb 2026).

6. Conceptual commonalities and outstanding limits

Across these works, the repeated move is to replace a pipeline of separate decisions with a single coordinated scaling space. In LLM serving, UniScale merges routing and TTS into one contextual bandit action. In robotic 3D reconstruction, UniScale merges normalized geometry, metric-scale recovery, and optional geometric priors into one feed-forward model. In search ranking, UniScale merges entire-space data construction with a heterogeneous Transformer explicitly designed for that data. This suggests a common research attitude: scale is not treated merely as model size, but as an operational variable spanning compute, supervision, geometry, or candidate space (Huang et al., 29 May 2026, Mahdavian et al., 26 Feb 2026, Yu et al., 25 Mar 2026).

The same comparison also makes the term’s limits clear. In current arXiv usage, UniScale does not denote a single cross-domain formalism. The LLM system is an online contextual bandit with verifier-aware reward shaping; the robotic system is a metric multi-view reconstruction model with a scene-level scale head; the search system is a data-engineering and ranking-architecture co-design. Their technical objects, losses, and guarantees are unrelated. The neighboring usages are similarly heterogeneous: UNSL is a multivariate neural scaling law, SKILD is a scale-ordered diffusion process, UniAug is a graph-structure augmentor, and $CP$9pscaling is a principled width-upscaling theory (Caballero et al., 25 May 2026, Chen et al., 25 May 2026, 2406.01899, Ma et al., 11 Feb 2026).

The open issues are correspondingly domain-specific. The LLM inference framework depends on verifier quality and a discretized action set; the robotic framework omits runtime and memory benchmarks and predicts a single scene-level scale; the search framework leaves several ES$M$00 steps procedural rather than closed-form and is validated on large-scale e-commerce search rather than on recommendation or advertising deployments. What unifies them is therefore not a shared implementation, but a shared design principle: scale is internalized into the model or training system, and unification is used to turn previously separate knobs into a single learnable structure (Huang et al., 29 May 2026, Mahdavian et al., 26 Feb 2026, Yu et al., 25 Mar 2026).