Paper deep dive
StableTTA: Training-Free Test-Time Adaptation that Improves Model Accuracy on ImageNet1K to 96%
Zheng Li, Jerry Cheng, Huanying Helen Gu
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 4/10/2026, 2:52:02 AM
Summary
StableTTA is a training-free test-time adaptation method designed to improve the stability and efficiency of ensemble-based image classification. By identifying and mitigating conflicts in aggregation strategies (hard voting, soft voting, logit averaging) caused by logit variance, the authors introduce specialized data augmentation (modified Mixup and CutMix) and a Non-Significant Suppression (NSS) logit processing technique. The method enables lightweight models like MobileNetV3 to outperform larger architectures on ImageNet-1K, achieving over 95% accuracy with significantly reduced memory and computational costs.
Entities (5)
Relation Signals (3)
StableTTA → improvesaccuracyon → ImageNet-1K
confidence 100% · Empirical results on ImageNet-1K show gains of 10.93--32.82% in top-1 accuracy
Non-Significant Suppression → componentof → StableTTA
confidence 95% · In this work, we apply NSS to logits averaging as a processing method
StableTTA → reducescomputationalcostfor → MobileNetV3
confidence 90% · StableTTA allows lightweight architectures to outperform ViT... reducing computational cost by approximately 89.1%
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Ensemble methods are widely used to improve predictive performance, but their effectiveness often comes at the cost of increased memory usage and computational complexity. In this paper, we identify a conflict in aggregation strategies that negatively impacts prediction stability. We propose StableTTA, a training-free method to improve aggregation stability and efficiency. Empirical results on ImageNet-1K show gains of 10.93--32.82\% in top-1 accuracy, with 33 models achieving over 95\% accuracy and several surpassing 96\%. Notably, StableTTA allows lightweight architectures to outperform ViT by 11.75\% in top-1 accuracy while using less than 5\% of parameters and reducing computational cost by approximately 89.1\% (in GFLOPs), enabling high-accuracy inference on resource-constrained devices.
Tags
Links
- Source: https://arxiv.org/abs/2604.04552v1
- Canonical: https://arxiv.org/abs/2604.04552v1
Trouble viewing inline? Open PDF directly →
Full Text
53,870 characters extracted from source content.
Expand or collapse full text
StableTTA: Training-Free Test-Time Adaptation that Improves Model Accuracy on ImageNet1K to 96% Zheng Li Department of Computer Science New York Institute of Technology New York, NY 10023 zli66@nyit.edu &Jerry Cheng Department of Computer Science New York Institute of Technology New York, NY 10023 jcheng18@nyit.edu Huanying Helen Gu Department of Computer Science New York Institute of Technology New York, NY 10023 hgu03@nyit.edu Abstract Ensemble methods are widely used to improve predictive performance, but their effectiveness often comes at the cost of increased memory usage and computational complexity. In this paper, we identify a conflict in aggregation strategies that negatively impacts prediction stability. We propose StableTTA, a training-free method to improve aggregation stability and efficiency. Empirical results on ImageNet-1K show gains of 10.93–32.82% in top-1 accuracy, with 33 models achieving over 95% accuracy and several surpassing 96%. Notably, StableTTA allows lightweight architectures to outperform ViT by 11.75% in top-1 accuracy while using less than 5% of parameters and reducing computational cost by approximately 89.1% (in GFLOPs), enabling high-accuracy inference on resource-constrained devices. Code is available at: https://github.com/LizhengMathAi/StableTTA , including a 3-minute reproduction demo. 1 Introduction Figure 1: Top: Milestone comparison. We show that StableTTA+MobileNetV3 significantly outperforms the base ViT in terms of accuracy (+11.75% top-1), memory usage (-97% parameters), and computational cost (-89.1% GFLOPs). Bottom: General comparison. StableTTA improves baseline models by 11-33% in top-1 accuracy, with 34 models achieving more than 95% accuracy. The proposed method yields consistent and significant improvements across all evaluated models. Ensemble methods have long been proposed to improve the predictive quality of machine learning models [6]. By aggregating diverse predictions, they typically yield more accurate and robust outputs [6, 1]. These methods are usually training-free, meaning they do not require additional model training or extensive tuning of training recipes [1]. As a result, they are widely adopted by non-professional users seeking straightforward performance improvements. However, despite their simplicity, ensemble methods are rarely used in highly challenging, industrial-scale applications. Instead, researchers tend to focus on designing new model architectures, as these approaches often lead to better performance, fewer parameters, and lower computational costs during inference [4]. A key limitation of ensemble methods lies in their reliance on multiple models acting as “experts” to generate predictions for aggregation. This design leads to a linear increase in both memory usage and computational cost, as each additional model contributes to the total resource demand. In practice, using 10 models may result in less than a 3% improvement in accuracy, while increasing memory and inference cost by nearly 10×. Test-time augmentation (TTA) has been explored as a way to mitigate this trade-off by using a single model with multiple augmented inputs. However, this approach still scales linearly in computational cost with respect to the number of augmentations [7, 10] and does not fundamentally resolve the efficiency issue. In this work, we identify and investigate an inherent conflict between different aggregation strategies. Based on this observation, we propose Training-Free Test-Time Adaptation (StableTTA), a hybrid method that combines image preprocessing with logit post-processing. Remarkably, this approach significantly improves the performance of popular vision models. Our experiments show gains of 10.93–32.82% in top-1 accuracy on the ImageNet-1K classification task. After applying StableTTA, 33 models achieve over 95% accuracy, and 4 models surpass 96%. Historically, improvements in vision models on ImageNet-1K [3] have been incremental, typically yielding only 1–2% gains in top-1 accuracy per advancement [13, 28, 7, 32, 4], with accuracies above 90% [4, 17] requiring additional training data such as ImageNet-21K [25] and JFT-300M [20]. In contrast, without extra data, our method enables nearly all commonly used models to exceed 90% accuracy, with most surpassing 95%. More importantly, StableTTA addresses the core weakness of traditional ensemble methods by significantly improving memory efficiency and reducing computational cost. This makes it particularly suitable for mobile and resource-constrained environments. For example, our experiments show that a lightweight model such as MobileNetV3, when enhanced with StableTTA, significantly outperforms the Vision Transformer, achieving +11.75% higher top-1 accuracy while reducing memory usage by 97% and computational cost by 89.1% (measured in GFLOPs). These results demonstrate that StableTTA enables high-accuracy inference even on low-resource devices, outperforming larger models without requiring additional training. 2 Related Work For image classification, model averaging is a widely used ensemble strategy to improve generalization performance [6]. As introduce by Breiman [1], outputs of multiple independently trained models are typically aggregated by majority voting on predictions or by averaging predicted probabilities. Generally, model averaging is not preferable to training a stronger, well-optimized single model [7]. Although it can provide modest performance gains, these improvements are often marginal relative to the substantial increase in resource requirements [7, 32]. Specifically, the total number of model parameters/FLOPs equals the sum of each model’s parameters/FLOPs. Test-time augmentation (TTA) offers a partial alternative that avoids the need for multiple models. TTA applies a set of transformations to the input image (e.g., flipping, cropping) and feeds the augmented images into the same model to obtain multiple logits for averaging [13, 7]. This approach preserves the model size. However, it still increases the inference-time computation cost. For instance, ResNet [7] and DenseNet [10] employ 10-crop evaluation during validation. This increased the total FLOPs ten times. Despite these efforts, TTA typically yields only a modest improvement of no more than 2% in top-1 accuracy. [27, 12, 14]. Although TTA has been widely studied, only a few works gain attention in top-tier research, mainly because its key limitation “the significant increase in inference computation” remains unsolved. 3 Background For a better understanding of the most commonly used ensemble methods, consider a C-class classification task. Let the logits z denote the output vector of a model f with parameters θ for input x: =f(;)∈ℝC. z=f( x; θ) ^C. After applying the softmax function, we obtain the probability distribution (y∣) p_ θ(y x) over the C classes: (y=k∣)=exp(k)∑j=1Cexp(j),k=1,…,C. p_ θ(y=k x)= ( z_k) _j=1^C ( z_j), k=1,…,C. The predicted class is given by y^=argmaxkk=argmaxk(y=k∣). y= *arg\,max_k z_k= *arg\,max_k p_ θ(y=k x). Majority voting, probability averaging, and logit averaging are commonly used aggregation strategies for ensemble methods, also known as hard voting, soft voting, and logit averaging, respectively [6, 22]. Consider an ensemble of N experts, the formal mathematical formulations are given as follows: • Hard Voting (Fig. 2(b), orange blocks): Each model casts a vote for a class, and the class receiving the majority of votes is selected. (Hard Voting)y^hard=argmaxk∑i=1Ny^(i)=k(Hard Voting) y_hard= *arg\,max_k _i=1^N 1_ y^(i)=k • Soft Voting (Fig. 2(b), cyan blocks): The predicted class probabilities are averaged, and the class with the highest mean probability is selected. (Soft Voting)y^soft=argmaxk1N∑i=1N(i)(y=k∣)(Soft Voting) y_soft= *arg\,max_k 1N _i=1^N p^(i)(y=k x) • Logit Averaging (Fig. 2(b), light blue blocks): The logits are averaged across models, and the class with the highest mean logit is selected. (Logit Averaging)y^logit=argmaxk1N∑i=1Nk(i)(Logit Averaging) y_logit= *arg\,max_k 1N _i=1^N z_k^(i) Figure 2: Limitations of Ensemble Methods and Conflicts in Aggregation Strategies. (a) Limitations: Illustration of model averaging and test-time augmentation (TTA) in image classification. Both methods significantly increase inference-time computational cost while providing only marginal improvements in accuracy. (b) Conflict: Given branch logits (i)∣i=1,2,3\ z^(i) i=1,2,3\, probabilities (i)=softmax((i)) p^(i)=softmax( z^(i)), and predictions y^(i)=argmax(i) y^(i)=argmax p^(i), different aggregation strategies can produce inconsistent results. In this example, logit averaging predicts class y^logit=1 y_logit=1, soft voting predicts class y^soft=2 y_soft=2, and hard voting predicts class y^hard=3 y_hard=3. (c) Explanation: When the logits ((1),(2),…)( z^(1), z^(2),…) are sparsely distributed in space, the conflict y^logit≠y^soft≠y^hard y_logit≠ y_soft≠ y_hard is more likely to occur due to the nonlinearity and non-bijective nature of the softmax function. In contrast, when the logits are more densely clustered, such conflicts are less likely. Limitations Fig. 2(a) illustrates the pipelines of model averaging and test-time augmentation (TTA) introduced in Section 2. Model averaging requires storing multiple models, and its inference-time computational cost is the sum of the costs of all models. In contrast, TTA uses a single model but increases the inference-time computational complexity by N (typically ranging from 5 to 64). Conflict In ensemble methods, different aggregation strategies may conflict with one another. Fig. 2(b) illustrates such an example: given three logit vectors ((1),(2),(3))( z^(1), z^(2), z^(3)), the predictions obtained from hard voting y^hard=3 y_hard=3, soft voting y^soft=2 y_soft=2, and logit averaging y^logit=1 y_logit=1 are inconsistent. This discrepancy arises because both the softmax and indicator functions are nonlinear, leading to inconsistencies across aggregation strategies. This mismatch is a key factor contributing to the modest improvements typically observed in classification accuracy with ensemble methods. Interpretation of the Conflict As shown in Fig. 2(c), these conflicts arise from the distribution of logits in the feature space. When the logits (1),(2),…\ z^(1), z^(2),…\ are sparsely distributed, the nonlinearity and non-bijective nature of the softmax and indicator functions can amplify their differences. In contrast, when the logits are more densely clustered, different aggregation strategies are more likely to yield consistent predictions. We provide both theoretical analysis and empirical evidence to support this observation. In Appendix A, we further validate this phenomenon using Monte Carlo simulation (see Fig. 5), which show that as the variance Var()Var( z) increases, the probability of prediction conflicts also increases. The empirical results are well fitted by our derived relationship between P(y^logit≠y^hard)P( y_logit≠ y_hard) and Var()Var( z). In summary, lower variance leads to fewer conflicts. This observation inspires us to reformulate the test-time augmentation method and to design a logit-processing method to achieve more stable prediction aggregation. 4 Method In this section, we aim to reduce Var()Var( z) through two approaches: (1) redesigning data augmentation strategies for TTA, and (2) introducing logit processing to further minimize logit variance. 4.1 Data Augmentation Principles for StableTTA Since neural network layers are continuous and differentiable, most neural networks satisfy the Hölder condition [35]: ‖−′‖=‖f(;)−f(′;)‖≤c⋅‖−′‖d\| z- z \|=\|f( x; θ)-f( x ; θ)\|≤ c·\| x- x \|^d where c and d are constants. This inequality indicates that the distance ‖−′‖\| z- z \| between logits is determined by the distance ‖−′‖\| x- x \| between inputs. Now, consider a sequence of data augmentations ψ1,ψ2,…\ _1, _2,…\ and the corresponding logit vector (i):=f(ψi();) z^(i):=f( _i( x); θ). By applying the Hölder condition, we obtain: ‖(i)−(j)‖=‖f(ψi();)−f(ψj();)‖≤c‖ψi()−ψj()‖d.\| z^(i)- z^(j)\|=\|f( _i( x); θ)-f( _j( x); θ)\|≤ c\| _i( x)- _j( x)\|^d. (1) Furthermore, when d∈(0,1]d∈(0,1], we have: Var()≤2d−1c2Var(ψ()).Var( z)≤ 2^d-1c^2Var(ψ( x)). To facilitate understanding Eq. (1), in Appendix B, we provide Fig. 6 for visualization, along with empirical results as supporting evidence. Now, reviewing the conflict discussed in Section 3, we draw the following conclusions: • Common augmentation techniques such as horizontal flipping and random cropping are not necessary in TTA, as the model is not sensitive to them. Pre-trained models already account for these transformations; thus, the logit vectors are typically identical after such augmentations. • Common augmentation techniques such as translation, rotation, and affine transformations [2] should be avoided in TTA, as they can induce an excessively large increase in Var‖ψ()‖Var\|ψ( x)\|. • Although random erasing yields relatively small values of Var(ψ())Var(ψ( x)), it provides limited diversity among augmented samples and is therefore excluded. Under these considerations, mixup [39] and CutMix [37] emerge as promising candidates. However, both require further modification to reduce Var(ψ())Var(ψ( x)): • Mixup: Instead of combining the input with a random image, we perform a weighted combination with a fixed image, which is randomly sampled and kept unchanged thereafter. • CutMix: The covering region is typically sampled from a Beta distribution to encourage diversity during training. In contrast, we fix the window size to one-quarter of the image to reduce Var(ψ())Var(ψ( x)). • Random Choice: Standard pipelines often compose multiple augmentations sequentially. Here, we randomly choose one of them in each inference step to better control Var(ψ())Var(ψ( x)). 4.2 Logits Processing for StableTTA Referring to the example in Fig. 2(b), a class that is not significant in individual predictions should not emerge as the final prediction solely because it attains the highest averaged logit. To address this issue, we propose Non-Significant Suppression (NSS) to averaging: NSS(,K):=⊙TopK(,K)+min()⋅(−TopK(,K)),K=1,…,CNSS( z,K):= z 1_TopK( z,K)+ ( z)·(1-1_TopK( z,K)), K=1,…,C (2) where, 1 is the all-ones vector, TopK(,K)∈0,1d1_TopK( z,K)∈\0,1\^d is the indicator vector of the Top-K indices (equal to 1 for indices in the Top-K set and 0 otherwise), and ⊙ denotes element-wise multiplication. In this work, we apply NSS to logits averaging as a processing method and prove that it reduces Var()Var( z). From the definition of NSS, we have: NSS(,K)i=i,i∈TopK(,K),min(),i∉TopK(,K),NSS( z,K)_i= cases z_i,&i ( z,K),\\ ( z),&i ( z,K), cases and ∂jNSS(,K)i=1, if i∈TopK(,K) and i=j,0, if i∈TopK(,K) and i≠j,1, if i∉TopK(,K) and j=argmin,0, if i∉TopK(,K) and j≠argmin. ∂ z_jNSS( z,K)_i= cases1,&~if~i ( z,K)~and~i=j,\\ 0,&~if~i ( z,K)~and~i≠ j,\\ 1,&~if~i ( z,K)~and~j=arg\,min z,\\ 0,&~if~i ( z,K)~and~j \,min z. cases Hence, almost everywhere, ‖∇NSS(,K)i‖2=∑j=1C(∂jNSS(,K)i)2=1. \| ( z,K)_i \|^2= _j=1^C ( ∂ z_jNSS( z,K)_i )^2=1. In particular, ‖∇NSS(,K)i‖2≤1. \| ( z,K)_i \|^2≤ 1. Then if ∼(,σ2) z ( μ,σ^2 I). Gaussian Poincaré inequality gives Var(NSS(,K)i)≤[σ2∥∇NSS(,K)i∥2]≤σ2=Var(i).Var(NSS( z,K)_i) [σ^2 \| ( z,K)_i \|^2 ]≤σ^2=Var( z_i). This implies that NSSNSS reduces the variance of each logit i z_i, thereby further avoiding the conflict described in Section 3. In Section 5.3, we discuss the impact of NSS and explore the parameter K, which represents the number of candidates; for additional details, see Appendix D. 4.3 Effectiveness of Augmentation and Logit Processing in StableTTA We summarize the overall formulation of StableTTA as follows: ←1N∑i=1NNSS(f(ψ(i)();),K),ψ(i)∈ψmixup∗,ψCutMix∗, z← 1N _i=1^NNSS(f(ψ^(i)( x); θ),K), ψ^(i)∈\ _mixup^*, _CutMix^*\, where ψmixup∗ _mixup^* and ψcutmix∗ _cutmix^* denote our proposed data augmentation strategies. Our proposed data augmentation strategies consistently yield an additional 4–10% improvement in top-1 accuracy compared to standard mixup and CutMix. These gains also surpass other standard augmentation techniques, such as flipping and cropping. Detailed ablation results are provided in Section 5.2 and Appendix C. Our logit processing method can only be applied when the required stability data augmentations are used (such as the proposed specialized mixup and CutMix). Nevertheless, as demonstrated in the sensitivity analysis in Section 3, it provides an additional 5–12% improvement in top-1 accuracy on already strong baseline models. Detailed experimental results are provided in Section 5.3 and Appendix D. It is important to note that NSS(,C)=⊙+min()⋅(−)=,NSS( z,C)= z 1+ ( z)·(1-1)= z, which implies that when K=CK=C, StableTTA is downgraded to enhance model performance using only our data augmentation strategies, without applying logit processing. 5 Experiments We evaluate the model-boosting capability of StableTTA in terms of accuracy, memory usage, and computational cost. To provide a comprehensive understanding, we first present a full comparison across a wide range of popular models on ImageNet-1K. We then provide visualizations demonstrating that StableTTA enables lightweight architectures to outperform larger models in terms of top-1 accuracy, memory efficiency, and computational cost. Finally, we present ablation studies and sensitivity analysis to further examine the effectiveness and robustness of our method. Dataset. ImageNet-1K [3] is a highly challenging and widely used benchmark dataset, which we adopt for evaluation. It contains 1.28 million training images and 50,000 validation images across 1,000 categories. Over the past years, well-established models have typically achieved only incremental improvements of 1–2% in top-1 accuracy per advancement [13, 28, 7, 32, 4]. Models. All pretrained models used in our experiments are publicly available in torchvision [21], including official implementations, subsequent improvements, and recent training recipes that achieve state-of-the-art performance [23]. We denote these enhanced variants with “†" in this paper. Evaluation. We directly evaluate model performance on a single NVIDIA RTX A6000 GPU. Unless otherwise specified, we use the number of experts N=32N=32, the number of candidates K=10K=10, and a batch size of 16 on the ImageNet-1K validation set. We compare four settings: (1) baseline models, (2) baseline with TTA using our recommend data augmentation strategy, and (3) baseline with StableTTA. Despite stochastic components, results are highly stable, with accuracy variation below 0.01. Both baseline and StableTTA use the same evaluation pipeline, including single-crop inference. 5.1 Main Results We cite the corresponding architecture and training recipe for each pretrained model. Table LABEL:tab:main_results demonstrates that StableTTA yields substantial and consistent improvements in top-1 and top-5 accuracy across popular ImageNet-1K models, while also reporting parameter counts and computational cost. Across all model families, StableTTA provides large accuracy gains, typically ranging from 10.93% to over 32.82% in top-1 accuracy. Notably, even early architectures such as AlexNet and lightweight models such as MobileNet and ShuffleNet achieve performance comparable to or exceeding that of modern large-scale models. Furthermore, a significant number of models surpass the 95% top-1 accuracy threshold, with several exceeding 96%, indicating a strong saturation of performance across diverse architectures. Importantly, Fig. 3 shows that these improvements are achieved without increasing model parameters or inference complexity in the serial setting, highlighting the efficiency of StableTTA. In the parallel setting, StableTTA increases inference cost by a factor of N; however, when applied to lightweight architectures, it remains significantly more economical than using large models. This demonstrates a favorable trade-off between accuracy, memory usage, and computational cost. These results suggest that StableTTA effectively bridges the gap between lightweight and high-performance models, enabling practical deployment in resource-constrained environments. Table 1: Comparison with popular ImageNet-1K classification benchmarks. We compare baseline performance with our proposed StableTTA, reporting model size, computational cost, and significant improvements in top-1 and top-5 validation accuracy. Peak Baseline (%) StableTTA (%) Weight Params GFLOPS Acc@1 Acc@5 Acc@1 Acc@5 AlexNet [13, 21] 61.1M 0.71 56.522 79.066 89.332 98.028 ConvNeXt_Tiny [17, 21] 28.6M 4.46 82.52 96.146 95.958 99.884 ConvNeXt_Small [17, 21] 50.2M 8.68 83.616 96.65 95.932 99.914 ConvNeXt_Base [17, 21] 88.6M 15.36 84.062 96.87 95.964 99.906 ConvNeXt_Large [17, 21] 197.8M 34.36 84.414 96.976 95.998 99.932 DenseNet121 [10, 21] 8.0M 2.83 74.434 91.972 94.160 99.370 DenseNet161 [10, 21] 28.7M 7.73 77.138 93.56 94.814 99.572 DenseNet169 [10, 21] 14.1M 3.36 75.6 92.806 94.600 99.510 DenseNet201 [10, 21] 20.0M 4.29 76.896 93.37 94.734 99.562 EfficientNet_B0 [32, 21] 5.3M 0.39 77.692 93.532 94.988 99.678 EfficientNet_B1 [32, 21] 7.8M 0.69 78.642 94.186 94.940 99.738 EfficientNet_B2 [32, 21] 9.1M 1.09 80.608 95.31 95.352 99.796 EfficientNet_B3 [32, 21] 12.2M 1.83 82.008 96.054 95.816 99.862 EfficientNet_B4 [32, 21] 19.3M 4.39 83.384 96.594 95.386 99.874 EfficientNet_B5 [32, 21] 30.4M 10.27 83.444 96.628 95.328 99.872 EfficientNet_B6 [32, 21] 43.0M 19.07 84.008 96.916 95.504 99.892 EfficientNet_B7 [32, 21] 66.3M 37.75 84.122 96.908 93.452 99.718 EfficientNet_V2_S [33, 21] 21.5M 8.37 84.228 96.878 96.046 99.926 EfficientNet_V2_M [33, 21] 54.1M 24.58 85.112 97.156 96.102 99.920 EfficientNet_V2_L [33, 21] 118.5M 56.08 85.808 97.788 96.164 99.892 GoogLeNet [29, 21] 6.6M 1.5 69.778 89.53 93.510 99.346 Inception_V3 [30, 21] 27.2M 5.71 77.294 93.45 94.906 99.734 MNASNet0_5 [31, 21] 2.2M 0.1 67.734 87.49 93.122 99.188 MNASNet0_75 [31, 21] 3.2M 0.21 71.18 90.496 93.868 99.484 MNASNet1_0 [31, 21] 4.4M 0.31 73.456 91.51 94.230 99.342 MNASNet1_3 [31, 21] 6.3M 0.53 76.506 93.522 94.876 99.728 MaxVit_T [34, 21] 30.9M 5.56 83.7 96.722 95.996 99.902 MobileNet_V2 [26, 21] 3.5M 0.3 71.878 90.286 93.784 99.262 MobileNet_V2† [26, 23] 3.5M 0.3 72.154 90.822 93.922 99.516 MobileNet_V3_Small [9, 21] 2.5M 0.06 67.668 87.402 92.830 99.094 MobileNet_V3_Large [9, 21] 5.5M 0.22 74.042 91.34 94.350 99.364 MobileNet_V3_Large† [9, 23] 5.5M 0.22 75.274 92.566 94.648 99.656 ResNeXt50_32x4d [36, 21] 25.0M 4.23 77.618 93.698 95.126 99.680 ResNeXt50_32x4d† [36, 23] 25.0M 4.23 81.198 95.34 95.978 99.916 ResNeXt101_32x8d [36, 21] 88.8M 16.41 79.312 94.526 94.862 99.648 ResNeXt101_32x8d† [36, 23] 88.8M 16.41 82.834 96.228 95.764 99.874 ResNet18 [7, 21] 11.7M 1.81 69.758 89.078 93.176 99.194 ResNet34 [7, 21] 21.8M 3.66 73.314 91.42 94.092 99.406 ResNet50 [7, 21] 25.6M 4.09 76.13 92.862 94.518 99.496 ResNet50† [7, 23] 25.6M 4.09 80.858 95.434 95.666 99.866 ResNet101 [7, 21] 44.5M 7.8 77.374 93.546 94.858 99.540 ResNet101† [7, 23] 44.5M 7.8 81.886 95.78 95.804 99.890 ResNet152 [7, 21] 60.2M 11.51 78.312 94.046 95.018 99.558 ResNet152† [7, 23] 60.2M 11.51 82.284 96.002 95.942 99.890 ShuffleNet_V2_X0_5 [19, 21] 1.4M 0.04 60.552 81.746 91.372 98.724 ShuffleNet_V2_X1_0 [19, 21] 2.3M 0.14 69.362 88.316 93.380 99.342 ShuffleNet_V2_X1_5 [19, 21] 3.5M 0.3 72.996 91.086 94.238 99.580 ShuffleNet_V2_X2_0 [19, 21] 7.4M 0.58 76.23 93.006 94.996 99.716 SqueezeNet1_0 [11, 21] 1.2M 0.82 58.092 80.42 89.030 97.878 SqueezeNet1_1 [11, 21] 1.2M 0.35 58.178 80.624 89.642 98.020 Swin_T [16, 21] 28.3M 4.49 81.474 95.776 95.798 99.864 Swin_S [16, 21] 49.6M 8.74 83.196 96.36 95.894 99.914 Swin_B [16, 21] 87.8M 15.43 83.582 96.64 95.832 99.916 Swin_V2_T [15, 21] 28.4M 5.94 82.072 96.132 95.798 99.864 Swin_V2_S [15, 21] 49.7M 11.55 83.712 96.816 96.024 99.902 Swin_V2_B [15, 21] 87.9M 20.32 84.112 96.864 95.890 99.938 VGG11_BN [28, 21] 132.9M 7.61 70.37 89.81 93.542 99.124 VGG11 [28, 21] 132.9M 7.61 69.02 88.628 93.236 99.038 VGG13_BN [28, 21] 133.1M 11.31 71.586 90.374 93.666 99.190 VGG13 [28, 21] 133.0M 11.31 69.928 89.246 93.400 99.084 VGG16_BN [28, 21] 138.4M 15.47 73.36 91.516 93.978 98.242 VGG16 [28, 21] 138.4M 15.47 71.592 90.382 93.542 99.164 VGG19_BN [28, 21] 138.4M 19.63 74.218 91.842 94.122 99.266 VGG19 [28, 21] 138.4M 19.63 72.376 90.876 93.868 99.246 ViT_B_16 [4, 21] 86.6M 17.56 81.072 95.318 95.850 99.886 ViT_B_32 [4, 21] 88.2M 4.41 75.912 92.466 95.000 99.778 ViT_L_16 [4, 21] 304.3M 61.55 79.662 94.638 95.314 99.860 ViT_L_32 [4, 21] 306.5M 15.38 76.972 93.07 95.132 99.804 Wide_ResNet101_2 [38, 21] 126.9M 22.75 78.848 94.284 95.146 99.666 Wide_ResNet101_2† [38, 23] 126.9M 22.75 82.51 96.02 95.914 99.898 Wide_ResNet50_2 [38, 21] 68.9M 11.4 78.468 94.086 95.168 99.654 Wide_ResNet50_2† [38, 23] 68.9M 11.4 81.602 95.758 95.736 99.868 Top-1 Validation Accuracy (%)Params (Millions)Peak GFLOPs (Series Mode)Total GFLOPs (Parallel Mode) Figure 3: Efficiency–Accuracy Trade-offs. Comparison of baseline (blue) and StableTTA (red) across parameters (left), peak GFLOPs in series mode (middle), and total GFLOPs in parallel mode (right). Series mode processes images sequentially, while parallel mode processes augmented inputs simultaneously. StableTTA improves accuracy while maintaining favorable efficiency, enabling lightweight models to outperform larger ones. Top-1 Validation Accuracy (%)Number of Experts (N)Number of Experts (N)Number of Experts (N) (a) Ablation Study (b) Sensitivity Analysis Figure 4: (a) TTA (with our augmentation) vs. StableTTA. (b) StableTTA is robust to K and mainly affected by N, but disabling logit processing (K=C)(K=C) will significantly reduce top-1 accuracy. 5.2 Ablation Study We compare standard test-time augmentation (TTA) using MixUp and CutMix with our proposed StableTTA. As shown in Fig. 4(a), StableTTA consistently outperforms standard TTA across different models and numbers of experts (N), achieving higher top-1 and top-5 accuracy while maintaining stable performance gains as N increases. We also observe a diminishing marginal effect as N increases, indicating that larger numbers of experts provide limited additional benefit. Therefore, we adopt N=32N=32 in Table 1 as a practical setting that balances performance and computational cost. We do not consider additional TTA strategies such as flipping, cropping, or image translation, as their improvements are modest (typically less than 2%) [27, 12, 14]. More detailed ablation results are provided in Appendix C. 5.3 Sensitivity Analysis We analyze the sensitivity of StableTTA to key hyperparameters, including the number of candidates K and the number of experts N. As shown in Fig. 4(b), StableTTA demonstrates strong robustness to the choice of N, with performance remaining stable across a wide range of values. In contrast, the number of candidates N has only a marginal effect on top-1 accuracy, indicating that StableTTA does not rely heavily on large ensemble sizes. Importantly, when logit processing is disabled (K=CK=C), the performance drops significantly, highlighting its critical role in achieving accuracy gains. These results suggest that StableTTA provides stable and reliable improvements without requiring careful hyperparameter tuning.. Additional sensitivity analysis results are provided in Appendix D. 5.4 Limitations and Future Work The proposed method can be executed in either sequential or parallel modes. In sequential mode, it does not require additional computational resources beyond a single model, but it increases inference latency approximately in proportion to N due to aggregation. In contrast, parallel execution achieves inference time comparable to that of a single model with a single forward pass, at the cost of N times the computational resources. This trade-off motivates our evaluation in Fig. 3: we compare StableTTA applied to lightweight architectures against large models, rather than comparing StableTTA-enhanced large models to their original counterparts. Using lightweight models ensures that the increased computational cost and runtime introduced by StableTTA remain lower than those of a single large model. For future work, we plan to extend StableTTA to other tasks, such as semantic segmentation [18] and object detection [24], and explore its integration with techniques such as model pruning [5] and knowledge distillation [8]. 6 Conclusion In this work, we reconstructed the efficiency–performance trade-off inherent in ensemble methods and observed a fundamental conflict in conventional aggregation strategies that limits prediction stability. To address this issue, we introduced StableTTA, a simple yet effective training-free approach that combines image preprocessing with logit post-processing to improve the stability of aggregation. Extensive experiments on ImageNet-1K demonstrate that StableTTA delivers substantial and consistent improvements across a wide range of models, boosting top-1 accuracy by 11–33% and enabling the majority of evaluated architectures to surpass 95% accuracy. Notably, our method allows lightweight models to outperform significantly larger architectures while drastically reducing memory footprint and computational cost. This highlights StableTTA’s potential as a practical solution for deploying high-accuracy vision systems in resource-constrained environments. Overall, StableTTA challenges the conventional reliance on larger models and costly ensembles by showing that simple aggregation can unlock significant performance gains. We hope this work encourages further exploration of training-free adaptation methods that enhance both efficiency and robustness across a broader range of applications, such as semantic segmentation and object detection. References [1] L. Breiman (1996) Bagging predictors. Machine learning 24 (2), p. 123–140. Cited by: §1, §2. [2] E. D. Cubuk, B. Zoph, D. Mane, V. Vasudevan, and Q. V. Le (2019) Autoaugment: learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 113–123. Cited by: 2nd item. [3] J. Deng, W. Dong, R. Socher, L. Li, K. Li, and L. Fei-Fei (2009) Imagenet: a large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, p. 248–255. Cited by: §1, §5. [4] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al. (2020) An image is worth 16x16 words: transformers for image recognition at scale. arXiv preprint arXiv:2010.11929. Cited by: §1, §1, §5, Table 1, Table 1, Table 1, Table 1. [5] S. Han, J. Pool, J. Tran, and W. Dally (2015) Learning both weights and connections for efficient neural network. Advances in neural information processing systems 28. Cited by: §5.4. [6] T. Hastie (2009) The elements of statistical learning: data mining, inference, and prediction. springer. Cited by: §1, §2, §3. [7] K. He, X. Zhang, S. Ren, and J. Sun (2016) Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 770–778. Cited by: §1, §1, §2, §2, §5, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1. [8] G. Hinton, O. Vinyals, and J. Dean (2015) Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531. Cited by: §5.4. [9] A. Howard, M. Sandler, G. Chu, L. Chen, B. Chen, M. Tan, W. Wang, Y. Zhu, R. Pang, V. Vasudevan, et al. (2019) Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, p. 1314–1324. Cited by: Table 1, Table 1, Table 1. [10] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger (2017) Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 4700–4708. Cited by: §1, §2, Table 1, Table 1, Table 1, Table 1. [11] F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally, and K. Keutzer (2016) SqueezeNet: alexnet-level accuracy with 50x fewer parameters and< 0.5 mb model size. arXiv preprint arXiv:1602.07360. Cited by: Table 1, Table 1. [12] I. Kim, Y. Kim, and S. Kim (2020) Learning loss for test-time augmentation. Advances in neural information processing systems 33, p. 4163–4174. Cited by: §2, §5.2. [13] A. Krizhevsky, I. Sutskever, and G. E. Hinton (2012) Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25. Cited by: §1, §2, §5, Table 1. [14] Z. Li, J. Cheng, and H. H. Gu (2025) LossTransform: reformulating the loss function for contrastive learning. Information 16 (12), p. 1068. Cited by: §2, §5.2. [15] Z. Liu, H. Hu, Y. Lin, Z. Yao, Z. Xie, Y. Wei, J. Ning, Y. Cao, Z. Zhang, L. Dong, et al. (2022) Swin transformer v2: scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 12009–12019. Cited by: Table 1, Table 1, Table 1. [16] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo (2021) Swin transformer: hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, p. 10012–10022. Cited by: Table 1, Table 1, Table 1. [17] Z. Liu, H. Mao, C. Wu, C. Feichtenhofer, T. Darrell, and S. Xie (2022) A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 11976–11986. Cited by: §1, Table 1, Table 1, Table 1, Table 1. [18] J. Long, E. Shelhamer, and T. Darrell (2015) Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 3431–3440. Cited by: §5.4. [19] N. Ma, X. Zhang, H. Zheng, and J. Sun (2018) Shufflenet v2: practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), p. 116–131. Cited by: Table 1, Table 1, Table 1, Table 1. [20] D. Mahajan, R. Girshick, V. Ramanathan, K. He, M. Paluri, Y. Li, A. Bharambe, and L. Van Der Maaten (2018) Exploring the limits of weakly supervised pretraining. In Proceedings of the European conference on computer vision (ECCV), p. 181–196. Cited by: §1. [21] TorchVision: pytorch’s computer vision library Note: https://github.com/pytorch/vision Cited by: §5, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1. [22] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, et al. (2011) Scikit-learn: machine learning in python. the Journal of machine Learning research 12, p. 2825–2830. Cited by: §3. [23] PyTorch Team (2021) How to train state-of-the-art models using torchvision’s latest primitives. Note: https://pytorch.org/blog/how-to-train-state-of-the-art-models-using-torchvision-latest-primitives/Accessed: 2026-03-25 Cited by: §5, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1. [24] S. Ren, K. He, R. Girshick, and J. Sun (2015) Faster r-cnn: towards real-time object detection with region proposal networks. Advances in neural information processing systems 28. Cited by: §5.4. [25] T. Ridnik, E. Ben-Baruch, A. Noy, and L. Zelnik-Manor (2021) Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972. Cited by: §1. [26] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L. Chen (2018) Mobilenetv2: inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 4510–4520. Cited by: Table 1, Table 1. [27] D. Shanmugam, D. Blalock, G. Balakrishnan, and J. Guttag (2021) Better aggregation in test-time augmentation. In Proceedings of the IEEE/CVF international conference on computer vision, p. 1214–1223. Cited by: §2, §5.2. [28] K. Simonyan and A. Zisserman (2014) Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556. Cited by: §1, §5, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1. [29] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich (2015) Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 1–9. Cited by: Table 1. [30] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna (2016) Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 2818–2826. Cited by: Table 1. [31] M. Tan, B. Chen, R. Pang, V. Vasudevan, M. Sandler, A. Howard, and Q. V. Le (2019) Mnasnet: platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 2820–2828. Cited by: Table 1, Table 1, Table 1, Table 1. [32] M. Tan and Q. Le (2019) Efficientnet: rethinking model scaling for convolutional neural networks. In International conference on machine learning, p. 6105–6114. Cited by: §1, §2, §5, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1, Table 1. [33] M. Tan and Q. Le (2021) Efficientnetv2: smaller models and faster training. In International conference on machine learning, p. 10096–10106. Cited by: Table 1, Table 1, Table 1. [34] Z. Tu, H. Talebi, H. Zhang, F. Yang, P. Milanfar, A. Bovik, and Y. Li (2022) Maxvit: multi-axis vision transformer. In European conference on computer vision, p. 459–479. Cited by: Table 1. [35] A. Virmaux and K. Scaman (2018) Lipschitz regularity of deep neural networks: analysis and efficient estimation. Advances in neural information processing systems 31. Cited by: §4.1. [36] S. Xie, R. Girshick, P. Dollár, Z. Tu, and K. He (2017) Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 1492–1500. Cited by: Table 1, Table 1, Table 1, Table 1. [37] S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y. Yoo (2019) Cutmix: regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, p. 6023–6032. Cited by: §4.1. [38] S. Zagoruyko and N. Komodakis (2016) Wide residual networks. arXiv preprint arXiv:1605.07146. Cited by: Table 1, Table 1, Table 1, Table 1. [39] H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz (2018) Mixup: beyond empirical risk minimization. In International Conference on Learning Representations, Cited by: §4.1. Appendix A Monte Carlo Simulation Consider the logit vector ∼(,σ2) z ( μ,σ^2 I) in a binary classification task. Then, (1−21>2) pmatrix z_1- z_2\\ 1_ z_1> z_2 pmatrix =[1−2Φ(1−22σ)], = bmatrix μ_1- μ_2\\ ( μ_1- μ_2 2σ ) bmatrix, (3) Var(1−21>2) pmatrix z_1- z_2\\ 1_ z_1> z_2 pmatrix =[2σ22σϕ(a)2σϕ(a)ϕ(a)−ϕ2(a)], = bmatrix2σ^2& 2σφ(a)\\ 2σφ(a)&φ(a)-φ^2(a) bmatrix, (4) where a=1−22σ.a= μ_1- μ_2 2σ. By the Central Limit Theorem, for large N, the sample averages satisfy (1N∑i=1N1(i)−2(i)1N∑i=1N1(i)>2(i))∼([1−2Φ(1−22σ)],1N[2σ22σϕ(a)2σϕ(a)ϕ(a)−ϕ2(a)]). pmatrix 1NΣ _i=1^N z_1^(i)- z_2^(i)\\ 1NΣ _i=1^N 1_ z_1^(i)> z_2^(i) pmatrix ( bmatrix μ_1- μ_2\\ ( μ_1- μ_2 2σ ) bmatrix, 1N bmatrix2σ^2& 2σφ(a)\\ 2σφ(a)&φ(a)-φ^2(a) bmatrix ). Define the standardized variables [XY]:=[1N∑i=1N(1(i)−2(i))−(1−2)2σ/N1N∑i=1N1(i)>2(i)−ϕ(a)ϕ(a)−ϕ2(a)/N], bmatrixX\\ Y bmatrix:= bmatrix 1NΣ _i=1^N( z_1^(i)- z_2^(i))-( μ_1- μ_2) 2σ/ N\\ 1NΣ _i=1^N 1_ z_1^(i)> z_2^(i)-φ(a) φ(a)-φ^2(a)/ N bmatrix, which jointly follow [XY]∼([00],[1Nϕ(a)1−ϕ(a)Nϕ(a)1−ϕ(a)1]). bmatrixX\\ Y bmatrix ( bmatrix0\\ 0 bmatrix, bmatrix1& Nφ(a)1-φ(a)\\ Nφ(a)1-φ(a)&1 bmatrix ). Conflict rate P(y^logit≠y^hard)P( y_logit≠ y_hard)Gaussian noise σ=Var()σ= Var( z) Figure 5: Monte Carlo simulation. The conflict probability increases as Var()Var( z) grows. In this simulation, we consider distributions: ∼(,σI)∣μ∈(1,0.9),(1,0.7),(1,0.5),σ∈[0.05,0.25]\ z ( μ,σ I) μ∈\(1,0.9),(1,0.7),(1,0.5)\,σ∈[0.05,0.25]\. The solid curves show Monte Carlo estimates of the relationship between σ and P(y^logit≠y^hard)P( y_logit≠ y_hard), while the dashed curves correspond to the theoretical (asymptotic) predictions. The empirical and theoretical results are closely matched. We now compute the probability of conflict between the logit-averaging and the hard voting method. First, P(y^logit=1,y^hard=2) ( y_logit=1, y_hard=2) =P(1N∑k=1N(1(k)−2(k))>0,1N∑i=1N1(k)>2(k)≤12) =P ( 1N _k=1^N( z_1^(k)- z_2^(k))>0, 1N _i=1^N 1_ z_1^(k)> z_2^(k)≤ 12 ) =P(X>−Na,Y≤(1/2−ϕ(a))Nϕ(a)−ϕ2(a)) =P (X>- Na,Y≤ (1/2-φ(a)) N φ(a)-φ^2(a) ) =Φ((1/2−ϕ(a))Nϕ(a)−ϕ2(a)) = ( (1/2-φ(a)) N φ(a)-φ^2(a) ) −Φ2(−aN,(1/2−ϕ(a))Nϕ(a)−ϕ2(a);ϕ(a)Φ(a)−Φ2(a)). - _2 (-a N, (1/2-φ(a)) N φ(a)-φ^2(a); φ(a) (a)- ^2(a) ). (5) Similarly, P(y^logit=2,y^hard=1) ( y_logit=2, y_hard=1) =P(1N∑k=1N(1(k)−2(k))≤0,1N∑i=1N1(k)>2(k)>12) =P ( 1N _k=1^N( z_1^(k)- z_2^(k))≤ 0, 1N _i=1^N 1_ z_1^(k)> z_2^(k)> 12 ) =P(X≤−Na,Y>(1/2−ϕ(a))Nϕ(a)−ϕ2(a)) =P (X≤- Na,Y> (1/2-φ(a)) N φ(a)-φ^2(a) ) =Φ(−Na) = (- Na ) −Φ2(−aN,(1/2−ϕ(a))Nϕ(a)−ϕ2(a);ϕ(a)Φ(a)−Φ2(a)). - _2 (-a N, (1/2-φ(a)) N φ(a)-φ^2(a); φ(a) (a)- ^2(a) ). (6) Therefore, the total conflict probability is P(y^logit≠y^hard) ( y_logit≠ y_hard) =P(y^logit=1,y^hard=2)+P(y^logit=2,y^hard=1) =P( y_logit=1, y_hard=2)+P( y_logit=2, y_hard=1) =Φ(−Na)+Φ((1/2−ϕ(a))Nϕ(a)−ϕ2(a)) = (- Na )+ ( (1/2-φ(a)) N φ(a)-φ^2(a) ) −2Φ2(−aN,(1/2−ϕ(a))Nϕ(a)−ϕ2(a);ϕ(a)Φ(a)−Φ2(a)). -2 _2 (-a N, (1/2-φ(a)) N φ(a)-φ^2(a); φ(a) (a)- ^2(a) ). (7) This expression indicates that the conflict rate increases as Var()=σ2Var( z)=σ^2 grows. Fig. 5 provides empirical support for this conclusion. The solid curves represent Monte Carlo estimates of the relationship between σ and P(y^logit≠y^hard)P( y_logit≠ y_hard), while the dashed curves correspond to the theoretical (asymptotic) predictions. We observe that the empirical and theoretical curves exhibit consistent trends, with closely matching derivatives. Appendix B Hölder Condition Figure 6 provides an intuitive visualization of how different data augmentation strategies influence the variance of logits under the Hölder continuity assumption. According to Eq. (1), the distance between logit vectors is bounded by the distance between augmented inputs. As illustrated in the left example, translation preserves the semantic structure of the image but introduces large pixel-wise differences between augmented samples, leading to a larger input distance ‖ψi()−ψj()‖\| _i( x)- _j( x)\| and consequently higher variance in the resulting logits. In contrast, the right example shows random erasing, which applies stronger noise but gives smaller overall pixel-wise differences across augmentations. This reduces the variance of the logit vector, producing more concentrated predictions that are better suited for stable aggregation. The figure thus highlights that not all augmentations contribute equally to logit stability. Figure 6: Varying Effects of Data Augmentation Methods on the Variance of Logits. Under Hölder continuity, larger changes in the input lead to larger changes in the logits. Left: Translation preserves the semantic content of the image but introduces large pixel-wise differences, resulting in a higher variance in the logit distribution. Right: Random erasing applies stronger modifications to the image, but the overall pixel-wise differences are smaller, leading to more concentrated logits that are better suited to stable aggregation. Table 2: Ablation study of StableTTA versus standard TTA under varying numbers of experts (N) across multiple architectures. StableTTA consistently outperforms TTA under MixUp and CutMix data augmentations. Weight Params N Acc@1 Acc@5 AlexNet 61.1M 4 65.590 87.022 AlexNet 61.1M 8 68.604 89.432 AlexNet 61.1M 16 70.496 90.510 AlexNet 61.1M 32 71.258 90.938 ResNet50† 25.6M 4 88.556 98.858 ResNet50† 25.6M 8 89.542 99.098 ResNet50† 25.6M 16 90.232 99.172 ResNet50† 25.6M 32 90.612 99.240 ViT_B_16 86.6M 4 87.052 98.470 ViT_B_16 86.6M 8 88.206 98.938 ViT_B_16 86.6M 16 89.048 99.030 ViT_B_16 86.6M 32 89.382 99.074 Table 3: Sensitivity analysis of StableTTA over K and N. Results show robustness to K, consistent gains with larger N, and significant drops when logit processing is disabled (K=CK=C). =N=4 =N=8 =N=16 =N=32 Model K Acc@1 Acc@5 Acc@1 Acc@5 Acc@1 Acc@5 Acc@1 Acc@5 AlexNet (61.1M) (0.71GFLOPS) 1 78.412 93.182 84.638 96.134 87.942 97.530 89.576 98.052 2 78.444 93.262 84.982 96.312 87.946 97.528 89.496 98.052 5 78.804 93.328 84.666 96.220 87.984 97.412 89.484 98.108 10 78.750 93.434 84.810 96.216 87.864 97.470 89.532 98.030 20 78.540 93.320 84.778 96.160 88.072 97.540 89.530 98.076 C 66.212 83.420 71.606 91.158 74.948 94.520 76.918 96.222 ResNet50† (25.6M) (4.09GFLOPS) 1 91.524 99.082 93.788 99.614 95.076 99.802 95.706 99.860 2 91.436 99.120 93.830 99.624 95.022 99.780 95.552 99.834 5 91.484 99.084 93.746 99.622 94.998 99.818 95.656 99.860 10 91.276 99.108 93.878 99.676 95.192 99.800 95.646 99.852 20 91.576 99.126 93.726 99.598 95.008 99.770 95.642 99.876 C 86.670 96.534 89.820 98.822 92.034 99.542 93.110 99.744 ViT_B_16 (86.6M) (17.56GFLOPS) 1 91.488 99.142 93.906 99.676 95.308 99.822 95.822 99.864 2 91.604 99.128 94.000 99.670 95.304 99.810 95.826 99.878 5 91.602 99.112 93.894 99.682 95.088 99.820 95.824 99.884 10 91.668 99.152 94.042 99.656 95.114 99.820 95.852 99.888 20 91.468 99.126 94.972 99.648 95.234 99.816 95.940 99.904 C 86.710 96.578 89.898 98.844 91.648 99.540 92.832 99.744 Appendix C Ablation Study Table 2 reports the performance comparison between standard TTA and our StableTTA under MixUp and CutMix augmentation strategies, across different architectures and varying numbers of experts (N). Consistent with the trends observed in Fig. 4(a), StableTTA outperforms TTA. Increasing N steadily improves both top-1 and top-5 accuracy for all models, including AlexNet, ResNet50, and ViT-B16. However, the performance gains gradually diminish as N increases, indicating a clear saturation effect. This table provides evidence emphasizing the advantages of StableTTA. Appendix D Sensitivity Analysis Table 3 presents a detailed sensitivity analysis of StableTTA with respect to the number of candidates K and experts N across different architectures. Consistent with the trends observed in Fig. 4(b), StableTTA exhibits strong robustness to the choice of K, with performance remaining largely stable across a wide range of values. Increasing N consistently improves both top-1 and top-5 accuracy, while the impact of varying K is relatively minor. Notably, when logit processing is disabled (K=CK=C), performance drops significantly across all models, reaffirming the critical role of our logit processing (NSS) in achieving accuracy gains. Overall, these results demonstrate that StableTTA delivers stable, reliable improvements without requiring careful hyperparameter tuning.