GS-Bias: Global-Spatial Test-Time Adaptation

• The paper introduces GS-Bias, which leverages learnable global and spatial bias vectors to adjust VLM outputs at test time.
• It utilizes output-side correction through augmented views and selective region analysis to combat distribution shifts efficiently.
• Empirical results demonstrate improved cross-dataset and domain generalization, achieving higher accuracy with lower memory and runtime cost.

Searching arXiv for the cited GS-Bias paper and closely related spatial/geographic bias work to ground the article. Global-Spatial Bias Learner (GS-Bias) is a single-image test-time adaptation paradigm for vision-LLMs (VLMs) in which two learnable, test-sample-specific bias vectors—a global bias and a spatial bias—are optimized for each test image and added directly to the pretrained model’s output logits. In its canonical formulation, developed for CLIP-based zero-shot classification, GS-Bias is intended to preserve the benefits of test-time adaptation while avoiding the memory and runtime costs of prompt tuning and the instability of training-free visual feature heuristics. The method treats adaptation as an output-side correction problem: it learns image-level semantic adjustment from augmented views and region-level semantic adjustment from spatial visual representations, while keeping the pretrained VLM frozen (Huang et al., 16 Jul 2025).

1. Definition and problem setting

GS-Bias is defined in the context of single-image test-time adaptation without access to training data. For a test image $x$, a pretrained VLM produces a global visual feature $F_g = E_v(x)$ and class text features $\{F_c^t\}_{c=1}^{C}$, and zero-shot prediction is based on cosine similarity between image and text features. The motivating problem is distribution shift: under cross-dataset transfer, corrupted styles, sketches, adversarial natural examples, or unseen fine-grained categories, vanilla CLIP zero-shot predictions degrade. Existing VLM test-time adaptation methods are framed as having an effectiveness–efficiency tradeoff: prompt-tuning approaches such as TPT and DiffTPT are effective but expensive because they require backpropagation through the whole VLM, while training-free visual enhancement methods such as MTA are efficient but can be unstable or weaker because they rely on handcrafted assumptions over visual features (Huang et al., 16 Jul 2025).

Within this setting, GS-Bias introduces two learnable variables for each test image: a global bias $B_g \in \mathbb{R}^{1\times C}$ and a spatial bias $B_s \in \mathbb{R}^{1\times C}$. Both are initialized to zero and optimized only for the current image. They are added at the logit output stage, not to prompts, text features, encoder weights, or internal activations. The design objective is explicit: retain sample-specific adaptation from augmented views, but move optimization away from “input-side” prompt tuning and toward lightweight output-space correction (Huang et al., 16 Jul 2025).

2. Mathematical formulation

For a test image $x$, the pretrained CLIP probability for class $y_c$ is

$P_{\text{CLIP}(y_c \mid x) = \frac{\exp\left(\cos(F_g, F_c^t)/T\right)} {\sum_{c=1}^{C}\exp\left(\cos(F_g, F_c^t)/T\right)},$

where $T$ is the softmax temperature. GS-Bias modifies this prediction by learning two additive bias terms with different semantic roles (Huang et al., 16 Jul 2025).

The global bias uses a set of augmented views $\{A_i(x)\}_{i=1}^{N}$. Following the confident-view selection mechanism used in TPT-style adaptation, GS-Bias aggregates view predictions and adds a shared class-wise bias $F_g = E_v(x)$0:

$F_g = E_v(x)$1

It then minimizes the entropy of the aggregated distribution and updates $F_g = E_v(x)$2 by gradient descent:

$F_g = E_v(x)$3

The resulting global-bias-adjusted output is

$F_g = E_v(x)$4

The spatial bias operates on region-level visual features. Let the visual encoder provide spatial features $F_g = E_v(x)$5, with region feature $F_g = E_v(x)$6. Region-level class probabilities are computed as

$F_g = E_v(x)$7

To select informative regions, GS-Bias computes a category-aware spatial map

$F_g = E_v(x)$8

then keeps the top-$F_g = E_v(x)$9 regions:

$\{F_c^t\}_{c=1}^{C}$0

A shared spatial bias $\{F_c^t\}_{c=1}^{C}$1 is added over the selected region predictions:

$\{F_c^t\}_{c=1}^{C}$2

and optimized by entropy minimization:

$\{F_c^t\}_{c=1}^{C}$3

The final adapted prediction is

$\{F_c^t\}_{c=1}^{C}$4

In this formulation, $\{F_c^t\}_{c=1}^{C}$5 captures image-level semantics and cross-view agreement, whereas $\{F_c^t\}_{c=1}^{C}$6 captures local semantic coherence from class-relevant regions. The pretrained VLM parameters remain frozen throughout; only $\{F_c^t\}_{c=1}^{C}$7 and $\{F_c^t\}_{c=1}^{C}$8 are optimized (Huang et al., 16 Jul 2025).

3. Inference-time procedure and efficiency profile

GS-Bias is a per-image method. The procedure is: freeze the pretrained VLM; generate augmented views of the test image; compute confident-view global predictions; extract spatial visual representations; score regions by semantic relevance to class text embeddings; optimize $\{F_c^t\}_{c=1}^{C}$9 and $B_g \in \mathbb{R}^{1\times C}$0 for a small number of steps; and return the original CLIP output plus the two learned biases. The implementation uses random cropping as augmentation. Reported settings are $B_g \in \mathbb{R}^{1\times C}$1 views for cross-dataset generalization and $B_g \in \mathbb{R}^{1\times C}$2 for domain generalization, with confident-view selection rates $B_g \in \mathbb{R}^{1\times C}$3 and $B_g \in \mathbb{R}^{1\times C}$4, respectively. The number of selected spatial regions is $B_g \in \mathbb{R}^{1\times C}$5, and experiments use 5 adaptation steps. Learning rates are $B_g \in \mathbb{R}^{1\times C}$6 for cross-dataset evaluation and $B_g \in \mathbb{R}^{1\times C}$7 for domain generalization (Huang et al., 16 Jul 2025).

The main backbone is CLIP ViT-B/16. Appendix results also report CLIP ResNet50, where spatial knowledge is taken from an intermediate feature map of size $B_g \in \mathbb{R}^{1\times C}$8. Because GS-Bias learns only two vectors of shape $B_g \in \mathbb{R}^{1\times C}$9, its memory footprint is small relative to prompt-tuning methods. On ImageNet with a single NVIDIA A800 GPU, reported memory usage is: CLIP 758 MiB, GS-Bias 1308 MiB, MTA 3576 MiB, and TPT 19997 MiB. The paper highlights that GS-Bias requires only 6.5% of TPT’s memory usage on ImageNet. Throughput is also favorable: TPT runs at 1.38 images/sec, MTA at 8.59 images/sec, MTA + ensemble prompt at 12.55 images/sec, and GS-Bias + ensemble prompt at 12.34 images/sec (Huang et al., 16 Jul 2025).

This efficiency profile follows directly from the location of adaptation. Prompt-tuning methods optimize prompt parameters and require full-network backward passes and a second forward pass after prompt updates. GS-Bias instead performs output-side optimization, avoiding full-network backpropagation and “secondary inference” while still learning image-specific corrections (Huang et al., 16 Jul 2025).

4. Empirical performance and ablations

On cross-dataset generalization with ten datasets—Flowers102, DTD, OxfordPets, StanfordCars, UCF101, Caltech101, Food101, SUN397, FGVC-Aircraft, and EuroSAT—GS-Bias + ensemble prompt achieves an average of 67.03, compared with 64.80 for TPT, 65.47 for DiffTPT+, 65.95 for MTA + ensemble, and 65.64 for CLIP + ensemble prompt. The reported gain over TPT is 2.23%. GS-Bias + ensemble is best on the table average while using only 8 augmented views, whereas TPT uses 64 and DiffTPT uses 128 (Huang et al., 16 Jul 2025).

On domain generalization over ImageNet, ImageNet-V2, ImageNet-Sketch, ImageNet-A, and ImageNet-R, GS-Bias + ensemble reaches an OOD average of 63.01, compared with 60.29 for TPT, 60.52 for DiffTPT+, and 62.53 for MTA + ensemble. The gain over TPT is 2.72%. Full GS-Bias + ensemble numbers are ImageNet 70.57, ImageNet-A 56.61, ImageNet-V2 64.62, ImageNet-R 80.49, and ImageNet-Sketch 50.33 (Huang et al., 16 Jul 2025).

Ablation studies show that both bias components contribute, but not equally. With CLIP + ensemble as baseline, the reported averages are: baseline 67.21, global bias only 67.44, spatial bias only 68.26, and both biases 68.80. The paper therefore interprets the two branches as complementary, with spatial bias contributing more strongly than global bias alone in the reported ablations. The sensitivity study over the number of selected regions indicates that performance improves as $B_s \in \mathbb{R}^{1\times C}$0 increases and peaks around $B_s \in \mathbb{R}^{1\times C}$1, after which irrelevant background regions begin to hurt performance. Adaptation-step analysis shows that ImageNet performance is already stable after the first step, while cross-dataset benchmarks benefit from a few steps but can overfit if adaptation is too prolonged (Huang et al., 16 Jul 2025).

The paper also reports compatibility with “historical stream” methods. GS-Bias combined with TDA reaches 68.30 cross-dataset and 70.96 ImageNet while retaining GS-Bias’s 1308M memory footprint, suggesting that the method can be combined with queue-based or memory-based test-time adaptation without changing its core design (Huang et al., 16 Jul 2025).

5. Relation to broader spatial-bias research

GS-Bias in the strict sense refers to the output-side TTA mechanism introduced for VLMs (Huang et al., 16 Jul 2025). Related literature, however, uses closely adjacent language to describe different phenomena. In work on multimodal urban reasoning, Urban-R1 is described as a reinforcement-learning alignment framework that functions as a GS-Bias-like mechanism because its reward-driven policy adaptation suppresses geospatial shortcutting, even though it has no separately parameterized module named GS-Bias (Wang et al., 18 Oct 2025). In this usage, the phrase refers to an implicit geospatial debiaser rather than a standalone bias vector.

Other nearby literatures treat “global-spatial bias” as a measurement problem. The DiST benchmark argues that resistance to style transfer is not the same as sensitivity to global spatial configuration, and that a valid benchmark should directly test whether global arrangement changes are encoded when local texture is preserved (Wen et al., 2023). Work on LVLM spatial robustness shows that semantically decisive content can be moved across image positions while the vision encoder remains robust, and attributes the resulting inconsistency to unbalanced design of position embeddings in the language-model component; its proposed Balanced Position Assignment assigns identical position embeddings to all image tokens to promote more balanced integration of visual information (Zhu et al., 26 Sep 2025). A separate study of VLMs finds a systematic left-first description bias in horizontally concatenated images and interprets it as a spatial processing regularity that is better explained by architectural factors than by reading direction or explicit annotation instructions (Chaudhary et al., 20 Dec 2025).

In geographic fairness and auditing, the term is even broader. One line of work shows that observed disparity between outcomes and sensitive attributes may be induced by shared spatial autocorrelation, and proposes Moran-based diagnostics plus eigenvector spatial filtering as a global spatial confounding-control layer before fairness evaluation (Majumdar et al., 2021). Another proposes a locational meta-referee that dynamically adjusts location-specific learning rates in few-shot meta-learning to improve parity of predictive quality across locations under region shift (Chen et al., 2024). GeoBS then generalizes geo-bias evaluation into an information-theoretic framework with explicit decomposition into multi-scalability, distance decay, and anisotropy through SG-SRE, DL-SRE, and DS-SRE (Wang et al., 27 Sep 2025).

This suggests that “GS-Bias” now spans at least two technical senses: a specific VLM test-time adaptation module built from global and spatial bias vectors, and a broader systems-level idea of learning, measuring, or correcting structured bias over space.

6. Limitations and open questions

The canonical GS-Bias paper identifies a practical limitation: the method depends on empirically selected hyperparameters, including the number of views $B_s \in \mathbb{R}^{1\times C}$2, selection rate $B_s \in \mathbb{R}^{1\times C}$3, number of selected spatial regions $B_s \in \mathbb{R}^{1\times C}$4, learning rates $B_s \in \mathbb{R}^{1\times C}$5, and adaptation steps. Although the reported experiments show reasonable robustness, these hyperparameters may require reconfiguration across datasets or even across test samples. The authors therefore suggest future work on sample-specific dynamic hyperparameter adjustment (Huang et al., 16 Jul 2025).

Further open questions arise when GS-Bias is read against adjacent spatial-bias literatures. A plausible implication is that output-side adaptation is only one part of the spatial-bias problem. Position-assignment work in LVLMs implies that some bias originates in the multimodal fusion mechanism itself, not only in the test-time calibration layer (Zhu et al., 26 Sep 2025). Geographic fairness work implies that spatial bias may have separable components linked to local heterogeneity, distance effects, directional anisotropy, and shared spatial structure, which are not explicitly modeled by the two-vector GS-Bias formulation (Majumdar et al., 2021, Wang et al., 27 Sep 2025). Meta-learning work on locational fairness further suggests that transferable spatial correction policies may require explicit region-shift conditioning rather than per-sample adaptation alone (Chen et al., 2024).

Within the narrower TTA setting, the central unresolved design question is how far output-side bias vectors can go before architectural intervention becomes necessary. GS-Bias demonstrates that logit-space correction can recover a strong efficiency–accuracy tradeoff for VLM test-time adaptation (Huang et al., 16 Jul 2025). Related work on spatial robustness, geospatial fairness, and explicit geographic auditing suggests that a fuller global-spatial bias learner may eventually need to combine three layers: measurement, representation balancing, and task-time adaptation.