This work proposes a principled VLM TTA method called \algname, and theoretically reveals that the InfoNCE loss can be neatly reformulated as a Wasserstein OT formulation, thereby unifying the objectives of the inference and adaptation of VLMs to achieve their mutual benefits.
Abstract
Vision-language models (VLMs) have demonstrated remarkable zero-shot capabilities yet remain sensitive to real-world distribution shifts during inference. Although significant efforts are devoted to adapting VLMs at test time, they rely heavily on noisy pseudo-labels predicted directly from raw embedding similarities during inference, which are unreliable under distribution shift and mislead the adaptation. To avoid noise amplification, existing works craft coarse-grained surrogate objectives during adaptation, which fail to explicitly model sample-level relationships across different modalities, creating objective mismatch with inference, thus leading to marginal performance improvement. In this work, we aim to bridge the detached objectives of inference and adaptation for VLMs, and propose a principled VLM TTA method called \algname. For VLM inference, we formulate the zero-shot image classification task as a cross-modal alignment problem encoded via a Wasserstein OT formulation, providing robust pseudo-labels at the sample-level to effectively adapt VLMs. For VLM adaptation, we adopt a soft-label InfoNCE loss to adapt VLMs based on the OT-induced pseudo-labels, leveraging fine-grained supervisions to explicitly model relationships of individual image-text pairs via contrastive learning, which empowers accurate inference at the same granularity. Moreover, we theoretically reveal that the InfoNCE loss can be neatly reformulated as a Wasserstein OT formulation, thereby unifying the objectives of the inference and adaptation of VLMs to achieve their mutual benefits. Extensive experiments demonstrate the effectiveness and efficiency of our methods, outperforming the best-performing methods by up to 7% with state-of-the-art efficiency.
Vision-Language-Action (VLA) models often fail to perform the same learned tasks under environmental shifts, such as changes in camera pose and shifts to a different but similar robot (e.g., from Panda to UR5e). Adapting these models to the shifted environment (i.e., target domain) often requires training on multiple demonstrations for each task, which are costly to collect. To reduce the burden of data curation and training, we propose an analogy-based method that adapts VLA models under environmental shifts through weight vector arithmetic with domain-specific information addition, named Domain ARiThmetic (DART). Unlike prior approaches, DART requires collecting only a single demonstration, enabling efficient adaptation. To accurately isolate domain-specific information for addition, DART performs subspace alignment between singular components in weight vectors to filter out noisy components. In both simulated and real-world experiments, DART outperforms existing VLA adaptation methods in one-shot scenarios across diverse visual and embodiment shifts. Code is available at https://github.com/snumprlab/dart.
Taewook Kang, Taeheon Kim, Donghyun Shin et al.· 0 citations
This work proposes RITA, a Robust test-tIme prompt-TAdaptation framework that shifts from sample-level estimates to distribution-level alignment, and employs optimal transport to align the distribution of augmented visual features with textual prototypes, mitigating adversarial outliers and rectifying cross-modal semantic misalignment.
Xingyu Zhu, Huanshen Wu, Shuo Wang et al.· 1 citation
Vision-Language-Action (VLA) models have made significant strides in embodied intelligence by integrating the powerful representations of pre-trained Vision-Language Models (VLMs). However, the massive parameter scale of VLAs imposes a heavy computational burden, and these models exhibit extreme sensitivity to parameter pruning. Current paradigms often treat the resulting performance degradation as inevitable, relying on fine-tuning or low-rank corrections to recover efficacy. We challenge this convention by questioning whether the removed parameters are truly redundant if VLA pruning necessitates performance recovery to be effective, or if this paradigm masks the indiscriminate pruning of critical parameters. We revisit parameter redundancy through the lens of VLM-to-VLA adaptation, first quantifying the spatial distribution of parameter divergence during adaptation to reveal structured patterns across different modules. Subsequently, we introduce controlled pruning as a diagnostic probe: by comparing the direct impact of removing different parameter subsets on VLA performance without any fine-tuning, we establish a causal link between adaptation-induced divergence signals and functional contributions. Based on the discovered modular heterogeneities, we design a multi-module joint pruning scheme. Evaluations on the LIBERO benchmark demonstrate that our approach reduces the parameters of OpenVLA and $\pi_{0.5}$ by 12\%--30\% while maintaining approximately 90\% of the original performance without any post-pruning recovery. In contrast, existing parameter pruning criteria result in total performance collapse when evaluated under the same recovery-free constraints. Our study reveals the parameter evolution mechanism in VLA adaptation and provides a new path for deploying efficient, robust robotic policies in resource-constrained environments.
Feng Zhang, Tao Huang, Siyu Xu et al.· arXiv.org· 1 citation
Vision-language models (VLMs) remain unreliable when predictions require fine-grained visual evidence. We identify a previously overlooked cause: spectral response rigidity. Despite substantial frequency variation across images and tasks, pretrained vision encoders exhibit persistent, encoder-specific layerwise spectral profiles that change only marginally under downstream fine-tuning. Since pretrained vision encoders only receive images, they cannot adapt spectral extraction to the evidence required by the current query. We therefore propose HAFI-VLM, which introduces a task-conditioned frequency pathway while preserving the pretrained semantic representation. Hierarchical Adaptive Frequency Injection (HAFI) retrieves complementary low-, mid-, and high-frequency evidence at multiple encoder depths using text-modulated, spatially aligned cross-attention. A Visual Enrichment Layer Adapter further recalibrates shallow LLM attention to effectively utilize the enriched visual tokens. Experiments on LLaVA-1.5 and Qwen2.5-VL demonstrate consistent improvements in general VQA, text-rich understanding, and hallucination robustness, outperforming representation-level enhancement methods and most resolution- or cropping-based approaches without additional high-resolution encoding. Mechanistic analyses show that HAFI restores task-dependent spectral allocation while retaining semantic attention, establishing frequency enrichment as a distinct and effective route for improving VLM perception.
Jin Cui, Chuanchang Su, Jiayi Lu et al.· 0 citations
Vision-language models (VLMs), such as contrastive language-image pre-training (CLIP), exhibit powerful zero-shot generalization capabilities. Parameter-efficient fine-tuning (PEFT) techniques, notably prompt learning, have been extensively explored to adapt these models to downstream tasks. However, their efficacy remains constrained when transferred to specialized domains like remote sensing. We argue that the bottleneck stems not merely from the limited parameters of prompts, but essentially from the disruption of the input’s original image–text features and the lack of deep cross-modal alignment. In particular, existing methods typically rely on global attention or coarse-grained feature mapping. This inadvertently corrupts the original input representations, thereby impairing the model’s inherent generalization. Furthermore, their isolated unimodal gradient updates fail to bridge the semantic gap inherent in complex remote sensing scenes. To address these challenges, we propose tokenwise prompt-free learning (Tiper), shifting the optimization paradigm from introducing external prompts to precisely recalibrating the critical tokens that govern classification outputs. In particular, Tiper employs a hierarchical learner to supersede global prompts. Crucially, this learner intervenes exclusively on the specific core tokens (i.e., the CLS token in the visual branch and the EOT token in the textual branch), leaving other original input representations unperturbed. This fine-grained strategy effectively balances domain adaptation with the preservation of inherent generalization. Finally, we design the learner as a cross-modal coupled bridge with shared weights, enabling it to synchronously receive gradient feedback from both modalities and fostering profound multimodal collaboration. Extensive experiments validate our method on eight public remote sensing datasets covering diverse scenes and resolutions. In the base-to-new generalization task, Tiper outperforms the strong baseline MaPLe with a significant 3.7% improvement in the harmonic mean (HM). Notably, without relying on any external large-scale domain models, Tiper surpasses the latest domain-specific prompt learning methods (e.g., domain-controlled prompt learning (DCPL), domain prompt learning with quaternion networks (DPLQ)), demonstrating its superior adaptability for remote sensing image scene classification.
Tengfei Gong, Junlin Wu, Yaxioong Chen et al.· IEEE Transactions on Geoscie...· 0 citations
Known for his clear and elegant writing style, Bertsekas shaped fields from control and optimization to large-scale computation and artificial intelligence.
