Frozen Vision-Language Models

Adverse weather removal (AWR) in real-world images remains challenging due to heterogeneous and unseen degradations, while distortion-driven training often yields overly smooth results. We propose PVRF, a unified framework that integrates zero-shot soft weather perceptions with velocity-constrained rectified-flow refinement. PVRF introduces an AWR-specific question answering module (AWR-QA) that uses frozen vision--language models (VLMs) to estimate soft probabilities of weather types and low-level attribute scores. These perceptions condition restoration networks via attribute-modulated normalization (AMN) and weather-weighted adapters (WWA), producing an anchor estimate for refinement. We then learn a terminal-consistent residual rectified flow with perception-adaptive source perturbation and a terminal-consistent velocity parameterization to stabilize learning near the terminal regime. Extensive experiments show that PVRF improves both fidelity and perceptual quality over state-of-the-art baselines, with strong cross-dataset generalization on single and combined degradations. Code will be released at https://github.com/dongw22/PVRF.

Cross-modal 3D medical image analysis requires voxelwise representations that remain anatomically consistent across imaging contrasts, scanners, and acquisition protocols. Recent work has shown that frozen 2D Vision Transformer (ViT) foundation models can support such representations, but typical pipelines extract features along a single anatomical axis and adapt those features inside a registration solver for one image pair at a time, leaving complementary viewing directions unused and producing representations that do not transfer to new volumes. We introduce VoxCor, a training-free fit--transform method for reusable volumetric feature representations from frozen 2D ViT foundation models. During an offline fitting phase, VoxCor combines triplanar ViT inference with a compact closed-form weighted partial least squares (WPLS) projection that uses fitting-time voxel correspondences to select modality-stable anatomical directions in the triplanar feature space. At transform time, new volumes are mapped by triplanar ViT inference and linear projection alone, without fine-tuning or registration. Voxel correspondences can then be queried directly by nearest-neighbor search. We evaluate VoxCor on intra-subject Abdomen MR--CT and inter-subject HCP T2w--T1w tasks using deformable registration, voxelwise k-nearest-neighbor segmentation, and segmentation-center landmark localization. VoxCor improves the hardest cross-subject, cross-modality transfer settings, reduces encoder sensitivity for dense correspondence transfer, and yields registration performance competitive with handcrafted descriptors and learned 3D features. This positions VoxCor as a reusable feature layer for downstream multimodal analysis beyond pairwise registration. Code, configuration files, and implementation details are publicly available on GitHub at \href{https://github.com/guneytombak/VoxCor}{guneytombak/VoxCor}.

Standard dual-encoder vision-language models that map images and text to deterministic points on a shared unit hypersphere through $\ell_2$ normalization typically expose neither \emph{aleatoric} uncertainty (cross-modal ambiguity) nor \emph{epistemic} uncertainty (lack of training-distribution support). Existing post-hoc methods either recover at most one of the two uncertainty components, or ignore the hyperspherical geometry of these models' embeddings. We propose \textbf{GeoFlowVLM} as a post-hoc adapter that learns the joint distribution of paired $\ell_2$-normalised dual-encoder VLM embeddings on the product hypersphere $\mathbb{S}^{d-1} \times \mathbb{S}^{d-1}$ via Riemannian flow matching with a single masked velocity field. A consistency result shows that, in the population limit, the trained network exposes the joint flow and both cross-modal conditional flows as valid Riemannian flow-matching velocity fields on their respective domains. We derive two quantities from this single model: a conditional retrieval entropy that quantifies aleatoric ambiguity with a decision-theoretic interpretation via a Fano-type bound, and a marginal-typicality epistemic score justified by an exact chain-rule decomposition of the joint NLL. This decomposition isolates a cross-modal pointwise-mutual-information term that is structurally discriminative rather than epistemic, and is empirically the only consistently uninformative standalone component. Empirically, the entropy tracks Recall@1 with near-ideal monotonic calibration across three retrieval benchmarks in both directions, and the marginal-typicality sum yields consistently calibrated selective accuracy across four zero-shot classification benchmarks.

Vision-Language-Action (VLA) models show strong potential for general-purpose robotic manipulation, yet their closed-loop reliability often degrades under local deployment conditions. Existing evaluations typically treat test episodes as independent zero-shot trials. However, real robots often operate repeatedly in the same or slowly changing environments, where successful executions provide environment-verified evidence of reliable behavior patterns. We study this persistent-deployment setting, asking whether a partially competent frozen VLA can improve its reliability by reusing its successful test-time experience. We propose an online success-memory guided test-time adaptation framework for generative VLAs. During deployment, the robot stores progress-calibrated successful observation-action segments in a long-term memory. At inference, it retrieves state-relevant action chunks, filters inconsistent candidates via trajectory-level consistency, and aggregates them into an elite action prior. To incorporate this prior into action generation, we introduce confidence-adaptive prior guidance, which injects the elite prior into an intermediate state of the flow-matching action sampler and adjusts the guidance strength based on retrieval confidence. This design allows the frozen VLA to exploit environment-specific successful experience while preserving observation-conditioned generative refinement. This retrieve-then-steer mechanism enables lightweight, non-parametric test-time adaptation without requiring parameter updates. Simulation and real-world experiments show improved task success and closed-loop stability, especially in long-horizon and multi-stage tasks.

Cameras and LiDAR degrade in rain, fog, and snow, while millimeter-wave radar remains largely unaffected. We align a radar encoder to frozen SigLIP vision embeddings and decode structured scene captions through a frozen vision-language model (VLM) with approximately 7M trainable parameters. On K-RADAR with held-out fog, light snow, and heavy snow sequences, all radar configurations outperform a camera baseline that collapses to over 90% hallucination. We identify a token-norm mismatch as the dominant failure mode when bridging radar to a frozen VLM and show that projector-output LayerNorm resolves it. Analysis of encoder complexity, caption format, and pooling strategy reveals tradeoffs that inform future radar-VLM pipeline design.

Adapting foundation models to medical segmentation typically requires either backbone fine-tuning or high-capacity task-specific decoders, both of which are difficult to fit reliably when annotations are scarce. We show that frozen DINOv3 features already contain useful structural and boundary cues for medical segmentation, and that the main bottleneck lies in how these features are read out. We propose DINO-MVR, a Multi-View Readout framework for annotation-efficient medical segmentation. DINO-MVR trains only lightweight MLP probes on features from the final three transformer blocks of a frozen DINOv3 backbone, without updating the backbone itself. At inference, each input is interpreted through complementary resolutions and test-time augmentations, whose probability maps are combined by entropy-weighted fusion and refined with simple spatial regularization. For volumetric inputs, Gaussian z-axis smoothing further improves inter-slice consistency. Under fixed evaluation protocols on endoscopy, dermoscopy, and MRI benchmarks, DINO-MVR achieves strong readout-only performance, including 0.895 Dice on Kvasir-SEG, 0.897 Dice on ISIC 2018, and 0.908 Dice on BraTS FLAIR whole-tumor segmentation. With only five annotated BraTS patients, it recovers 98.4% of the performance obtained by the 40-patient BraTS reference run. These results suggest that frozen self-supervised vision backbones can support accurate medical segmentation when paired with an effective multi-view readout.

Frozen vision foundation models do not merely extract features; they organize images through a learned coordinate system. We ask whether that coordinate system remains internally coherent within a single input. This leads to Representational Consistency: the study of whether a frozen foundation model represents one sample coherently across its semantic subregions. We introduce Dimensional Coactivation (DCA), a per-dimension instrument for measuring this coherence. DCA compares semantic regions by asking whether the same feature dimensions coactivate across them. Unlike classical similarity measures, it deliberately avoids centering, L2 normalization, and full Gram coupling. These operations are useful when comparing different models or distributions, but they are mismatched to the intra-sample setting, where the coordinate system is fixed and raw magnitude carries signal. Deepfake detection provides a natural validation task. Synthetic faces may reproduce plausible eyes, noses, and mouths while breaking the representational structure that links those regions in real faces. Using frozen DINOv3 features, DCA exposes this break: an eyes-mouth-nose fingerprint achieves 0.9106 AUC on CelebDF-v2 and 0.9289 on DFD under FF++ c23 cross-dataset transfer. The design is also sharply validated by ablation: reintroducing centering collapses CelebDF-v2 AUC to 0.459, L2 normalization reduces it to 0.862, and cross-dimension coupling reduces it to 0.478. Finally, replacing DINOv3 with FaRL collapses CelebDF-v2 AUC to 0.582. DCA therefore depends on a stable per-dimension coordinate system, not on region extraction alone. These results position DCA as an instrument for measuring intra-sample representational coherence in frozen foundation models, with deepfake detection as the first validation task.

Grounding traffic accidents in real CCTV footage is a rare-event problem where training on labeled accident video is often prohibited, yet accurate joint localization in time, space, and collision type is required. We present a no-fine-tuning pipeline that elicits this joint output from frozen vision-language models through two ideas. First, a coarse-to-fine two-pass decomposition: a full-video pass at 1 fps produces a coarse (t, x, y, c) tuple, then a second pass at 5 fps within a +/- 3 s window refines time and location, with two deterministic confidence gates that revert to the coarse estimate on boundary hedges or edge-clamped coordinates. Second, a specialist role assignment: Qwen3-VL-Plus handles grounding, Gemini 3.1 Flash-Lite handles typing on a centered video clip. On the ACCIDENT@CVPR 2026 benchmark (2,027 real CCTV videos) we reach ACC^S = 0.539 (95% CI [0.525, 0.553]): +0.127 over the benchmark paper's best-of-baselines oracle (0.412), +0.143 over the strongest single-VLM baseline (Molmo-7B, 0.396), and +0.250 over the naive baseline (0.289). The VLM path uses up to three API calls per video (17% fall back to physics on API failures); the full run costs ~$20.

While Vision-Language Models (VLMs) have achieved state-of-the-art performance in general visual tasks, their perceptual robustness remains remarkably brittle when confronted with optical illusions. These failures are often attributed to shortcut heuristics, where models prioritize linguistic priors and memorized prototypes over direct visual evidence. In this work, we propose Structured Qualitative Inference (SQI), a training-free, data-centric framework designed to fortify visual grounding in frozen VLMs. SQI addresses perceptual anomalies through three systematic modules: (1) Axiomatic Constraint Injection, which suppresses erroneous metric estimations and quantitative hallucinations; (2) Hierarchical Scene Decomposition, which decouples target visual manifolds from complex background distractors; and (3) Counterfactual Self-Verification, an adversarial reasoning step that mitigates confirmation bias. By orchestrating these qualitative constraints at inference time, SQI effectively aligns high-level linguistic reasoning with low-level visual perception. Our framework was evaluated on the DataCV 2026 Challenge (Task I: Classic Illusion Understanding), where it ranked 2nd place overall. Experimental results demonstrate that SQI not only significantly enhances accuracy across diverse illusion categories but also provides superior diagnostic interpretability without any model fine-tuning. Our success underscores the potential of structured qualitative grounding as a robust paradigm for developing next-generation, illusion-resistant vision-language systems.

Have you ever post-trained a generalist vision-language-action (VLA) policy on a small demonstration dataset, only to find that it stops responding to new instructions and is limited to behaviors observed during post-training? We identify this phenomenon as lock-in: after low-data, supervised fine-tuning (SFT), the policy becomes overly specialized to the post-training data and fails to generalize to novel instructions, manifesting as concept lock-in (fixation on training objects/attributes) and spatial lock-in (fixation on training spatial targets). Many existing remedies introduce additional supervision signals, such as those derived from foundation models or auxiliary objectives, or rely on augmented datasets to recover generalization. In this paper, we show that the policy's internal pre-trained knowledge is sufficient: DeLock mitigates lock-in by preserving visual grounding during post-training and applying test-time contrastive prompt guidance to steer the policy's denoising dynamics according to novel instructions. Across eight simulation and real-world evaluations, DeLock consistently outperforms strong baselines and matches or exceeds the performance of a state-of-the-art generalist policy post-trained with substantially more curated demonstrations.

Composed image retrieval, multi-turn composed image retrieval, and composed video retrieval all share a common paradigm: composing the reference visual with modification text to retrieve the desired target. Despite this shared structure, the three tasks have been studied in isolation, with no prior work proposing a unified framework, let alone a zero-shot solution. In this paper, we propose UniCVR, the first unified zero-shot composed visual retrieval framework that jointly addresses all three tasks without any task-specific human-annotated data. UniCVR strategically combines two complementary strengths: Multimodal Large Language Models (MLLMs) for compositional query understanding and Vision-Language Pre-trained (VLP) models for structured visual retrieval. Concretely, UniCVR operates in two stages. In Stage I, we train the MLLM as a compositional query embedder via contrastive learning on a curated multi-source dataset of approximately 3.5M samples, bridging the heterogeneous embedding spaces between the MLLM and the frozen VLP gallery encoder. A cluster-based hard negative sampling strategy is proposed to strengthen contrastive supervision. In Stage II, we introduce an MLLM-guided dual-level reranking mechanism that applies adaptive budgeted subset scoring to a small number of top-ranked candidates, and then exploits the resulting relevance signals through a dual-level re-scoring scheme, producing more accurate final rankings with minimal computational overhead. Extensive experiments across five benchmarks covering all three tasks demonstrate that UniCVR achieves cutting-edge performance, validating its effectiveness and generalizability. Our data and code will be released upon acceptance.

We present a system for automated detection, localization, and scoring of arrow punctures on 40,cm indoor archery target faces, trained on only 48 annotated photographs (5{,}084 punctures). Our pipeline combines three components: a color-based canonical rectification stage that maps perspective-distorted photographs into a standardized coordinate system where pixel distances correspond to known physical measurements; a frozen self-supervised vision transformer (DINOv3 ViT-L/16) paired with AnyUp guided feature upsampling to recover sub-millimeter spatial precision from $32 \times 32$ patch tokens; and lightweight CenterNet-style detection heads for arrow-center heatmap prediction. Only 3.8,M of 308,M total parameters are trainable. Across three cross-validation folds, we achieve a mean F1 score of $0.893 \pm 0.011$ and a mean localization error of $1.41 \pm 0.06$,mm, comparable to or better than prior fully-supervised approaches that require substantially more training data. An ablation study shows that the CenterNet offset regression head, typically essential for sub-pixel refinement, provides negligible detection improvement while degrading localization in our setting. This suggests that guided feature upsampling already resolves the spatial precision lost through patch tokenization. On downstream archery metrics, the system recovers per-image average arrow scores with a median error of 1.8% and group centroid positions to within a median of 4.00,mm. These results demonstrate that frozen foundation models with minimal task-specific adaptation offer a practical paradigm for dense prediction in small-data regimes.

Visual foundation models provide strong perceptual features for robotics, but their dense representations lack explicit object-level structure, limiting robustness and controllability in manipulation tasks. We propose STORM (Slot-based Task-aware Object-centric Representation for robotic Manipulation), a lightweight object-centric adaptation module that augments frozen visual foundation models with a small set of task-aware slots for robotic manipulation. Rather than fully tuning large backbones on the task, STORM employs an efficient two-stage training strategy: few layers of object-centric representation are first trained on top of the frozen backbone through visual--semantic pretraining using language embeddings, then jointly adapted with a downstream manipulation policy for task alignement. This staged learning prevents degenerate slot formation and preserves semantic consistency while aligning perception with task objectives. Experiments on object discovery benchmarks and robotic manipulation tasks show that STORM improves control performance and generalization to visual shifts (distractors, textures, lighting) compared to directly using frozen or fine-tuned foundation model features, or existing object-centric representations. STORM serves not only as an efficient mechanism for refining generic foundation model features, but also as a novel way of injecting beneficial structural and semantic bias into policy learning.