The reliability of VLM-as-a-Judge is critical for the automatic evaluation of vision-language models (VLMs).Despite recent progress, our analysis reveals that VLM-as-a-Judge often pays limited attention to the image when making decisions.Instead, they often blindly favor the more informative answer, even when they can recognize it conflicts with the image content.We call this problem informativeness bias, which significantly undermines judge reliability.To address it, we propose BIRCH (Balanced Informativeness and CoRrectness with a Truthful AnCHor), a judging paradigm that first corrects inconsistencies with the image content in candidate answers, and then compares the answers against this corrected version.This shifts the judge's focus from informativeness to image-grounded correctness.Experiments on multiple models and benchmarks show that BIRCH reduces informativeness bias by up to 17%, resulting in performance gains of up to 9.8%.Our work reveals an overlooked but fundamental flaw in current VLM-as-a-Judge systems and highlights the need for more principled designs.Preference: Answer A Answer A: The image does not provide explicit information about the name of the park.Answer B: The benches reside in Concord Pacific Place Park.The park's name is inscribed on the bench.

Recent advancements in prompt-based learning have significantly advanced image and video class-incremental learning. However, the prompts learned by these methods often fail to capture the diverse and informative characteristics of videos, and struggle to generalize effectively to future tasks and classes. To address these challenges, this paper proposes modeling the distribution of space-time prompts conditioned on the input video using a diffusion model. This generative approach allows the proposed model to naturally handle the diverse characteristics of videos, leading to more robust prompt learning and enhanced generalization capabilities. Additionally, we develop a simple yet effective mechanism to transfer the token relationship modeling capabilities of pre-trained image transformers to spatio-temporal modeling in videos. Our approach has been thoroughly evaluated across four established benchmarks, showing remarkable improvements over existing state-of-the-art methods in video class-incremental learning. Code is available at https://github.com/Renovamen/CoSTEP.

Recent studies reveal that vision-language models (VLMs) become more susceptible to harmful requests and jailbreak attacks after integrating the vision modality, exhibiting greater vulnerability than their text-only LLM backbones. To uncover the root cause of this phenomenon, we conduct an in-depth analysis and identify a key issue: multimodal inputs introduce an modality-induced activation shift toward a “safer” direction compared to their text-only counterparts, leading VLMs to systematically overestimate the safety of harmful inputs. We refer to this issue as safety perception distortion. To mitigate such distortion, we propose Activation Shift Disentanglement and Calibration (ShiftDC), a training-free method that decomposes and calibrates the modality-induced activation shift to reduce its impact on safety. By isolating and removing the safety-relevant component, ShiftDC restores the inherent safety alignment of the LLM backbone while preserving the vision-language capabilities of VLMs. Experiments demonstrate that ShiftDC significantly enhances safety alignment without impairing model utility.

Single-cell RNA sequencing (scRNA-seq) technology enables the analysis of gene expression in individual cells, allowing for a deeper exploration of heterogeneity in organisms and complex diseases. Cell clustering is a crucial step in single-cell analysis, enabling the identification of cellular heterogeneity. However, the high dimensionality, sparsity, and dropout events in single-cell data have brought enormous challenges to clustering analysis. Building on the proven success of deep generative models in learning meaningful representations from low-dimensional latent spaces, we introduce scDVAE, a novel deep generative approach that leverages a variational autoencoder with disentangled latent representations for single-cell clustering. First, each latent representation generated by the encoder is disentangled into clustering features and generative features. In this way, the clustering features can enhance the performance of the clustering task without interference from the generative task. Second, we employ a Student's t-mixture model as the prior distribution for the clustering features to enhance the robustness of our method against dropout events. In addition, we introduce a hybrid data augmentation strategy to generate augmented scRNA-seq data, which enhances dataset diversity while also helping to reduce noise. Our experimental studies on 10 real-world datasets demonstrate that scDVAE significantly improves clustering performance compared to state-of-the-art methods.

Continual learning aims to alleviate catastrophic forgetting when handling consecutive tasks under non-stationary distributions. Gradient-based meta-learning algorithms have shown the capability to implicitly solve the transfer-interference trade-off problem between different examples. However, they still suffer from the catastrophic forgetting problem in the setting of continual learning, since the past data of previous tasks are no longer available. In this work, we propose a novel efficient meta-learning algorithm for solving the online continual learning problem, where the regularization terms and learning rates are adapted to the Taylor approximation of the parameter's importance to mitigate forgetting. The proposed method expresses the gradient of the meta-loss in closed-form and thus avoid computing second-order derivative which is computationally inhibitable. We also use Proximal Gradient Descent to further improve computational efficiency and accuracy. Experiments on diverse benchmarks show that our method achieves better or on-par performance and much higher efficiency compared to the state-of-the-art approaches.

Image aesthetic captioning is a multi-modal task that is to generate aesthetic critiques for images. In contrast to common image captioning tasks, where different captions aimed at providing factual descriptions of a same image are always similar, captions with respect to different aesthetic attributes of the same image can be totally different in an aesthetic captioning task. Such inter-aspect differences are always overlooked, which leads to the lack of diversity and coherence of the captions generated by most of the existing image aesthetic captioning systems. In this paper, we propose a novel model to generate aesthetic captions for food images. Our model redefines food image aesthetic captioning as a compositional task that consists of two separated modules, i.e., a single-aspect captioning and an unsupervised text compression. The first module is guaranteed to generate the captions and learn feature representations of each aesthetic attribute. Then, the second module is supposed to study the associations among all feature representations and automatically aggregate captions of all aesthetic attributes to a final sentence. We also collect a dataset which contains pair-wise image-comment data related to six aesthetic attributes. Two new evaluation criteria are introduced to comprehensively assess the quality of the generated captions. Experiments on the dataset demonstrate the effectiveness of the proposed model.