While large vision-language models (LVLMs) have shown impressive capabilities in generating plausible responses correlated with input visual contents, they still suffer from hallucinations, where the generated text inaccurately reflects visual contents. To address this, recent approaches apply contrastive decoding to calibrate the model's response via contrasting output distributions with original and visually distorted samples, demonstrating promising hallucination mitigation in a training-free manner. However, the potential of changing information in visual inputs is not well-explored, so a deeper investigation into the behaviors of visual contrastive decoding is of great interest. In this paper, we first explore various methods for contrastive decoding to change visual contents, including image downsampling and editing. Downsampling images reduces the detailed textual information while editing yields new contents in images, providing new aspects as visual contrastive samples. To further study benefits by using different contrastive samples, we analyze probability-level metrics, including entropy and distribution distance. Interestingly, the effect of these samples in mitigating hallucinations varies a lot across LVLMs and benchmarks. Based on our analysis, we propose a simple yet effective method to combine contrastive samples, offering a practical solution for applying contrastive decoding across various scenarios. Extensive experiments are conducted to validate the proposed fusion method among different benchmarks.

Text-to-image (T2I) models have shown remarkable progress, but their potential to generate harmful content remains a critical concern in the ML community. While various safety mechanisms have been developed, the field lacks systematic tools for evaluating their effectiveness against real-world misuse scenarios. In this work, we propose ICER, a novel red-teaming framework that leverages Large Language Models (LLMs) and a bandit optimization-based algorithm to generate interpretable and semantic meaningful problematic prompts by learning from past successful red-teaming attempts. Our ICER efficiently probes safety mechanisms across different T2I models without requiring internal access or additional training, making it broadly applicable to deployed systems. Through extensive experiments, we demonstrate that ICER significantly outperforms existing prompt attack methods in identifying model vulnerabilities while maintaining high semantic similarity with intended content. By uncovering that successful jailbreaking instances can systematically facilitate the discovery of new vulnerabilities, our work provides crucial insights for developing more robust safety mechanisms in T2I systems.

The recent development of Video-based Large Language Models (VideoLLMs), has significantly advanced video summarization by aligning video features and, in some cases, audio features with Large Language Models (LLMs). Each of these VideoLLMs possesses unique strengths and weaknesses. Many recent methods have required extensive fine-tuning to overcome the limitations of these models, which can be resource-intensive. In this work, we observe that the strengths of one VideoLLM can complement the weaknesses of another. Leveraging this insight, we propose a novel video summarization framework inspired by the Mixture of Experts (MoE) paradigm, which operates as an inference-time algorithm without requiring any form of fine-tuning. Our approach integrates multiple VideoLLMs to generate comprehensive and coherent textual summaries. It effectively combines visual and audio content, provides detailed background descriptions, and excels at identifying keyframes, which enables more semantically meaningful retrieval compared to traditional computer vision approaches that rely solely on visual information, all without the need for additional fine-tuning. Moreover, the resulting summaries enhance performance in downstream tasks such as summary video generation, either through keyframe selection or in combination with text-to-image models. Our language-driven approach offers a semantically rich alternative to conventional methods and provides flexibility to incorporate newer VideoLLMs, enhancing adaptability and performance in video summarization tasks.

Recent advancements in post-hoc and inherently interpretable methods have markedly enhanced the explanations of black box classifier models. These methods operate either through post-analysis or by integrating concept learning during model training. Although being effective in bridging the semantic gap between a model's latent space and human interpretation, these explanation methods only partially reveal the model's decision-making process. The outcome is typically limited to high-level semantics derived from the last feature map. We argue that the explanations lacking insights into the decision processes at low and mid-level features are neither fully faithful nor useful. Addressing this gap, we introduce the Multi-Level Concept Prototypes Classifier (MCPNet), an inherently interpretable model. MCPNet autonomously learns meaningful concept prototypes across multiple feature map levels using Centered Kernel Alignment (CKA) loss and an energy-based weighted PCA mechanism, and it does so without reliance on predefined concept labels. Further, we propose a novel classifier paradigm that learns and aligns multi-level concept prototype distributions for classification purposes via Class-aware Concept Distribution (CCD) loss. Our experiments reveal that our proposed MCPNet while being adaptable to various model architectures, offers comprehensive multi-level explanations while maintaining classification accuracy. Additionally, its concept distribution-based classification approach shows improved generalization capabilities in few-shot classification scenarios.

The recent renaissance of deep learning techniques has While image data starts to enjoy the simple-but-effective brought a magic leap to various fields, such as computer vision, self-supervised learning scheme built upon masking and natural language processing, and robotics. Learning self-reconstruction objective thanks to the introduction of from a large-scale labeled/supervised dataset, which is one tokenization procedure and vision transformer backbone, of the key factors leading to the success of deep learning, convolutional neural networks as another important and however, has now turned out to be a significant limitation widely-adopted architecture for image data, though having on its extensions to more fields. In addition to the expensive contrastive-learning techniques to drive the self-supervised cost of time and human resources to collect training learning, still face the difficulty of leveraging such straightforward datasets for different tasks and their corresponding labels, and general masking operation to benefit their the supervised learning scenario typically would suffer from learning process significantly. In this work, we aim to the issue of overfitting on the training dataset, thus leading alleviate the burden of including masking operation into to worse generalizability of the learnt models. These problems the contrastive-learning framework for convolutional neural bring challenges for the application of deep learning networks as an extra augmentation method. In addition techniques but also give rise to the research topic of selfsupervised to the additive but unwanted edges (between masked and learning, wherein it aims to learn to extract informative unmasked regions) as well as other adverse effects caused feature representations from an unlabelled dataset by the masking operations for ConvNets, which have been via leveraging the underlying structure of data and building discussed by prior works, we particularly identify the potential the supervisory signals from the data itself. The discovered problem where for one view in a contrastive samplepair representations are typically more general and can be further the randomly-sampled masking regions could be overly utilized or fine-tuned to various downstream tasks.

Text-to-image diffusion models, e.g. Stable Diffusion (SD), lately have shown remarkable ability in high-quality content generation, and become one of the representatives for the recent wave of transformative AI. Nevertheless, such advance comes with an intensifying concern about the misuse of this generative technology, especially for producing copyrighted or NSFW (i.e. not safe for work) images. Although efforts have been made to filter inappropriate images/prompts or remove undesirable concepts/styles via model fine-tuning, the reliability of these safety mechanisms against diversified problematic prompts remains largely unexplored. In this work, we propose Prompting4Debugging (P4D) as a debugging and red-teaming tool that automatically finds problematic prompts for diffusion models to test the reliability of a deployed safety mechanism. We demonstrate the efficacy of our P4D tool in uncovering new vulnerabilities of SD models with safety mechanisms. Particularly, our result shows that around half of prompts in existing safe prompting benchmarks which were originally considered "safe" can actually be manipulated to bypass many deployed safety mechanisms, including concept removal, negative prompt, and safety guidance. Our findings suggest that, without comprehensive testing, the evaluations on limited safe prompting benchmarks can lead to a false sense of safety for text-to-image models.

Most of the existing algorithms for zero-shot classification problems typically rely on the attribute-based semantic relations among categories to realize the classification of novel categories without observing any of their instances. However, training the zero-shot classification models still requires attribute labeling for each class (or even instance) in the training dataset, which is also expensive. To this end, in this paper, we bring up a new problem scenario: "Can we derive zero-shot learning for novel attribute detectors/classifiers and use them to automatically annotate the dataset for labeling efficiency?". Basically, given only a small set of detectors that are learned to recognize some manually annotated attributes (i.e., the seen attributes), we aim to synthesize the detectors of novel attributes in a zero-shot learning manner. Our proposed method, Zero-Shot Learning for Attributes (ZSLA), which is the first of its kind to the best of our knowledge, tackles this new research problem by applying the set operations to first decompose the seen attributes into their basic attributes and then recombine these basic attributes into the novel ones. Extensive experiments are conducted to verify the capacity of our synthesized detectors for accurately capturing the semantics of the novel attributes and show their superior performance in terms of detection and localization compared to other baseline approaches. Moreover, we demonstrate the application of automatic annotation using our synthesized detectors on Caltech-UCSD Birds-200-2011 dataset. Various generalized zero-shot classification algorithms trained upon the dataset re-annotated by ZSLA shows comparable performance with those trained with the manual ground-truth annotations. Please refer to our project page for source code: https://yuhsuanli.github.io/ZSLA/

On playing video games, different players usually have their own playstyles. Recently, there have been great improvements for the video game AIs on the playing strength. However, past researches for analyzing the behaviors of players still used heuristic rules or the behavior features with the game-environment support, thus being exhausted for the developers to define the features of discriminating various playstyles. In this paper, we propose the first metric for video game playstyles directly from the game observations and actions, without any prior specification on the playstyle in the target game. Our proposed method is built upon a novel scheme of learning discrete representations that can map game observations into latent discrete states, such that playstyles can be exhibited from these discrete states. Namely, we measure the playstyle distance based on game observations aligned to the same states. We demonstrate high playstyle accuracy of our metric in experiments on some video game platforms, including TORCS, RGSK, and seven Atari games, and for different agents including rule-based AI bots, learning-based AI bots, and human players.

There exist many forms of deep latent variable models, such as the variational autoencoder and adversarial autoencoder. Regardless of the specific class of model, there exists an implicit consensus that the latent distribution should be regularized towards the prior, even in the case where the prior distribution is learned. Upon investigating the effect of latent regularization on image generation our results indicate that in the case where a sufficiently expressive prior is learned, latent regularization is not necessary and may in fact be harmful insofar as image quality is concerned. We additionally investigate the benefit of learned priors on two common problems in computer vision: latent variable disentanglement, and diversity in image-to-image translation.