VidHal: Benchmarking Temporal Hallucinations in Vision LLMs

Vision Large Language Models (VLLMs) are widely acknowledged to be prone to hallucination. Existing research addressing this problem has primarily been confined to image inputs, with limited exploration of video-based hallucinations. Furthermore, current evaluation methods fail to capture nuanced errors in generated responses, which are often exacerbated by the rich spatiotemporal dynamics of videos. To address this, we introduce VidHal, a benchmark specially designed to evaluate video-based hallucinations in VLLMs. VidHal is constructed by bootstrapping video instances across common temporal aspects. A defining feature of our benchmark lies in the careful creation of captions which represent varying levels of hallucination associated with each video. To enable fine-grained evaluation, we propose a novel caption ordering task requiring VLLMs to rank captions by hallucinatory extent. We conduct extensive experiments on VidHal and comprehensively evaluate a broad selection of models. Our results uncover significant limitations in existing VLLMs regarding hallucination generation. Through our benchmark, we aim to inspire further research on 1) holistic understanding of VLLM capabilities, particularly regarding hallucination, and 2) extensive development of advanced VLLMs to alleviate this problem.

VideoAutoArena: An Automated Arena for Evaluating Large Multimodal Models in Video Analysis through User Simulation

Large multimodal models (LMMs) with advanced video analysis capabilities have recently garnered significant attention. However, most evaluations rely on traditional methods like multiple-choice questions in benchmarks such as VideoMME and LongVideoBench, which are prone to lack the depth needed to capture the complex demands of real-world users. To address this limitation-and due to the prohibitive cost and slow pace of human annotation for video tasks-we introduce VideoAutoArena, an arena-style benchmark inspired by LMSYS Chatbot Arena's framework, designed to automatically assess LMMs' video analysis abilities. VideoAutoArena utilizes user simulation to generate open-ended, adaptive questions that rigorously assess model performance in video understanding. The benchmark features an automated, scalable evaluation framework, incorporating a modified ELO Rating System for fair and continuous comparisons across multiple LMMs. To validate our automated judging system, we construct a 'gold standard' using a carefully curated subset of human annotations, demonstrating that our arena strongly aligns with human judgment while maintaining scalability. Additionally, we introduce a fault-driven evolution strategy, progressively increasing question complexity to push models toward handling more challenging video analysis scenarios. Experimental results demonstrate that VideoAutoArena effectively differentiates among state-of-the-art LMMs, providing insights into model strengths and areas for improvement. To further streamline our evaluation, we introduce VideoAutoBench as an auxiliary benchmark, where human annotators label winners in a subset of VideoAutoArena battles. We use GPT-4o as a judge to compare responses against these human-validated answers. Together, VideoAutoArena and VideoAutoBench offer a cost-effective, and scalable framework for evaluating LMMs in user-centric video analysis.

Joint Vision-Language Social Bias Removal for CLIP

Vision-Language (V-L) pre-trained models such as CLIP show prominent capabilities in various downstream tasks. Despite this promise, V-L models are notoriously limited by their inherent social biases. A typical demonstration is that V-L models often produce biased predictions against specific groups of people, significantly undermining their real-world applicability. Existing approaches endeavor to mitigate the social bias problem in V-L models by removing biased attribute information from model embeddings. However, after our revisiting of these methods, we find that their bias removal is frequently accompanied by greatly compromised V-L alignment capabilities. We then reveal that this performance degradation stems from the unbalanced debiasing in image and text embeddings. To address this issue, we propose a novel V-L debiasing framework to align image and text biases followed by removing them from both modalities. By doing so, our method achieves multi-modal bias mitigation while maintaining the V-L alignment in the debiased embeddings. Additionally, we advocate a new evaluation protocol that can 1) holistically quantify the model debiasing and V-L alignment ability, and 2) evaluate the generalization of social bias removal models. We believe this work will offer new insights and guidance for future studies addressing the social bias problem in CLIP.

SCAN: Bootstrapping Contrastive Pre-training for Data Efficiency

While contrastive pre-training is widely employed, its data efficiency problem has remained relatively under-explored thus far. Existing methods often rely on static coreset selection algorithms to pre-identify important data for training. However, this static nature renders them unable to dynamically track the data usefulness throughout pre-training, leading to subpar pre-trained models. To address this challenge, our paper introduces a novel dynamic bootstrapping dataset pruning method. It involves pruning data preparation followed by dataset mutation operations, both of which undergo iterative and dynamic updates. We apply this method to two prevalent contrastive pre-training frameworks: \textbf{CLIP} and \textbf{MoCo}, representing vision-language and vision-centric domains, respectively. In particular, we individually pre-train seven CLIP models on two large-scale image-text pair datasets, and two MoCo models on the ImageNet dataset, resulting in a total of 16 pre-trained models. With a data pruning rate of 30-35\% across all 16 models, our method exhibits only marginal performance degradation (less than \textbf{1\%} on average) compared to corresponding models trained on the full dataset counterparts across various downstream datasets, and also surpasses several baselines with a large performance margin. Additionally, the byproduct from our method, \ie coresets derived from the original datasets after pre-training, also demonstrates significant superiority in terms of downstream performance over other static coreset selection approaches.

The VLLM Safety Paradox: Dual Ease in Jailbreak Attack and Defense

The vulnerability of Vision Large Language Models (VLLMs) to jailbreak attacks appears as no surprise. However, recent defense mechanisms against these attacks have reached near-saturation performance on benchmarks, often with minimal effort. This simultaneous high performance in both attack and defense presents a perplexing paradox. Resolving it is critical for advancing the development of trustworthy models. To address this research gap, we first investigate why VLLMs are prone to these attacks. We then make a key observation: existing defense mechanisms suffer from an \textbf{over-prudence} problem, resulting in unexpected abstention even in the presence of benign inputs. Additionally, we find that the two representative evaluation methods for jailbreak often exhibit chance agreement. This limitation makes it potentially misleading when evaluating attack strategies or defense mechanisms. Beyond these empirical observations, our another contribution in this work is to repurpose the guardrails of LLMs on the shelf, as an effective alternative detector prior to VLLM response. We believe these findings offer useful insights to rethink the foundational development of VLLM safety with respect to benchmark datasets, evaluation methods, and defense strategies.

UnStar: Unlearning with Self-Taught Anti-Sample Reasoning for LLMs

The key components of machine learning are data samples for training, model for learning patterns, and loss function for optimizing accuracy. Analogously, unlearning can potentially be achieved through anti-data samples (or anti-samples), unlearning method, and reversed loss function. While prior research has explored unlearning methods and reversed loss functions, the potential of anti-samples remains largely untapped. In this paper, we introduce UnSTAR: Unlearning with Self-Taught Anti-Sample Reasoning for large language models (LLMs). Our contributions are threefold; first, we propose a novel concept of anti-sample-induced unlearning; second, we generate anti-samples by leveraging misleading rationales, which help reverse learned associations and accelerate the unlearning process; and third, we enable fine-grained targeted unlearning, allowing for the selective removal of specific associations without impacting related knowledge - something not achievable by previous works. Results demonstrate that anti-samples offer an efficient, targeted unlearning strategy for LLMs, opening new avenues for privacy-preserving machine learning and model modification.

Strong Preferences Affect the Robustness of Value Alignment

Value alignment, which aims to ensure that large language models (LLMs) and other AI agents behave in accordance with human values, is critical for ensuring safety and trustworthiness of these systems. A key component of value alignment is the modeling of human preferences as a representation of human values. In this paper, we investigate the robustness of value alignment by examining the sensitivity of preference models. Specifically, we ask: how do changes in the probabilities of some preferences affect the predictions of these models for other preferences? To answer this question, we theoretically analyze the robustness of widely used preference models by examining their sensitivities to minor changes in preferences they model. Our findings reveal that, in the Bradley-Terry and the Placket-Luce model, the probability of a preference can change significantly as other preferences change, especially when these preferences are dominant (i.e., with probabilities near 0 or 1). We identify specific conditions where this sensitivity becomes significant for these models and discuss the practical implications for the robustness and safety of value alignment in AI systems.

STAR: Skeleton-aware Text-based 4D Avatar Generation with In-Network Motion Retargeting

The creation of 4D avatars (i.e., animated 3D avatars) from text description typically uses text-to-image (T2I) diffusion models to synthesize 3D avatars in the canonical space and subsequently applies animation with target motions. However, such an optimization-by-animation paradigm has several drawbacks. (1) For pose-agnostic optimization, the rendered images in canonical pose for naive Score Distillation Sampling (SDS) exhibit domain gap and cannot preserve view-consistency using only T2I priors, and (2) For post hoc animation, simply applying the source motions to target 3D avatars yields translation artifacts and misalignment. To address these issues, we propose Skeleton-aware Text-based 4D Avatar generation with in-network motion Retargeting (STAR). STAR considers the geometry and skeleton differences between the template mesh and target avatar, and corrects the mismatched source motion by resorting to the pretrained motion retargeting techniques. With the informatively retargeted and occlusion-aware skeleton, we embrace the skeleton-conditioned T2I and text-to-video (T2V) priors, and propose a hybrid SDS module to coherently provide multi-view and frame-consistent supervision signals. Hence, STAR can progressively optimize the geometry, texture, and motion in an end-to-end manner. The quantitative and qualitative experiments demonstrate our proposed STAR can synthesize high-quality 4D avatars with vivid animations that align well with the text description. Additional ablation studies shows the contributions of each component in STAR. The source code and demos are available at: \href{https://star-avatar.github.io}{https://star-avatar.github.io}.

Do Vision-Language Transformers Exhibit Visual Commonsense? An Empirical Study of VCR

Visual Commonsense Reasoning (VCR) calls for explanatory reasoning behind question answering over visual scenes. To achieve this goal, a model is required to provide an acceptable rationale as the reason for the predicted answers. Progress on the benchmark dataset stems largely from the recent advancement of Vision-Language Transformers (VL Transformers). These models are first pre-trained on some generic large-scale vision-text datasets, and then the learned representations are transferred to the downstream VCR task. Despite their attractive performance, this paper posits that the VL Transformers do not exhibit visual commonsense, which is the key to VCR. In particular, our empirical results pinpoint several shortcomings of existing VL Transformers: small gains from pre-training, unexpected language bias, limited model architecture for the two inseparable sub-tasks, and neglect of the important object-tag correlation. With these findings, we tentatively suggest some future directions from the aspect of dataset, evaluation metric, and training tricks. We believe this work could make researchers revisit the intuition and goals of VCR, and thus help tackle the remaining challenges in visual reasoning.

Multi-Modal Recommendation Unlearning

Unlearning methods for recommender systems (RS) have emerged to address privacy issues and concerns about legal compliance. However, evolving user preferences and content licensing issues still remain unaddressed. This is particularly true in case of multi-modal recommender systems (MMRS), which aim to accommodate the growing influence of multi-modal information on user preferences. Previous unlearning methods for RS are inapplicable to MMRS due to incompatibility of multi-modal user-item behavior data graph with the matrix based representation of RS. Partitioning based methods degrade recommendation performance and incur significant overhead costs during aggregation. This paper introduces MMRecUN, a new framework for multi-modal recommendation unlearning, which, to the best of our knowledge, is the first attempt in this direction. Given the trained recommendation model and marked forget data, we devise Reverse Bayesian Personalized Ranking (BPR) objective to force the model to forget it. MMRecUN employs both reverse and forward BPR loss mechanisms to selectively attenuate the impact of interactions within the forget set while concurrently reinforcing the significance of interactions within the retain set. Our experiments demonstrate that MMRecUN outperforms baseline methods across various unlearning requests when evaluated on benchmark multi-modal recommender datasets. MMRecUN achieves recall performance improvements of up to $\mathbf{49.85%}$ compared to the baseline methods. It is up to $\mathbf{1.3}\times$ faster than the \textsc{Gold} model, which is trained on retain data from scratch. MMRecUN offers advantages such as superior performance in removing target elements, preservation of performance for retained elements, and zero overhead costs in comparison to previous methods.