Understanding human preferences is crucial for improving foundation models and building personalized AI systems. However, preferences are inherently diverse and complex, making it difficult for traditional reward models to capture their full range. While fine-grained preference data can help, collecting it is expensive and hard to scale. In this paper, we introduce Decomposed Reward Models (DRMs), a novel approach that extracts diverse human preferences from binary comparisons without requiring fine-grained annotations. Our key insight is to represent human preferences as vectors and analyze them using Principal Component Analysis (PCA). By constructing a dataset of embedding differences between preferred and rejected responses, DRMs identify orthogonal basis vectors that capture distinct aspects of preference. These decomposed rewards can be flexibly combined to align with different user needs, offering an interpretable and scalable alternative to traditional reward models. We demonstrate that DRMs effectively extract meaningful preference dimensions (e.g., helpfulness, safety, humor) and adapt to new users without additional training. Our results highlight DRMs as a powerful framework for personalized and interpretable LLM alignment.

Recent studies show that diffusion models (DMs) are vulnerable to backdoor attacks. Existing backdoor attacks impose unconcealed triggers (e.g., a gray box and eyeglasses) that contain evident patterns, rendering remarkable attack effects yet easy detection upon human inspection and defensive algorithms. While it is possible to improve stealthiness by reducing the strength of the backdoor, doing so can significantly compromise its generality and effectiveness. In this paper, we propose UIBDiffusion, the universal imperceptible backdoor attack for diffusion models, which allows us to achieve superior attack and generation performance while evading state-of-the-art defenses. We propose a novel trigger generation approach based on universal adversarial perturbations (UAPs) and reveal that such perturbations, which are initially devised for fooling pre-trained discriminative models, can be adapted as potent imperceptible backdoor triggers for DMs. We evaluate UIBDiffusion on multiple types of DMs with different kinds of samplers across various datasets and targets. Experimental results demonstrate that UIBDiffusion brings three advantages: 1) Universality, the imperceptible trigger is universal (i.e., image and model agnostic) where a single trigger is effective to any images and all diffusion models with different samplers; 2) Utility, it achieves comparable generation quality (e.g., FID) and even better attack success rate (i.e., ASR) at low poison rates compared to the prior works; and 3) Undetectability, UIBDiffusion is plausible to human perception and can bypass Elijah and TERD, the SOTA defenses against backdoors for DMs. We will release our backdoor triggers and code.

We introduce Non-Euclidean-MDS (Neuc-MDS), an extension of classical Multidimensional Scaling (MDS) that accommodates non-Euclidean and non-metric inputs. The main idea is to generalize the standard inner product to symmetric bilinear forms to utilize the negative eigenvalues of dissimilarity Gram matrices. Neuc-MDS efficiently optimizes the choice of (both positive and negative) eigenvalues of the dissimilarity Gram matrix to reduce STRESS, the sum of squared pairwise error. We provide an in-depth error analysis and proofs of the optimality in minimizing lower bounds of STRESS. We demonstrate Neuc-MDS's ability to address limitations of classical MDS raised by prior research, and test it on various synthetic and real-world datasets in comparison with both linear and non-linear dimension reduction methods.

We consider the problem of multi-objective alignment of foundation models with human preferences, which is a critical step towards helpful and harmless AI systems. However, it is generally costly and unstable to fine-tune large foundation models using reinforcement learning (RL), and the multi-dimensionality, heterogeneity, and conflicting nature of human preferences further complicate the alignment process. In this paper, we introduce Rewards-in-Context (RiC), which conditions the response of a foundation model on multiple rewards in its prompt context and applies supervised fine-tuning for alignment. The salient features of RiC are simplicity and adaptivity, as it only requires supervised fine-tuning of a single foundation model and supports dynamic adjustment for user preferences during inference time. Inspired by the analytical solution of an abstracted convex optimization problem, our dynamic inference-time adjustment method approaches the Pareto-optimal solution for multiple objectives. Empirical evidence demonstrates the efficacy of our method in aligning both Large Language Models (LLMs) and diffusion models to accommodate diverse rewards with only around 10% GPU hours compared with multi-objective RL baseline.

Off-road autonomy validation presents unique challenges due to the unpredictable and dynamic nature of off-road environments. Traditional methods focusing on sequentially sweeping across the parameter space for variability analysis struggle to comprehensively assess the performance and safety of off-road autonomous systems within the imposed time constraints. This paper proposes leveraging scalable digital twin simulations within high-performance computing (HPC) clusters to address this challenge. By harnessing the computational power of HPC clusters, our approach aims to provide a scalable and efficient means to validate off-road autonomy algorithms, enabling rapid iteration and testing of autonomy algorithms under various conditions. We demonstrate the effectiveness of our framework through performance evaluations of the HPC cluster in terms of simulation parallelization and present the systematic variability analysis of a candidate off-road autonomy algorithm to identify potential vulnerabilities in the autonomy stack's perception, planning and control modules.

In the field of portrait video generation, the use of single images to generate portrait videos has become increasingly prevalent. A common approach involves leveraging generative models to enhance adapters for controlled generation. However, control signals (e.g., text, audio, reference image, pose, depth map, etc.) can vary in strength. Among these, weaker conditions often struggle to be effective due to interference from stronger conditions, posing a challenge in balancing these conditions. In our work on portrait video generation, we identified audio signals as particularly weak, often overshadowed by stronger signals such as facial pose and reference image. However, direct training with weak signals often leads to difficulties in convergence. To address this, we propose V-Express, a simple method that balances different control signals through the progressive training and the conditional dropout operation. Our method gradually enables effective control by weak conditions, thereby achieving generation capabilities that simultaneously take into account the facial pose, reference image, and audio. The experimental results demonstrate that our method can effectively generate portrait videos controlled by audio. Furthermore, a potential solution is provided for the simultaneous and effective use of conditions of varying strengths.

Current soft prompt methods yield limited performance when applied to small-sized models (fewer than a billion parameters). Deep prompt-tuning, which entails prepending parameters in each layer for enhanced efficacy, presents a solution for prompting small-sized models, albeit requiring carefully designed implementation. In this paper, we introduce the Lottery Ticket Prompt-learning (LTP) framework that integrates winning tickets with soft prompts. The LTP offers a simpler implementation and requires only a one-time execution. We demonstrate LTP on cross-lingual tasks, where prior works rely on external tools like human-designed multilingual templates and bilingual dictionaries, which may not be feasible in a low-resource regime. Specifically, we select a subset of parameters that have been changed the most during the fine-tuning with the Masked Language Modeling objective. Then, we prepend soft prompts to the original pre-trained language model and only update the selected parameters together with prompt-related parameters when adapting to the downstream tasks. We verify the effectiveness of our LTP framework on cross-lingual tasks, specifically targeting low-resource languages. Our approach outperforms the baselines by only updating 20\% of the original parameters.

Deep neural networks have demonstrated remarkable performance in various tasks. With a growing need for sparse deep learning, model compression techniques, especially pruning, have gained significant attention. However, conventional pruning techniques can inadvertently exacerbate algorithmic bias, resulting in unequal predictions. To address this, we define a fair pruning task where a sparse model is derived subject to fairness requirements. In particular, we propose a framework to jointly optimize the pruning mask and weight update processes with fairness constraints. This framework is engineered to compress models that maintain performance while ensuring fairness in a single execution. To this end, we formulate the fair pruning problem as a novel constrained bi-level optimization task and derive efficient and effective solving strategies. We design experiments spanning various datasets and settings to validate our proposed method. Our empirical analysis contrasts our framework with several mainstream pruning strategies, emphasizing our method's superiority in maintaining model fairness, performance, and efficiency.

Large Language Models (LLMs), representing a significant achievement in artificial intelligence (AI) research, have demonstrated their ability in a multitude of tasks. This project aims to explore the capabilities of GPT-3.5, a leading example of LLMs, in processing the sentiment analysis of Internet memes. Memes, which include both verbal and visual aspects, act as a powerful yet complex tool for expressing ideas and sentiments, demanding an understanding of societal norms and cultural contexts. Notably, the detection and moderation of hateful memes pose a significant challenge due to their implicit offensive nature. This project investigates GPT's proficiency in such subjective tasks, revealing its strengths and potential limitations. The tasks include the classification of meme sentiment, determination of humor type, and detection of implicit hate in memes. The performance evaluation, using datasets from SemEval-2020 Task 8 and Facebook hateful memes, offers a comparative understanding of GPT responses against human annotations. Despite GPT's remarkable progress, our findings underscore the challenges faced by these models in handling subjective tasks, which are rooted in their inherent limitations including contextual understanding, interpretation of implicit meanings, and data biases. This research contributes to the broader discourse on the applicability of AI in handling complex, context-dependent tasks, and offers valuable insights for future advancements.

Clustering continues to be a significant and challenging task. Recent studies have demonstrated impressive results by applying clustering to feature representations acquired through self-supervised learning, particularly on small datasets. However, when dealing with datasets containing a large number of clusters, such as ImageNet, current methods struggle to achieve satisfactory clustering performance. In this paper, we introduce a novel method called Contrastive representation Disentanglement for Clustering (CDC) that leverages contrastive learning to directly disentangle the feature representation for clustering. In CDC, we decompose the representation into two distinct components: one component encodes categorical information under an equipartition constraint, and the other component captures instance-specific factors. To train our model, we propose a contrastive loss that effectively utilizes both components of the representation. We conduct a theoretical analysis of the proposed loss and highlight how it assigns different weights to negative samples during the process of disentangling the feature representation. Further analysis of the gradients reveals that larger weights emphasize a stronger focus on hard negative samples. As a result, the proposed loss exhibits strong expressiveness, enabling efficient disentanglement of categorical information. Through experimental evaluation on various benchmark datasets, our method demonstrates either state-of-the-art or highly competitive clustering performance. Notably, on the complete ImageNet dataset, we achieve an accuracy of 53.4%, surpassing existing methods by a substantial margin of +10.2%.