Xuefeng Du

Xuefeng Du, Chaowei Xiao, Yixuan Li

The surge in applications of large language models (LLMs) has prompted concerns about the generation of misleading or fabricated information, known as hallucinations. Therefore, detecting hallucinations has become critical to maintaining trust in LLM-generated content. A primary challenge in learning a truthfulness classifier is the lack of a large amount of labeled truthful and hallucinated data. To address the challenge, we introduce HaloScope, a novel learning framework that leverages the unlabeled LLM generations in the wild for hallucination detection. Such unlabeled data arises freely upon deploying LLMs in the open world, and consists of both truthful and hallucinated information. To harness the unlabeled data, we present an automated membership estimation score for distinguishing between truthful and untruthful generations within unlabeled mixture data, thereby enabling the training of a binary truthfulness classifier on top. Importantly, our framework does not require extra data collection and human annotations, offering strong flexibility and practicality for real-world applications. Extensive experiments show that HaloScope can achieve superior hallucination detection performance, outperforming the competitive rivals by a significant margin. Code is available at https://github.com/deeplearningwisc/haloscope.

Xuefeng Du, Zhen Fang, Ilias Diakonikolas et al.

Using unlabeled data to regularize the machine learning models has demonstrated promise for improving safety and reliability in detecting out-of-distribution (OOD) data. Harnessing the power of unlabeled in-the-wild data is non-trivial due to the heterogeneity of both in-distribution (ID) and OOD data. This lack of a clean set of OOD samples poses significant challenges in learning an optimal OOD classifier. Currently, there is a lack of research on formally understanding how unlabeled data helps OOD detection. This paper bridges the gap by introducing a new learning framework SAL (Separate And Learn) that offers both strong theoretical guarantees and empirical effectiveness. The framework separates candidate outliers from the unlabeled data and then trains an OOD classifier using the candidate outliers and the labeled ID data. Theoretically, we provide rigorous error bounds from the lens of separability and learnability, formally justifying the two components in our algorithm. Our theory shows that SAL can separate the candidate outliers with small error rates, which leads to a generalization guarantee for the learned OOD classifier. Empirically, SAL achieves state-of-the-art performance on common benchmarks, reinforcing our theoretical insights. Code is publicly available at https://github.com/deeplearning-wisc/sal.

Haoyue Bai, Xuefeng Du, Katie Rainey et al.

Out-of-distribution (OOD) learning often relies heavily on statistical approaches or predefined assumptions about OOD data distributions, hindering their efficacy in addressing multifaceted challenges of OOD generalization and OOD detection in real-world deployment environments. This paper presents a novel framework for OOD learning with human feedback, which can provide invaluable insights into the nature of OOD shifts and guide effective model adaptation. Our framework capitalizes on the freely available unlabeled data in the wild that captures the environmental test-time OOD distributions under both covariate and semantic shifts. To harness such data, our key idea is to selectively provide human feedback and label a small number of informative samples from the wild data distribution, which are then used to train a multi-class classifier and an OOD detector. By exploiting human feedback, we enhance the robustness and reliability of machine learning models, equipping them with the capability to handle OOD scenarios with greater precision. We provide theoretical insights on the generalization error bounds to justify our algorithm. Extensive experiments show the superiority of our method, outperforming the current state-of-the-art by a significant margin.

Leitian Tao, Xuefeng Du, Sharon Li

Reward modeling, crucial for aligning large language models (LLMs) with human preferences, is often bottlenecked by the high cost of preference data. Existing textual data synthesis methods are computationally expensive. We propose a novel framework LENS for synthesizing preference data directly in the LLM's latent embedding space. Our method employs a Variational Autoencoder (VAE) to learn a structured latent representation of response embeddings. By performing controlled perturbations in this latent space and decoding back to the embedding space, we efficiently generate diverse, semantically consistent synthetic preference pairs, bypassing costly text generation and annotation. We provide theoretical guarantees that our synthesized pairs approximately preserve original preference ordering and improve reward model generalization. Empirically, our latent-space synthesis significantly outperforms text-based augmentation on standard benchmarks, achieving superior results while being 18x faster in generation and using a 16,000x smaller model. Our work offers a scalable and effective alternative for enhancing reward modeling through efficient data augmentation. Code is publicly available at https://github.com/deeplearning-wisc/lens

Seongheon Park, Xuefeng Du, Min-Hsuan Yeh et al.

Hallucinations in LLMs pose a significant concern to their safe deployment in real-world applications. Recent approaches have leveraged the latent space of LLMs for hallucination detection, but their embeddings, optimized for linguistic coherence rather than factual accuracy, often fail to clearly separate truthful and hallucinated content. To this end, we propose the Truthfulness Separator Vector (TSV), a lightweight and flexible steering vector that reshapes the LLM's representation space during inference to enhance the separation between truthful and hallucinated outputs, without altering model parameters. Our two-stage framework first trains TSV on a small set of labeled exemplars to form compact and well-separated clusters. It then augments the exemplar set with unlabeled LLM generations, employing an optimal transport-based algorithm for pseudo-labeling combined with a confidence-based filtering process. Extensive experiments demonstrate that TSV achieves state-of-the-art performance with minimal labeled data, exhibiting strong generalization across datasets and providing a practical solution for real-world LLM applications.

Changdae Oh, Zhen Fang, Shawn Im et al.

Multimodal large language models (MLLMs) have shown promising capabilities but struggle under distribution shifts, where evaluation data differ from instruction tuning distributions. Although previous works have provided empirical evaluations, we argue that establishing a formal framework that can characterize and quantify the risk of MLLMs is necessary to ensure the safe and reliable application of MLLMs in the real world. By taking an information-theoretic perspective, we propose the first theoretical framework that enables the quantification of the maximum risk of MLLMs under distribution shifts. Central to our framework is the introduction of Effective Mutual Information (EMI), a principled metric that quantifies the relevance between input queries and model responses. We derive an upper bound for the EMI difference between in-distribution (ID) and out-of-distribution (OOD) data, connecting it to visual and textual distributional discrepancies. Extensive experiments on real benchmark datasets, spanning 61 shift scenarios, empirically validate our theoretical insights.

Xuefeng Du, Haohan Wang, Zhenxi Zhu et al.

Motivation: Cryo-Electron Tomography (cryo-ET) is a 3D bioimaging tool that visualizes the structural and spatial organization of macromolecules at a near-native state in single cells, which has broad applications in life science. However, the systematic structural recognition and recovery of macromolecules captured by cryo-ET are difficult due to high structural complexity and imaging limits. Deep learning based subtomogram classification have played critical roles for such tasks. As supervised approaches, however, their performance relies on sufficient and laborious annotation on a large training dataset. Results: To alleviate this major labeling burden, we proposed a Hybrid Active Learning (HAL) framework for querying subtomograms for labelling from a large unlabeled subtomogram pool. Firstly, HAL adopts uncertainty sampling to select the subtomograms that have the most uncertain predictions. Moreover, to mitigate the sampling bias caused by such strategy, a discriminator is introduced to judge if a certain subtomogram is labeled or unlabeled and subsequently the model queries the subtomogram that have higher probabilities to be unlabeled. Additionally, HAL introduces a subset sampling strategy to improve the diversity of the query set, so that the information overlap is decreased between the queried batches and the algorithmic efficiency is improved. Our experiments on subtomogram classification tasks using both simulated and real data demonstrate that we can achieve comparable testing performance (on average only 3% accuracy drop) by using less than 30% of the labeled subtomograms, which shows a very promising result for subtomogram classification task with limited labeling resources.

Xuefeng Du, Jingfeng Zhang, Bo Han et al.

In adversarial training (AT), the main focus has been the objective and optimizer while the model has been less studied, so that the models being used are still those classic ones in standard training (ST). Classic network architectures (NAs) are generally worse than searched NAs in ST, which should be the same in AT. In this paper, we argue that NA and AT cannot be handled independently, since given a dataset, the optimal NA in ST would be no longer optimal in AT. That being said, AT is time-consuming itself; if we directly search NAs in AT over large search spaces, the computation will be practically infeasible. Thus, we propose a diverse-structured network (DS-Net), to significantly reduce the size of the search space: instead of low-level operations, we only consider predefined atomic blocks, where an atomic block is a time-tested building block like the residual block. There are only a few atomic blocks and thus we can weight all atomic blocks rather than find the best one in a searched block of DS-Net, which is an essential trade-off between exploring diverse structures and exploiting the best structures. Empirical results demonstrate the advantages of DS-Net, i.e., weighting the atomic blocks.

Xuefeng Du, Pengtao Xie

In human learning, an effective learning methodology is small-group learning: a small group of students work together towards the same learning objective, where they express their understanding of a topic to their peers, compare their ideas, and help each other to trouble-shoot problems. In this paper, we aim to investigate whether this human learning method can be borrowed to train better machine learning models, by developing a novel ML framework -- small-group learning (SGL). In our framework, a group of learners (ML models) with different model architectures collaboratively help each other to learn by leveraging their complementary advantages. Specifically, each learner uses its intermediately trained model to generate a pseudo-labeled dataset and re-trains its model using pseudo-labeled datasets generated by other learners. SGL is formulated as a multi-level optimization framework consisting of three learning stages: each learner trains a model independently and uses this model to perform pseudo-labeling; each learner trains another model using datasets pseudo-labeled by other learners; learners improve their architectures by minimizing validation losses. An efficient algorithm is developed to solve the multi-level optimization problem. We apply SGL for neural architecture search. Results on CIFAR-100, CIFAR-10, and ImageNet demonstrate the effectiveness of our method.

Lin Lan, Pinghui Wang, Xuefeng Du et al.

We study the problem of node classification on graphs with few-shot novel labels, which has two distinctive properties: (1) There are novel labels to emerge in the graph; (2) The novel labels have only a few representative nodes for training a classifier. The study of this problem is instructive and corresponds to many applications such as recommendations for newly formed groups with only a few users in online social networks. To cope with this problem, we propose a novel Meta Transformed Network Embedding framework (MetaTNE), which consists of three modules: (1) A \emph{structural module} provides each node a latent representation according to the graph structure. (2) A \emph{meta-learning module} captures the relationships between the graph structure and the node labels as prior knowledge in a meta-learning manner. Additionally, we introduce an \emph{embedding transformation function} that remedies the deficiency of the straightforward use of meta-learning. Inherently, the meta-learned prior knowledge can be used to facilitate the learning of few-shot novel labels. (3) An \emph{optimization module} employs a simple yet effective scheduling strategy to train the above two modules with a balance between graph structure learning and meta-learning. Experiments on four real-world datasets show that MetaTNE brings a huge improvement over the state-of-the-art methods.

Huikai Shao, Dexing Zhong, Xuefeng Du

Deep palmprint recognition has become an emerging issue with great potential for personal authentication on handheld and wearable consumer devices. Previous studies of palmprint recognition are mainly based on constrained datasets collected by dedicated devices in controlled environments, which has to reduce the flexibility and convenience. In addition, general deep palmprint recognition algorithms are often too heavy to meet the real-time requirements of embedded system. In this paper, a new palmprint benchmark is established, which consists of more than 20,000 images collected by 5 brands of smart phones in an unconstrained manner. Each image has been manually labeled with 14 key points for region of interest (ROI) extraction. Further, the approach called Deep Distillation Hashing (DDH) is proposed as benchmark for efficient deep palmprint recognition. Palmprint images are converted to binary codes to improve the efficiency of feature matching. Derived from knowledge distillation, novel distillation loss functions are constructed to compress deep model to further improve the efficiency of feature extraction on light network. Comprehensive experiments are conducted on both constrained and unconstrained palmprint databases. Using DDH, the accuracy of palmprint identification can be increased by up to 11.37%, and the Equal Error Rate (EER) of palmprint verification can be reduced by up to 3.11%. The results indicate the feasibility of our database, and DDH can outperform other baselines to achieve the state-of-the-art performance. The collected dataset and related source codes are publicly available at http://gr.xjtu.edu.cn/web/bell/resource.