As synthetic data proliferates across the Internet, it is often reused to train successive generations of generative models. This creates a "self-consuming loop" that can lead to training instability or *model collapse*. Common strategies to address the issue---such as accumulating historical training data or injecting fresh real data---either increase computational cost or require expensive human annotation. In this paper, we empirically analyze the latent space dynamics of self-consuming diffusion models and observe that the low-dimensional structure of latent representations extracted from synthetic data degrade over generations. Based on this insight, we propose *Latent Space Filtering* (LSF), a novel approach that mitigates model collapse by filtering out less realistic synthetic data from mixed datasets. Theoretically, we present a framework that connects latent space degradation to empirical observations. Experimentally, we show that LSF consistently outperforms existing baselines across multiple real-world datasets, effectively mitigating model collapse without increasing training cost or relying on human annotation.

The rapid advancement of large language models (LLMs) has led to growing interest in using synthetic data to train future models.However, this creates a self-consuming retraining loop, where models are trained on their own outputs and may cause performance drops and induce emerging biases.In real-world applications, previously deployed LLMs may influence the data they generate, leading to a dynamic system driven by user feedback.For example, if a model continues to underserve users from a group, less query data will be collected from this particular demographic of users.In this study, we introduce the concept of Self-Consuming Performative Loop (SCPL) and investigate the role of synthetic data in shaping bias during these dynamic iterative training processes under controlled performative feedback.This controlled setting is motivated by the inaccessibility of real-world user preference data from dynamic production systems, and enables us to isolate and analyze feedback-driven bias evolution in a principled manner.We focus on two types of loops, including the typical retraining setting and the incremental finetuning setting, which is largely underexplored.Through experiments on three real-world tasks, we find that the performative loop increases preference bias and decreases disparate bias.We design a reward-based rejection sampling strategy to mitigate the bias, moving towards more trustworthy self-improving systems.

Instruction tuning is critical for adapting large language models (LLMs) to downstream tasks, and recent studies have demonstrated that small amounts of human-curated data can outperform larger datasets, challenging traditional data scaling laws. While LLM-based data quality rating systems offer a cost-effective alternative to human annotation, they often suffer from inaccuracies and biases, even in powerful models like GPT-4. In this work, we introduce $DS^2$, a **D**iversity-aware **S**core curation method for **D**ata **S**election. By systematically modeling error patterns through a score transition matrix, $DS^2$ corrects LLM-based scores and promotes diversity in the selected data samples. Our approach shows that a curated subset (just 3.3\% of the original dataset) outperforms full-scale datasets (300k samples) across various machine-alignment benchmarks, and matches or surpasses human-aligned datasets such as LIMA with the same sample size (1k samples). These findings challenge conventional data scaling assumptions, highlighting that redundant, low-quality samples can degrade performance and reaffirming that ``more can be less''.

Unlearning in Large Language Models (LLMs) is essential for ensuring ethical and responsible AI use, especially in addressing privacy leak, bias, safety, and evolving regulations. Existing approaches to LLM unlearning often rely on retain data or a reference LLM, yet they struggle to adequately balance unlearning performance with overall model utility. This challenge arises because leveraging explicit retain data or implicit knowledge of retain data from a reference LLM to fine-tune the model tends to blur the boundaries between the forgotten and retain data, as different queries often elicit similar responses. In this work, we propose eliminating the need to retain data or the reference LLM for response calibration in LLM unlearning. Recognizing that directly applying gradient ascent on the forget data often leads to optimization instability and poor performance, our method guides the LLM on what not to respond to, and importantly, how to respond, based on the forget data. Hence, we introduce Forget data only Loss AjustmenT (FLAT), a "flat" loss adjustment approach which addresses these issues by maximizing $f$-divergence between the available template answer and the forget answer only w.r.t. the forget data. The variational form of the defined $f$-divergence theoretically provides a way of loss adjustment by assigning different importance weights for the learning w.r.t. template responses and the forgetting of responses subject to unlearning. Empirical results demonstrate that our approach not only achieves superior unlearning performance compared to existing methods but also minimizes the impact on the model’s retained capabilities, ensuring high utility across diverse tasks, including copyrighted content unlearning on Harry Potter dataset and MUSE Benchmark, and entity unlearning on the TOFU dataset.

Detecting anomalies in temporal data has gained significant attention across various real-world applications, aiming to identify unusual events and mitigate potential hazards. In practice, situations often involve a mix of segment-level labels (detected abnormal events with segments of time points) and unlabeled data (undetected events), while the ideal algorithmic outcome should be point-level predictions. Therefore, the huge label information gap between training data and targets makes the task challenging. In this study, we formulate the above imperfect information as noisy labels and propose NRdetector, a noise-resilient framework that incorporates confidence-based sample selection, robust segment-level learning, and data-centric point-level detection for multivariate time series anomaly detection. Particularly, to bridge the information gap between noisy segment-level labels and missing point-level labels, we develop a novel loss function that can effectively mitigate the label noise and consider the temporal features. It encourages the smoothness of consecutive points and the separability of points from segments with different labels. Extensive experiments on real-world multivariate time series datasets with 11 different evaluation metrics demonstrate that NRdetector consistently achieves robust results across multiple real-world datasets, outperforming various baselines adapted to operate in our setting.

Large language models (LLMs) have advanced to encompass extensive knowledge across diverse domains. Yet controlling what a large language model should not know is important for ensuring alignment and thus safe use. However, accurately and efficiently unlearning knowledge from an LLM remains challenging due to the potential collateral damage caused by the fuzzy boundary between retention and forgetting, and the large computational requirements for optimization across state-of-the-art models with hundreds of billions of parameters. In this work, we present \textbf{Embedding-COrrupted (ECO) Prompts}, a lightweight unlearning framework for large language models to address both the challenges of knowledge entanglement and unlearning efficiency. Instead of relying on the LLM itself to unlearn, we enforce an unlearned state during inference by employing a prompt classifier to identify and safeguard prompts to forget. We learn corruptions added to prompt embeddings via zeroth order optimization toward the unlearning objective offline and corrupt prompts flagged by the classifier during inference. We find that these embedding-corrupted prompts not only lead to desirable outputs that satisfy the unlearning objective but also closely approximate the output from a model that has never been trained on the data intended for forgetting. Through extensive experiments on unlearning, we demonstrate the superiority of our method in achieving promising unlearning at \textit{nearly zero side effects} in general domains and domains closely related to the unlearned ones. Additionally, we highlight the scalability of our method to 100 LLMs, ranging from 0.5B to 236B parameters, incurring no additional cost as the number of parameters increases. We have made our code publicly available at \url{https://github.com/chrisliu298/llm-unlearn-eco}.

Security problems of deep reinforcement learning draw much attention recently. Previous works on adversary attack mainly focus on preventing the targeted agent from choosing the most desirable action at each step, which may not reduce the cumulative reward effectively. In this paper, we first investigate how changing features affect the cumulative reward achieved by an agent. The static reward impact map is introduced to quantify the influence on the reward of each feature experimentally. By focusing on tasks with the static reward impact map, an adversarial attack method against deep reinforcement learning aiming to minimize the cumulative reward is proposed. Features with the large reward impact are perturbed in crafting an adversarial sample. Deep Q-network is selected to demonstrate the performance of our attack method in the experiments. The results indicate that our proposed method achieves better performance than the existing one-time attack method and the random attack in terms of the cumulative reward and the successful attack rate under both white-box and black-box settings.

Tumor samples clustering based on biomolecular data is a hot issue of cancer classifications discovery. How to extract the valuable information from high dimensional genomic data is becoming an urgent problem in tumor samples clustering. In this paper, we introduce manifold regularization into low-rank representation model and present a novel method named Mixed-norm Laplacian regularized Low-Rank Representation (MLLRR) to identify the differentially expressed genes for tumor clustering based on gene expression data. Then, in order to advance the accuracy and stability of tumor clustering, we establish the clustering model based on Penalized Matrix Decomposition (PMD) and propose a novel cluster method named MLLRR-PMD. In this method, the cancer clustering research includes three steps. First, the matrix of gene expression data is decomposed into a low rank representation matrix and a sparse matrix by MLLRR. Second, the differentially expressed genes are identified based on the sparse matrix. Finally, the PMD is applied to cluster the samples based on the differentially expressed genes. The experiment results on simulation data and real genomic data illustrate that MLLRR method enhances the robustness to outliers and achieves remarkable performance in the extraction of differentially expressed genes.

Motivation: Species and gene trees represent how species and individual loci within their genomes evolve from their most recent common ancestors. These trees are central to addressing several questions in biology relating to, among other issues, species conservation, trait evolution and gene function. Consequently, their accurate inference from genomic data is a major endeavor. One approach to their inference is to co-estimate species and gene trees from genome-wide data. Indeed, Bayesian methods based on this approach already exist. However, these methods are very slow, limiting their applicability to datasets with small numbers of taxa. The more commonly used approach is to first infer gene trees individually, and then use gene tree estimates to infer the species tree. Methods in this category rely significantly on the accuracy of the gene trees which is often not high when the dataset includes closely related species. Results: In this work, we introduce a simple, yet effective, iterative method for co-estimating gene and species trees from sequence data of multiple, unlinked loci. In every iteration, the method estimates a species tree, uses it as a generative process to simulate a collection of gene trees, and then selects gene trees for the individual loci from among the simulated gene trees by making use of the sequence data. We demonstrate the accuracy and efficiency of our method on simulated as well as biological data, and compare them to those of existing competing methods. Availability and implementation: The method has been implemented in PhyloNet, which is publicly available at http://bioinfocs.rice.edu/phylonet.

A data-centric network architecture, Named Data Networking (NDN) has been developed to meet applications' growing demands of network effciency and resilience. Currently, the deployment of NDN in real network environments requires careful system design to not only enable NDN but also support IP traffc, considering IP network has been prevalent for decades and almost all the equipments and applications are IP-based. In this paper, we take the most popular local area network (LAN) technology, Ethernet, as an example to investigate incremental deployment of NDN. Assuming a local network with both NDN and IP traffc, we mainly layout three deployment scenarios: NDN-enabled hosts and all Ethernet switches, NDN-enabled hosts and all Dual-Stack switches (i.e., it can process both NDN and IP traffc), and a hybrid network with both Dual-Stack switches and Ethernet switches. We examine the technical issues involved in each scenario and present solutions. In particular, in the hybrid scenario, we propose heuristics to optimize the placement of Dual-Stack switches. Compared with traditional Ethernet, introducing Dual-Stack switches can improve network effciency and resiliency by utilizing more links, reducing each link's traffc load, and taking shorter paths, at the same time also maintaining the functionality of IP-based applications.

With the advent of sequencing technology, numerous gene expression data are generated. Identifying differentially expressed genes play an important role in the gene therapy of cancer patients. As an useful mathematical tool, nonnegative matrix factorization (NMF) has been successfully used for identifying differentially expressed genes. In this paper, a novel method named robust graph regularized sparse orthogonal nonnegative matrix factorization (RGSON) is proposed and used for identifying differentially expressed genes, which introduces manifold learning, L1and orthogonal constraints into the objective function. In particular, L2,1-norm minimization is enforced on the objective function to improve the robustness of the algorithm. To prove the validity of the algorithm, experiments on the real genomic dataset are conducted. The results show that RGSON performs more effective than many other methods for identifying differentially expressed genes.

Low-rank representation (LRR) via rank minimization is a high efficiency method for capturing low-dimensional structure embedded in high-dimensional data. However, minimizing the rank of a matrix is NP-hard. In this paper, robust truncated nuclear norm low-rank representation regularized by L2,1-norm method (RTLRR) is proposed. The truncated nuclear norm is introduced to replace the nuclear norm to approximate the rank function. At the same time, L2,1-norm is used to regularize the sparse matrix to achieve better sparse effect of the algorithm. The proposed method is divided into two steps. Firstly, we do singular value decomposition (SVD) to the original data matrix. Then we apply the truncated nuclear norm and L2,1-norm constraints to subproblems and use inexact augmented Lagrange multiplier method to solve subproblems. Finally, the genes with high scores will be identified as differentially expressed genes according to the sparse matrix. The results on The Cancer Genome Atlas (TCGA) data illustrate that the effectiveness of RTLRR method outperforms many methods.

Sparse Principal Component Analysis (SPCA) is a method that can get the sparse loadings of the principal components (PCs), and it may formulate PCA as a regression-type optimization problem by using the elastic net. But the selected features are different with each PC and generally independent. A new method named SPCA has been proposed for removing these detect, which replaces the elastic net with L2,1-norm penalty. The results of the method on gene expression data are still unknown. Therefore, we will take a test to prove this point in this paper. Firstly, this method is applied to the simulated data for obtaining an optimal parameter. Secondly, the L2,1SPCA method is applied to the gene expression data, that is the head and neck squamous carcinoma data (HNSC). Thirdly, the characteristic genes are selected according the PCs. The results consist of very lower P-value and very higher hit count, which shows the method of L2,1SPCA can obtain higher recognition accuracy and higher relevancy to the genes. Finally, the experimental results demonstrate that the L2,1SPCA works well and has good performances in the gene expression data.

Robust Principal Component Analysis (RPCA) is an efficient method in the selection of differentially expressed genes. However, nuclear norm minimizes all singular values simultaneously, so it may not be the best solution to replace the low-rank function. In this paper, the truncated nuclear norm is introduced. And a new method named Truncated nuclear norm regularized Robust Principal Component Analysis (TRPCA) is proposed. The method decomposes the observation matrix of genomic data into a low-rank matrix and a sparse matrix. The differentially expressed genes can be selected according to the sparse matrix. The experimental results on the The Cancer Genome Atlas (TCGA) data illustrate that the TRPCA method outperforms other state-of-the-art methods in the selection of differentially expressed genes.